$ cat blog/README.md
Blog
A public log of what's changing across my repositories — releases, meaningful commits, and new projects. Only public activity appears here; routine chores, typo fixes, and automated merge noise are left out.
ls _posts/
Three READMEs that had drifted from the repositories: red-ink, slop-audit, diaz-modulus-lean
A pass over three READMEs that no longer matched what the repositories are. red-ink’s title now names the repository next to the game — “Sublevel Three (red-ink)” — so the page is findable under the name the project uses (b102897); slop-audit had kept “(private)” next to its own link long after it went public in September, and the line is gone (6178652).
guess-the-move goes live on GitHub Pages: Lighthouse from 63 to 92, three browsers in CI, and a link-preview card
The four days after guess-the-move went public went into making it publishable. Lighthouse on the real build scored 63 and now scores 92, with a budget check alongside it (#20, 8e1bedd); the browser specs run in Firefox and WebKit as well as Chrome (#21, 8033601); the comparison table orders its rows from best to worst with Stockfish first (#23, ef7f823); and a first play no longer waits for the whole dataset — a 24-position sample of about 20 KB gzipped is fetched first and the full file swaps in behind it (#24, 07cd9e7).
guess-the-move is public: a chess trainer that shows what Stockfish, a real player and Maia would do
guess-the-move went public on 7 November and plays at carlok.github.io/guess-the-move: a position comes up, you guess a move, and the app reveals three references and how far your guess is from each on a normalized scale — Stockfish’s best move, the move a real player of known rating actually played, and Maia’s move at that rating. The shipped set is 2,000 real positions from Lichess standard rated games of September 2026 (CC0), players rated 1200 to 2000, every legal move evaluated by Stockfish 19 at 400,000 nodes and stored with Maia-3 move probabilities; no player names are stored, but each position keeps its Lichess game link so a game can be taken out on request.
topshift-trend: the Distroless runtime moves up to Debian 13 and a Python 3.13 builder
Four days after moving production to Distroless Python, the topshift-trend image is on the next base: a Python 3.13 builder stage, Distroless Debian 13 for the runtime, and pip 26.2.2 in the build (60668b0, #12). The point of the earlier change holds — no shell and no package manager in the shipped layer — and the interpreter in the venv still has to match the one that built it, which is what the pinned builder stage is for.
red-ink sprint 30: the ambience was four loops of filtered noise, so now it is a mains hum and a crowd that answers blows
The ambience in red-ink was four looping noise sources — a low-passed room, two band-passed crowd murmurs, wind in bout 4 — with the crowd swelling on every hit, and filtered noise has no pitch or syllables, so it read as hiss. Sprint 30 replaced it: the room is a mains hum built from a fundamental and two partials, each two sine oscillators a few tenths of a hertz apart at base frequencies of 50, 41 and 56 Hz depending on the place, every bed event is now tonal, and the crowd is crowdReact, four to nine sawtooth voices through two vowel formants that answer a blow — 60 percent for a heavy, 30 for a dodge, 15 for a jab, 10 for a block, always for the knockout — at least 0.9 s apart, with nothing playing between blows (4a438a4, #29).
red-ink sprint 29: a training room with a dummy and a move-cost strip, medals with edges, and a limiter the mix actually needed
red-ink built ten backlog items in sprint 29 and moved the regression from 61 tests to 70. The new training room puts a dummy in front of a strip of move costs — jab 10, heavy 24, dodge 16 — that strikes out what the stamina on hand cannot afford, with a guard switch on a button and T; a test drives 240 frames of heavies and requires the dummy never to attack, move or die (#26). Medals got real edges — Untouched, Sharp at 85 percent, Quick at 26 / 31 / 40 / 46 / 38 / 39 seconds, no medal from Practice or the second pass, save version 3 — and the alert mark that warns of an incoming blow got a dark rim after measurement found it at 1.6 to 1 against bout 4’s pale floor, which the rim moved to 2.6 (f238b34).
dratify and cdclkit: CI now runs the gate CONTRIBUTING promised, on the Python floor and against two other solvers
cdclkit had been skipping two steps its CONTRIBUTING.md said ran as part of make gate — make lint and bench/compare.py --quick — and CI now runs both, with a new agreement job comparing the solver’s answers against MiniSat and CaDiCaL (#13, f54f106). The same pass adds .github/workflows/audit.yml for cargo audit over native/Cargo.lock, on every lockfile change and weekly, and a Dependabot pip entry for the build tooling, with dratify and cdclkit-native excluded because their versions move only with a release.
caciarabot: CodeQL's secret-logging alert reopens on a probability, and the fix enumerates the whole config
CodeQL alert 1 in caciarabot came back — the py/clear-text-logging-sensitive-data finding reopened after the startup log line added on 4 November, this time tracing bot_config.llm_secret_probability into settings_loaded. The value is a float; the rule matched the word in its name. Instead of naming fields one at a time, logging_utils.py now walks the dataclass with dataclasses.asdict, so the whole effective configuration is logged at startup and a field added later is included without anyone remembering to add it (28b1391), with a new tests/test_logging_utils.py covering the rendering. It is the same false-positive shape as the September segreto count fix — a rule named for credentials firing on the word rather than on one — so the enumeration retires the next instance before it is filed.
red-ink sprint 28: foes read a string of jabs, and the bout table is retuned to hold
A jab in red-ink lands 0.10 s after it starts, and five of the six foes wait cfg.react — 0.24 to 0.445 s at Normal — before they answer anything, so they could never answer one: about 1,500 to 2,200 jabs per 100 fights went through against Dag, Pike, Voss, Chen and Halvorsen with none blocked or dodged. balance-run can now say that out loud: strike() counts the player’s jabs by how the foe received them (jabHit, jabBlock, jabDodge in G.log) and each row prints jabAnswered, the share blocked or dodged. Baseline, 200 fights per bout on seed 1: 30.4 / 0 / 0 / 0 / 0 / 0 percent — only Reyes, whose 0.08 s reaction is under a jab’s 0.20 s, ever answers one (2147429).
deck-lovers: Jest coverage writes to test-results/, so the Actions runner stops hitting EACCES
Jest in deck-lovers was told to write HTML coverage to an absolute /reports/converter-js and /reports/server-js, a path that exists inside the project’s own test image but not on a GitHub Actions runner, where npm test -- --ci then fails writing the report. Both coverageDirectory values now point at ../test-results/* relative to each workspace (0a532fa), while the Podman test image keeps passing its own --coverageDirectory overrides from Dockerfile.js.test, so the container path is unchanged. The change sits on daily-care/2026-11-04-jest-coverage-path as #7, open at the time of writing.
caciarabot: the running bot now says which settings it loaded, and /stats shows what actually fired
The complaint was that CACIARABOT_LLM_REPLY_PROBABILITY=0.05 still replies too often, and the roll turned out to be correct — 5.01% over 200,000 simulated messages, one roll per message, no second handler — which left the value the running process actually holds as the suspect. That is hard to inspect in caciarabot now that production runs Distroless: no shell, no printenv. A .env edit followed by podman compose restart also keeps the old value, because restart does not re-read env_file; only down then up -d does, which deploy/update.sh already does (729dedc).
red-ink sprint 27: a crash card you can copy from, and a save that survives a bad value and a second tab
The sprint 27 branch of red-ink catches what used to freeze the page in silence: the first uncaught error or rejected promise pauses a fight and, 300 ms later, puts up the full-screen card with “Something broke” and a report — Babylon version, user agent, state, bout, language, quality level, viewport, the last five playtest-log entries, the error and its first stack lines — plus Reload and Copy details, with “Keep playing” offered only while frames still arrive; resource load errors, ResizeObserver noise, extension scripts and cross-origin “Script error.” are ignored, and nothing is sent anywhere (c904b72).
red-ink sprint 26: a second pass through the six bouts, WASD by physical key, and a save that admits when it failed
Once all six bouts are won, red-ink’s title card offers “Second pass: bout N”: the same six fighters at the player’s own difficulty, each with one change — Reyes iron jaw at 1.4× health, Dag quick hands with reaction 0.8× and cooldown 0.85×, Pike and Voss heavier hands at 1.1× and 1.25× damage, Chen relentless at 0.8× cooldown, Halvorsen iron jaw at 1.06× health — with the HUD reading “BOUT n / 6 · II”, wins counting only in progress.pass2, and the save moving to version 2 with a migrate step that adds the field (e2a32bf). Measured with the new balance-run --pass2, 200 bot meets on Normal and seed 1: the first pass is identical at 90.5 / 67.5 / 44.0 / 36.0 / 30.0 / 22.0 while the second reads 71.5 / 52.0 / 16.5 / 16.0 / 12.0 / 11.5, and the plan’s Hard multipliers had to be dropped — Hard alone takes the bot to 0 against Pike and Halvorsen — as did block and dodge changes, which did not move the bot at all.
red-ink sprint 25: practice mode, an exportable playtest log, and an Italian pass that found English in the HUD
red-ink gained a Practice link on the title card, shown once a bout has been won: it opens the opponent’s dossier with Another opponent and Back, a practice fight starts at full health and grants no wins, unlocks or saved progress, and its card says so and offers Fight again or Choose opponent (80d568d). Every finished bout now writes one entry to a playtest log — result, time, hits taken, blocks, dodges, punches thrown and landed, difficulty, who controlled it, block mode and the practice flag — keeping the newest 200 in browser storage, with Export playtest log and a two-tap Clear beside it in the title settings; the export is a JSON download, the clipboard if that fails, and nothing is sent anywhere (a84e6df).
red-ink sprint 24: a voice from the browser, autoplay, and punches thrown and landed on the result cards
The robot voice of sprint 23 is now the last fallback. Each word takes the first source that works: a clip listed in assets/voice/voice.js, then the browser’s own speech engine — a male voice of the game’s language, pitch 0.7, macOS novelty voices like Bahh and Fred skipped — then the formant synthesiser (b93cdd8). Nothing is bundled because Apple’s licence keeps the macOS voices to personal, non-commercial use and the Piper voices found were non-commercial or unclear, while the repo’s audio is claimed CC0; a Voice checkbox in the title settings and the pause menu turns it off, a voice list offers a Hear button, FIGHT is spoken about 0.2 s before the banner, and the banners themselves are localised (COMBATTI!, A TERRA). Regression stands at 31/31 with a stubbed speech engine; on this Mac the real list picks Daniel for English and Eddy for Italian, and how that sounds on an actual device is not something a headless probe can measure.
deck-lovers moves its server runtime to Distroless, with a healthcheck that needs no shell
deck-lovers now builds its server in two stages: python:3.11-slim-bookworm for the dependencies and gcr.io/distroless/python3-debian12:nonroot for the runtime, with the builder pinned to bookworm so its glibc matches the runtime’s (#6, 4c22adb). The catch was Python’s own path: Debian’s /usr/bin/python3 searches only dist-packages, so the packages pip installed with --prefix /usr/local were not importable and the container exited on No module named uvicorn until PYTHONPATH was set.
topshift-trend moves production to Distroless Python, and the venv had to be built by the right python
topshift-trend now builds its production image on gcr.io/distroless/python3-debian12:nonroot in a multi-stage Containerfile, with the virtualenv built on python:3.11-slim-bookworm to match the Distroless interpreter and requires-python relaxed to >=3.11 for it (#11, ead6607).
roundstorm takes zod 4.6.7 for an advisory; dratify's README catches up with its test suite
roundstorm moved its transitive zod to 4.6.7 for GHSA-mrhx-c7xc-m3w5, a lockfile-only change — package-lock.json, three lines added and eighteen removed, refresh of the optional libc metadata included (f23e9bb).
red-ink sprint 23: 80s synth timbres, three result themes, and a formant voice that says FIGHT
The sequencer in red-ink was rebuilt in 80s voices: a detuned-saw lead with chorus and a dotted delay, a gated-reverb snare cut at 180 ms, claps and toms, an octave-pulsing bass with a sub, a chorused pad under a sidechain pump, open hats from two looping noise chains and 16th arpeggios over an i–VI–III–VII line (9ceb543). Six places keep their own tempo and mode — E minor 84 bpm, A minor 100, D dorian 110, C minor 92 with a DX bell, F minor 118 with a march snare, G minor 126 — and the phrygian and harmonic modes are gone. The audio probe measures the change: the snare is −7.7 dB at 100 ms and −64 dB at 300 ms where it used to be a click, the title card sits at −27.7 dBFS rms (was −26.1) and the tense moment at −25.3 (was −23.9), while the nodes created per 10 s in that moment went from 162 to 285 in The Sump, 375 at The Blind Map and 471 at The Reconciler — over the plan’s limit of 250, since every note costs a few nodes, though the hats now cost none.
diaz-modulus-lean: a companion note shows Roy's rank bound cannot reach u⁴ at rank at most three
A new companion note in diaz-modulus-lean, tex/diaz-roy-rank-three.tex with its checks under scripts/roy_rank/, closes the question a scoping study of 23 October had left open and guessed would take days: whether matrices larger than the rank-one case — that is, beyond the strong six exponentials theorem stopping at u³ — can break Roy’s rank bound (1992, Cor. 1) far enough to exclude u⁴ (f5b4d6c). The answer is no at rank at most three. Nothing here is formal: no Lean file changed and the library stays at 618 results, with note v1.41 still the newest release.
alien-hunt: a way to test whether the unavailability term chooses anything
In the five runs of alien-hunt whose records show what was actually hunted — the fourth on PyJWT and four on joserfc — the hunted set is exactly the candidates at or above the coherence floor, and the cap of 20 never cut one. So the floor chose and the unavailability term only set the order, which is the question the paper’s author put on the table: compare the combined score against coherence alone and against a random draw, under the same investigation budget (a802c44).
red-ink is public: Sublevel Three, a warehouse fighting game in one HTML file
red-ink went public on 1 November and plays at carlok.github.io/red-ink: Sublevel Three, a browser fighting game where you audit pallet counts on the night shift at Halvorsen Freight until Pike hands you a keycard and points at a door that is not on the floor plan. Six named bouts happen on a taped circle in a warehouse — Reyes teaches, Dag answers a miss with a marked heavy, Pike repeats your last three punches — with a hub between fights, a loading-dock beat, and EN/IT card copy drawn from a single STR table capped at 55 words per card (#1, daea07a).
gusinistro starred alien-hunt
gusinistro put a star on alien-hunt on 1 November — its first, and the twenty-seventh star across these repositories. The profile carries no name and no bio, and only two repositories, both of them working experiments rather than portfolios: cctv-dvr-spacetimedb, a DVR and CCTV monitoring system in TypeScript on SpacetimeDB with some AI features, and voxel4d, a C++17 proof of concept for sparse voxel reconstruction, DDA traversal and Doppler-field sampling.
diaz-modulus-lean: Waldschmidt's five exponentials theorem machine-checked — notes v1.40 and v1.41
Two rounds on 1 November took diaz-modulus-lean from 596 to 618 proved results. Note v1.40 adds Waldschmidt’s five exponentials theorem — 1988, Cor. 2.2, DALAG §11.3.3 — proved the way DALAG’s remark after Cor. 11.16 does it, from Roy’s strong six exponentials theorem together with Baker’s theorem, with the two consequences the sources print: one of 2^π and 3^π is transcendental (Waldschmidt 1988, (2.2.3)) and DALAG Exercise 11.8(b) (4abe905). The sharp six exponentials theorem of Cor. 2.1 is not formalised, and the release says so.
Contributed to pymodbus and libmodbus: two Modbus servers disagreeing on the wire, four bugs
Four protocol-conformance bugs went upstream to the two Modbus implementations on 30 October, each found by differential testing one against the other: the same raw request put on the wire, the two answers compared byte for byte. pymodbus-dev/pymodbus#3074 was merged on 31 October by janiversen (three files, +39/−2): a Write Single Coil request (0x05) with a value other than 0xFF00 or 0x0000 used to be carried out — any non-zero value turned the coil on — and answered with an echoed 0xFF00 the client never sent, where the specification’s state diagram answers exception 03. WriteSingleCoilRequest now keeps the decoded value and datastore_update answers ILLEGAL_VALUE before touching the datastore.
hidden-intersections is public: a line that misses a circle still meets it, and the classification that follows
hidden-intersections went public on 31 October, a reply in code to Inigo Quilez’s Hidden intersections short: a line that misses a circle still meets it at two complex conjugate points, and sweeping a family of parallel lines glues the real curve and its phantom branches into one locus at the tangency points. Circles and lines become points of P³(C) under a quadratic form Q, so inverting one circle in another is a linear reflection; mirroring makes the circles a complex kei and the mirrors generate PO(4,C), a circle meets its mirror image exactly where it meets the mirror, and one number sorts everything — κ = ⟨C,L⟩²/(Q(C)Q(L)), which is 1 at tangency, cos²θ for crossing circles, and can be negative (2b17299).
hermes-podman: pinned to Hermes v0.22.5, and the three things that make a healthy upgrade look broken
hermes-podman now pins docker.io/nousresearch/hermes-agent at v0.22.5@sha256:9c5561579f3f… in compose.yaml and in the quadlet deploy/hermes.container together, one commit for both sites, with the digest taken from the Docker Hub tag record rather than typed (8c044e1). The jump is from v0.21.6, so it crosses the release boundary where docs/UPGRADE.md had nothing to say.
pacenotch: the daily update check is on by default, and the notice links the build that matches your machine
pacenotch now checks for a newer release once a day on new installs, through DefaultSettings rather than a change to anyone’s saved configuration (ed93612, #3). The notice is the part worth reading: internal/update parses the release’s assets, matches the GUI build for the current GOOS/GOARCH by prefix and falls back to the release page when nothing matches, so the tray and the About window can offer a download action and the release notes instead of a link to a page where the reader picks their own binary.
diaz-modulus-lean: Roy's strong six exponentials theorem now holds with no hypothesis — note v1.39
The last hypothesis in diaz-modulus-lean is gone. Since notes v1.36–v1.38 Roy’s theorem was machine-checked from one explicit hypothesis, (ME) — the case G⁻ = {e}, G⁺ = G of the multiplicity estimate of DALAG Th. 8.1, for G = 𝔾ₐ^{A₀} × 𝔾ₘ^{A₁}, written in coordinates, with only its Hilbert side proved. Thirty-two nodes on 30 October prove its multiplicity side and assemble it as DiazModulus.multiplicityEstimateElem, so Diaz.strongSixExponentials holds unconditionally and Diaz/RoyUnconditional.lean restates each of the 25 results that had taken Roy’s theorem as a hypothesis without it (3e513e4, note-v1.39). The mirror stands at 596 of 596 results, the appendix at 226 identifiers, and the companion note at 44 pages.
contributed to The Alien Space of Science and joserfc: where the idea went, and a one-second JWT bug
Two things went upstream from alien-hunt. On alejandrohdez00/alien-space-of-science — the paper and code the sampler is adapted from — issue #2 says where the idea was taken: what replaces the paper’s trained coherence and availability models, the two results that survived a reproducer, and the two that did not. It asks for nothing.
alien-hunt is public: sampling the defects a plain review of a repository does not report
alien-hunt went public on 30 October: a Python research prototype, MIT, that looks for defects a plain review of a repository does not already produce and then checks whether a reproducer can actually observe them. It is the cheap version of the sampler in The Alien Space of Science (code, MIT), with a model’s 1–5 coherence rating standing in for the paper’s trained models: score = (1 − β)·z(coherence) + β·z(unavailability), where unavailability is one minus the maximum cosine similarity to the claims three plain reviews and any local scanners had already made.
LeanFrontier: Field Note 28 — the series has been on GPT-6 since 17 October, and the notes had not said so
Field Note 28, Same Agent, New Model., is mostly a correction to the two notes before it. The two submissions it describes are already covered here — the Stern–Brocot depth formula of 21 October and the forced Abel identity with its Lagrange–Green sum of 29 October — and what the note adds is that every submission since 17 October, all twelve of them, credits GPT-6, from the same contributor and the same agent, while the last one credited to GPT-5.6 Sol was accepted on 7 October. Notes 26 and 27 described that work without mentioning the change.
LeanFrontier: the discrete Lagrange–Green identity, with forcing, in any commutative ring
#546 proves the discrete Lagrange–Green identity for two forced solutions of a second-order recurrence with variable coefficients: with u(n+2) = p n·u(n+1) − q n·u n + f n and v driven by g, the Casoratian satisfies D(n+1) = q n·D n + u(n+1)·g n − f n·v(n+1), and unrolling it gives the closed Green-source sum — D n = (∏ i<n, q i)·D 0 + ∑ j<n (∏ j<i<n, q i)·(u(j+1)·g j − f j·v(j+1)) (b52dee6).
forgepulse v0.2.0 moves production to Distroless, adds a built-in healthcheck, and prints its own version
forgepulse released
v0.2.0 on 28 October
(34d03d2). The production
image is now gcr.io/distroless/cc-debian12:nonroot with a built-in
forgepulse-server healthcheck instead of a shell-less Alpine base, and the
README gained the section the move needs: an existing data volume must be
chowned to uid 65532 once, before the container will read it
(#8, c97d705,
256f853). The header
now shows what is running — v0.2.0 · main@abc1234 — and /api/health
reports the same pair, so a stale deploy is visible without reading the image
digest.
caciarabot: a Distroless runtime halves the image, and CI now runs the config validator
caciarabot’s production stage moved from python:3.12-slim to gcr.io/distroless/python3-debian12:nonroot (#8, 2bfc4e2): no shell, no package manager, no curl, entrypoint still the exec-form python -m caciarabot.main, running as UID 65532 with /data prepared and owned in a slim runtime-prep stage. The built image drops from 213 MB to 114 MB.
carlok/pacenotch was forked by matteodesimone, and his tray-icon patch shipped as v0.2.2 the same day
carlok/pacenotch was forked by
matteodesimone on 28 October —
Matteo De Simone, in Italy, who forked LeanFrontier
in August and used it the way the corpus expects, as a branch to submit from.
This time the fork came back as a patch to a different project: PR #1
noticed that in macOS light mode the menu-bar icon uses LightBars(), whose
on-pace fill 0xB25000 is an amber chosen to stay readable as text on white —
at menu-bar size almost the same as the ahead red, while the window shows
yellow for the same row. TrayBars(dark) keeps the palettes but gives the
light tray 0xFFCC00, Apple’s system yellow, since an icon carries no text;
the window palette is untouched, and a new test pins the tray’s on-pace fill to
yellow in both modes.
matrix-podman-stack: an integration suite that runs the real stack, and pip-audit in CI
matrix-podman-stack grew the layer the unit tests could not reach. docs/testing-plan.md inventories the gaps, and the new suite closes them by running the actual containers: a Postgres 16→18 migration checked end to end through /health and through Matrix users, a room and messages that survive it, a Garage in-place upgrade driven with boto3, the NATS/Valkey bus hooks, and the monitoring profile including the Grafana datasource proxy query (#27, 1148def).
hermes-podman: a Renovate config that keeps compose and the quadlet on one tag@digest
hermes-podman pins its image by tag and digest in two places — compose.yaml through the docker-compose manager, deploy/hermes.container through the quadlet manager — and until now nothing kept the two in step. A branch adds renovate.json with docker:pinDigests so both move together (f4f2b52): minor and major Hermes bumps arrive grouped in one pull request, patches are skipped, and the GitHub Actions versions are batched weekly.
chrikrah follows the work
chrikrah — Christopher Krah, in Germany, who writes at plexito.de about AI agents breaking in production and sends fixes upstream to the agent, eval and LLM tooling he uses, with Plexito and UiPath in the profile — is the eighteenth follower of this account. That is close to the audience these repositories are built for: someone who reads agent failures for a living and files the bug rather than the complaint.
matrix-podman-stack: Postgres 18 with a migration script, a Synapse health smoke test, and a bot bus on nats-py 2.16 and redis 8.1
A day after going public, matrix-podman-stack moved both Postgres services to 18.6-bookworm with the official /var/lib/postgresql volume layout, and shipped upgrade-postgres.sh for data that is already on 16 — the reason the change is one pull request and not a compose edit (#23, 2669435). A migration that completes is not the same as a Synapse that starts, so verify-synapse-postgres.sh now walks the stack’s own /health endpoint after the upgrade and records what it saw (#25, d0ac04b).
storygen, eclipse-3d-sim, collective-starship-game, agility-trainer and solids-hunter: one vitest advisory, five repositories
The same test-runner advisory reached five of the game repositories, and each took vitest 4.1.12 for it: GHSA-xp26-3454-mjgp, a Windows backslash traversal that bypasses @vitest/mocker’s redirect allow-list. storygen bumped vitest (#7, 6a9e3a5) and then vite to 6.4.5 (#9, a2ad1a5); eclipse-3d-sim did both (#3, #4); collective-starship-game took vitest 4.1.12 together with the transitive packages behind it (#7, e7c157c).
matrix-podman-stack is public: a private Matrix server for Podman, where the bot accounts are configuration
matrix-podman-stack is now public: Synapse and its Postgres, Element Web, an optional mautrix-telegram bridge with a second Postgres, optional Prometheus and Grafana, and the helper scripts that provision the chat’s automation accounts. It is built for a small private group of humans and bots rather than for federation — no federation, no TURN, no bridge besides Telegram — and every container publishes on 127.0.0.1 only, because nginx stays on the host and terminates TLS there. Images are pinned (Synapse v1.163.2, Element v1.12.32, Postgres 16.9-bookworm, mautrix-telegram v0.2610.0), and the optional parts are compose profiles: telegram, monitoring, bus, object, vectors, llm. The stack itself landed as #1.
lspace-det-sigma: the k32 figures are drawn from the note's own braid words, and SnapPy says the drawing is the knot
The two figures of the k32 note in lspace-det-sigma — H(2,3) with its unknotting, and one bundle crossing X₃ with every pair (i,j) marked under its crossing — are now generated by note/figures/make_figures.py straight from the same words the note’s proof uses, rather than drawn by hand (a76eea1). The revision also tags the PDF sent on 11 October as k32-v1, so the version referees read is recoverable.
forgepulse: two certificate-trust advisories closed, and the two that cannot be
forgepulse took the transitive TLS bump for RUSTSEC-2026-0381 and RUSTSEC-2026-0382: webpki-roots 1.0.9 → 1.0.11 and rustls-webpki 0.103.15 → 0.103.16, lockfile only, no manifest change (#6, 2782e74). cargo audit goes from three vulnerabilities to one — RUSTSEC-2023-0071 in rsa, which has no upstream fix — plus a tokio unsoundness warning, RUSTSEC-2026-0386, that waits on a patched release; the workspace’s 54 tests pass either side of the bump.
aristowrap moves to Python 3.15, and CI follows the image instead of trailing it
aristowrap now builds from python:3.15-slim-bookworm and tests 3.15 in CI (#11, a81a665). Dependabot had already opened the base-image bump on its own, but the workflow matrix still covered only 3.11–3.14, so the change that landed applies the image and extends the matrix, the Programming Language :: Python :: 3.15 classifier and the docs in one commit — requires-python stays at >=3.11.
dratify: parse_dimacs gives a body literal the header's bound, and counts the tautologies CNF.add drops
Two holes in parse_dimacs closed on 25 October. The header guard refused p cnf 3000000 0, but the body form 1 3000000 0 — twelve characters — set CNF.nvars to 3,000,000, which is what sizes the checker’s arrays; a literal is now bounded like the header, by max(len(text), HEADER_VARS_FLOOR), checked before it is stored, with the ValueError naming the line and the literal (CWE-789, #5, 7bf8057).
contributed to urllib3: Timeout.read_timeout should honour total before a connection starts
Opened urllib3#5336, a one-property fix in Timeout.read_timeout. The docstring says that when a read timeout and a total are both given “the shorter timeout will be applied”, but before start_connect() has been called the property returns read as it is, so the longer one wins: Timeout(total=1, read=2).read_timeout is 2. After the connect clock starts it already returns min(total - elapsed, read), so the patch makes the not-started branch agree — no time has elapsed, hence min(total, read), the same value it returns immediately after start_connect().
portcullis: the remote script is read until EOF, so Phase 2's tail stops vanishing from the log
portcullis had a logging bug that only showed on a real run: execute_remote_script stopped reading as soon as the exit status was ready and nothing looked buffered, but the exit status can arrive before the last output chunks — so the tail of Phase 2 and the last 11 verify.sh checks, including its results line, were simply missing from the log of an otherwise successful provision (bb81c01).
pickmydegree: code-scanning alerts #3 and #4 fixed — a workflow token that was too wide and an HTML stripper that a browser would parse differently
Two pickmydegree code-scanning alerts were open and neither was reachable through the API — querying the open alerts returned 403 Resource not accessible by integration — so the findings were reproduced locally instead, with CodeQL CLI 2.27.3 against the default code-scanning and security-and-quality suites for JavaScript, Python and GitHub Actions (84a241a).
dratify 0.1.9 and 0.1.10: a repeated literal is the same clause, and a proof is a stream of clauses; cdclkit 0.1.7 carries the first fix into the native half
dratify 0.1.9 fixes a false rejection in both checkers: they counted literal occurrences, so 1 1 0 — the clause (1) — never became unit at the root and a later step that needed it was refused, while 1 0 / 0 verified and 1 1 0 / 0 did not; d 1 1 0 also got past the ignore-unit-deletions rule and deleted the unit. Repeats are now removed at the top of check_step keeping first occurrences so the RAT pivot is unchanged, false rejections only and nothing invalid accepted (8a06896). tests/naive.py counted the same way, so the reference could not see the bug; it now treats clauses as sets, and its property test and the agreement fuzz generate repeated literals, with new end-to-end tests on both engines.
contributed to telaviv/strong-six-exponentials: a licence request, so a theorem becomes a dependency
strong-six-exponentials is someone else’s Lean 4 formalisation of Roy’s strong six exponentials theorem — public since 18 October, on the same Lean and Mathlib revision — and it is the one hypothesis standing between diaz-modulus-lean and a batch of unconditional results: the library’s theorems take that theorem as the hypothesis hSSE, and the two statements are a short bridge apart. So the contribution to that project is a question rather than a patch, issue #1 opened on 23 October and addressed to Shawn Krisman: the repository has no LICENSE file while its file headers say Apache 2.0, so a licence with a named copyright holder is what would let the work depend on it through Lake, with credit in the README and the companion note.
roundstorm adopts the Contributor Covenant, with GitHub as the enforcement channel
roundstorm now carries a
CODE_OF_CONDUCT.md:
the Contributor Covenant 2.1 in full, with enforcement routed through GitHub —
a report goes to the maintainer’s GitHub contact rather than to an address that
has to be kept current — and a link to it from CONTRIBUTING.md
(#63,
af4bdab).
pickmydegree: a catalogue that rebuilds itself from the universities' own data, a game and accessibility pass, and a deploy on every push
pickmydegree — the degree chooser that narrows a university choice through tournament-style pairwise comparisons — grew the thing it had been missing: a catalogue that can be rebuilt instead of hand-edited. catalog/ is a separate subproject, outside the Vite build and the Docker image, whose catalog.py uses only the standard library to fetch PoliTo’s open-data CSV and course pages, UniTo’s paginated course listing (with a five-second crawl delay) and the MUR Offerta formativa file as a cross-check, merge them into out/courses.json, and diff the result against src/data/degrees.json through mapping.json; only enrich --write touches the site’s data, and only the university and url fields (6d6f459). The first refresh is for 2026/27: 100 cards gain a university and a course URL (76 UniTo, 24 PoliTo), one new card arrives — Economics and Finance with Data Science — the welcome text moves to 2026-27, and nothing is deleted: deg-033, deg-017 and deg-028 are flagged in out/TODO-possibly-discontinued.md for a decision rather than dropped.
diaz-modulus-lean: Theorem F settles which hulls of a rational z carry a configuration, and it agrees with Theorem E
Two rounds on 23 October took diaz-modulus-lean from 475 to 503 proved results, each with its own version of the companion note: v1.34 with R20 and v1.35 with R21a, the blueprint running to section 10.154 of chapter 10 and the audit to 138 theorems (26a7aac, 8cf2a67).
cdclkit: the scaling test steps 16x so a linear engine still fails it; dratify 0.1.8 is a documentation release
The scaling test in cdclkit watched for a regression by quadrupling the variables and expecting more than linear time, and a macOS runner duly reported 8.3x for 4x the variables on native runs of 4 and 32 ms — cache growth and timer noise, with a linear engine already sitting at 4.5x. Rather than tighten a threshold against noise, the test now steps by 16x with a limit of 64, where linear is about 20x and quadratic about 256x (b59350d). It was checked against the engine before the fix: cdclkit 0.1.4 measures 252x on native runs and 218x in Python at the new step, so the test still catches what it exists to catch.
contributed to pyasn1 and python-multipart: three ordering and chunking bugs
Three pull requests went upstream on 2026-10-21, each with a failing test first. On pyasn1/pyasn1, BitString ordering breaks on equal-length values: __lt__ and __gt__ compare length then value, but __le__ and __ge__ test len(a) <= len(b) or ..., already true when the lengths match, so the value is never reached — BitString(binValue='11') <= BitString(binValue='10') and the same pair with > both come back true. Comparing lengths strictly in the two operators fixes it, two tests go in and the suite moves from 1261 to 1263.
LeanFrontier: Stern–Brocot depth is the sum of the continued-fraction coefficients
#542 proves that for every Boolean path into the Stern–Brocot tree, the sum of the regular continued-fraction coefficients of the rational the path represents is the number of path edges plus one — (regularCoefficients path).sum = path.length + 1 — quantifying the classical correspondence between a rational’s depth in the tree and the total size of the quotients in its Euclidean expansion, with the empty-root convention and the leading zero for fractions below one handled uniformly (1e72baf).
forgepulse, neon-bumper-cars and cross-tetris: RUSTSEC advisories, a yanked chacha20, and a GITHUB_TOKEN that was too wide
Three repositories closed their open security alerts in one pass. forgepulse pins source-map-js to 1.2.2 through a pnpm override for GHSA-68fv-2mgg-jv7q — the same parser denial of service the games took on 17 October — moves quinn/quinn-proto for RUSTSEC-2026-0338, -0340 and -0343, takes chacha20 0.10.2 because 0.10.1 was yanked, and enables sqlx’s features explicitly with default-features off so the dependency set is the one declared (57f0c29, #4).
platosdf: the Flask engine API closes eleven CodeQL findings without trusting a filename
platosdf, the genetic evolver for invariant signed-distance-field solids, cleared its CodeQL backlog in the Python engine that serves grammars, meshes and the gallery. The GitHub API for alert details answered 403, so the findings were reproduced locally with CodeQL CLI 2.27.2 on the python-security-extended and javascript-security-extended suites, which is what the repo’s dynamic workflow runs anyway: eleven Python findings, among them py/flask-debug for the debug=True in __main__, py/path-injection twice — a user-supplied grammar name joined onto a path in engine/grammar_store.py, and the epoch/filename pair in gallery_file — and py/stack-trace-exposure on the JSON error bodies (9158692, #1).
LeanFrontier: Field Note 27 — three submissions that build on the corpus, and a definition consumed in three days
Field Note 27, Three Days to a Consumer., records a day whose three submissions all build on the corpus by both of the series’ reuse measures. Banks’ theorem and the four standalone subjects were the previous week’s shape; here three submissions from qazW12345 take up definitions that were already accepted. #532 proves the characteristic-polynomial shift law: if f(n+2) = s·f(n+1) + t·f(n) then its weighted binomial transform satisfies g(n+2) = (s + 2a)·g(n+1) + (t − sa − a²)·g(n), which is X² − sX − t translated to P(X − a), in any commutative ring and with no division — and it consumes binomialTransform from #512, accepted three days earlier.
staffarda: the walkthrough of the abbey guesthouse is on YouTube
The walkthrough of staffarda — the model introduced when the repository went public — is now published on YouTube: 2 min 43 s, the three source photos and then one continuous 1080p take through the hall, in through the doorway and slaloming between the four columns at eye height, rendered in EEVEE by blender/walkthrough.py (952162b). The README links it behind a clickable thumbnail. The soundtrack is “Antiphona et Magnificat” by Schola Gregoriana (WSD SAC Ołtarzew), from Wikimedia Commons under CC BY-SA 3.0 — and because of that ShareAlike music the video as a whole is CC BY-SA 4.0, while the code stays MIT and the renders and documentation CC BY 4.0.
staffarda: the model revised against the photos, and a walkthrough mode that renders the same moment in EEVEE and Cycles
The staffarda model was revised against the three source photos before the walkthrough was rendered: weathered materials instead of clean ones, wall corbels made higher and smaller, four column bases added, and the large windows kept as they are today with the arched heads left behind a hypothesis switch rather than assumed (2cc47a3, 97bb8d5) — the same policy as the rest of the reconstruction, where what the photos show and what the literature suggests are separate settings.
staffarda is public: a Blender reconstruction of the Staffarda Abbey guesthouse from three photos
staffarda is public: a procedural Blender model of the ground floor of the foresteria — the guesthouse — of Staffarda Abbey in Revello, Piedmont, the two-nave hall with four stone columns and a brick rib vault that once served as the pilgrims’ refectory. It was rebuilt from three phone photos taken on 16 October, with no survey, point cloud or photo projection: every wall, rib, capital and brick is generated by a Python script, and blender/build_foresteria.py keeps every dimension, colour and option as a constant at the top (883d0bc). The README is explicit about what that is worth — dimensions come from the perspective in the photos and are good to about 10–15%, and the result is a plausible, stylized reconstruction rather than a measured one.
storygen, eclipse-3d-sim and pickmydegree: a path injection, an XSS in Vue's SSR, and what the browser was remembering
Three of the older browser projects cleared their security backlog. storygen — short story videos rendered inside a container — closes six open CodeQL security-extended findings by hardening /api/image/{filename} with a basename check, a realpath normalisation and a root-prefix test that CodeQL recognises rather than a string comparison, and its CSP tests now parse the header into directives and compare URL sources with urlparse instead of substring-matching the whole string (4313707, #6).
forgepulse: human attention v2 scores every repo from fresh inputs
forgepulse’s human-attention score had three ways of lying. Empty referrer lists were read as unknown rather than as zero, so 55% of repos scored N/A; referrer snapshots never expired, and a 53-day-old one still ranked fifth on the chart; and repos under 30 days old had no star or fork baseline at all, so the newest projects were scored against nothing (affe833).
diaz-modulus-lean: notes v1.31–v1.33 formalise Theorem E, decide which degree-one spaces Theorem B sees, and make Theorem D a barrier
Three batches took diaz-modulus-lean from 428 to 475 proved results on Lean’s three standard axioms, each released as its own version of the companion note: v1.31 at mid-morning, v1.32 in the afternoon and v1.33 in the evening, with the blueprint at 217, 242 and 262 nodes and scripts/Audit.lean growing through 125, 130 and 133 theorems (04a756e, c9b26a6, 9b81477).
consculpta: a sound engine that reads the genome, sampled instruments, and a genome panel that opens itself
A day after going public, consculpta grew a sound engine instead of a soundtrack: src/audio/score.js maps a genome to key, mode, pace, swing, timbre and texture, scan.js reads the melody off the specimen’s own silhouette by marching inward along directions swept across the side facing the camera — the outermost inside radius sets the pitch, empty directions rest — and synth.js voices it with Web Audio pads, plucks, bells and a membrane thump for the drum block (10d9949). Because it follows Spin and Morph, the music moves when the sculpture does.
roundstorm: a trimmed transcript leaves one line per dropped message
The transcript cap in roundstorm used to cut silently: once the history budget bound, the oldest messages were simply gone. #61 now leaves one line per dropped message — - Alice, round 2 [object]: first sentence..., 140 characters at most, inside a 6,000-character or 15%-of-budget allowance whichever is smaller, newest lines first (866e3f5). There is no model call in it, so no added cost or latency, and the trade the note on the pull request names is that a first sentence is a crude gist. A simulated run of five agents over twelve rounds puts total prompt volume 22% below uncapped, 18% with the bare marker alone, with each turn bounded — but it is still not measured on a real five-by-five, which issue #33 asks for, so the pull request says Refs rather than Closes.
LeanFrontier: Field Note 26 — seven submissions in thirteen hours, Banks' theorem, and a recursion closed thirteen days later
Field Note 26, The Roadmap Was Read., records nine days without a submission — the last seven of them after the contribution roadmap was refreshed — ended by seven arriving in thirteen hours. Two take up the roadmap’s top directions: Banks’ 1992 theorem on sensitive dependence is proved for any continuous, topologically transitive self-map of an infinite metric space with dense periodic points, with sensitivity defined explicitly because Mathlib has no such notion, and importing nothing from the corpus — so the corpus’s own transitivity and dense-periodic-point results for the tent and logistic maps are one application away, still unapplied. The other closes the duplication Field Note 22 flagged on 4 October, when two submissions written two hours apart each defined the same way of peeling a Stern–Brocot path into runs and add-only review could not reconcile them: the quotients read from a path are now the run lengths of its encoding with one added to the last, and a companion proves the encoding injective and length-preserving (#511, #509).
neon-bumper-cars and collective-starship-game: proxy-addr IP spoofing, a source-map-js denial of service, and rate limiting from a CodeQL alert
Two of the browser games picked up security fixes. neon-bumper-cars moves its transitive proxy-addr from 2.0.7 to 2.0.8, which is GHSA-jqcg-44mw-7w3h / CVE-2026-90711 — IP spoofing through an IPv4-mapped IPv6 trust subnet, reached through express — shipped as Dependabot’s bump and then a hand-checked patch that confirms the root lockfile resolves to the patched version (d3498db). collective-starship-game takes source-map-js 1.2.2 for the high-severity denial of service in GHSA-68fv-2mgg-jv7q (CVE-2026-93749), and the same pull request answers CodeQL’s js/missing-rate-limiting on the SPA fallback route — sendFile with no limit in front of it — by putting express-rate-limit at 600 requests per minute per IP across the whole Express app, ahead of the static and Vite handlers (8d4ed28, #6).
consculpta is public: 3D sculptures bred from formulas over simplex noise, raymarched in WebGL 2
consculpta went public on 17 October: a browser-only breeder of 3D sculptures, Apache-2.0, live at carlok.github.io/consculpta after 41 commits and twenty-five pull requests on its first day. Each specimen is the region where a formula over 3D simplex noise rises above an iso level inside a sphere, cube or drum; the formula is a tree of noise, trigonometry and space transforms, and every offspring changes its parent’s tree by tweaking a number, swapping a function, grafting a branch, wrapping a node or cutting one out, with roughly one in twelve switching to strands — thin filaments traced where the surface crosses a rotated copy of itself. Before anything is compiled the page evaluates each formula on 520 sample points to normalise it, estimate a safe raymarching step and discard the empty or the solid, which is what makes a litter of six cheap enough to breed in a click.
portcullis: ENABLE_IPV6 decides whether the VM gets an IPv6 address at all
portcullis — the two-phase Hetzner provisioner that locks a VM down in about thirty seconds and then hardens it against a 58-check list — now takes ENABLE_IPV6 as a first-class setting (default true), validated by preflight.py in the words it accepts elsewhere (true/false, 1/0, yes/no, on/off, blank meaning the default) and reaching the API through build_public_net(), which always asks for IPv4 and for IPv6 only when the flag says so (78ecce3). Set to false, the firewall follows: firewall_sources() drops ::/0 and keeps 0.0.0.0/0, on both the port-22 rule created with the server and the lockdown rule that replaces it once Phase 1 has closed 22, so the only rule the VM ever has is IPv4-only and no address is allocated to be forgotten. What the guest does is unchanged — the hardening scripts disable IPv6 inside the OS either way — so with true the address is reserved and unused unless someone deliberately re-enables it, and the README says exactly that instead of leaving the gap between “allocated” and “reachable” to be discovered. The .env.example entry, the config table and tests for the parser, both firewall shapes and the public-network object came with it.
diaz-modulus-lean: note v1.30 makes the polynomial hulls a proposition, and the research note classifies every polynomial in u and ū
The Diaz companion note is at v1.30, which states the R16 batch as a single result: Proposition 5.21, ten new nodes and milestones 137–139, after which diaz-modulus-lean holds all 428 results on Lean’s three standard axioms and the blueprint 211 nodes, with 195 appendix identifiers checked against the board (2136117). Clause (a) is Theorem C: for u ∉ Q̄ with uū algebraic and P ∈ Q̄[X], the hull H₀ + Q̄P(u) + Q̄P(ū) carries a 2×3 configuration exactly when deg P is 2 or 3 — the configurations themselves are Diaz’s (2007, Théorème 7(1); 2004, Théorème 2), and the new half is the converse, which normalises a configuration to Laurent polynomials so that the orders at 0 and the degrees pinch every product’s support into [-1,1], the order sets of the hull being {−k, −1, 0, 1, m} and a sizes-2-and-3 sumset inside one of those being rigid. Clause (b) is that real polynomial expressions in Re u and Im u with algebraic coefficients carry no configuration in their hull, so Roy’s theorem cannot exclude (Re u)² one number at a time; (c) and (d) are the inversion lemma — over K ⊆ F, a configuration in V₀ + Kw + Kw⁻¹ with V₀ ⊆ F and w transcendental over F has its products in V₀ — and its barrier, that ℒ̃ + Q̄u + Q̄ū carries no configuration when u is transcendental over the field of ℒ̃, so Roy’s theorem alone cannot refute such a point. All ten results are routine by the blueprint’s own rubric, and none of them is a step towards Diaz’s conjecture: they say what Roy’s theorem cannot do.
curve-symmetry-lean moves to Lean 4.35.0, and six statements need new parentheses
curve-symmetry-lean moved to Lean 4.35.0 and Mathlib v4.35.0, 651 Mathlib commits past the 5ed2965 of v4.34.0, and the move is more than two lock files: five modules changed, one for the SetLike.exists_of_lt deprecation now spelled IsConcreteLE.exists_of_lt, and six statements across PlaceUniformizers and InfinityDifferentials where a singleton {f a (b)} written across lines began to parse as the start of a set-builder and needed to be written {(f a (b))} — a change of spelling that elaborates to the same statement (a2c13b7). No declaration was added, removed or renamed, the audit stays at 2,058, and none of the 30 modules shared with the public sharp-symmetry-bounds-lean repository changed. verification/lean-v4.35.md records the move, the Prove2Me copy stays on Lean v4.33.1 with its mission record untouched, and a follow-up commit records the passing CI runs (18d3984).
cdclkit and dratify: python-sat 1.9.dev16 fixes the truncated CaDiCaL proofs
The upstream bug that dratify’s fflush workaround existed for is fixed: python-sat 1.9.dev16 resolves pysathq/pysat#233, the truncated CaDiCaL proof files that made most of PySAT’s refutations fail to verify until the writer was flushed. Re-measured on the same 50 uuf100 instances, 50 of 50 now verify on all four CaDiCaL versions with no flush at all, against 4–8 of 50 on 1.9.dev15; the workaround stays, marked as needed only for older versions, and the changelog catches up (af19023). cdclkit’s roadmap records the confirmation, which closes the item the same chase had opened there (5a8fe71).
roundstorm: a failed request stops being silent, and an informational Windows cell
Two follow-ups landed in roundstorm behind 0.1.1. Six raw fetch(apiUrl(...)) calls in App.tsx now go through api() with a call helper, and the behaviour change that comes with it is the point: rename, duplicate and agent delete used to ignore a failed response silently, and now show the server’s error and stop (#59, 9ae96b1). The same pull request writes down the event convention logEvent had been following by hand — room, deliberation and agent ids are columns, other entities’ ids belong in the payload — moves the two events that concerned a room but carried no roomId (deliberation.interrupted, round.restarted) onto it so they appear in the room’s log, and gives backup.created one shape, { via, path, bytes } with path null for a download, on both routes, with a test.
roundstorm 0.1.1: Claude's tools become an allow-list, and a cancelled turn leaves nothing running
roundstorm shipped 0.1.1, seven pull requests of hardening and fixes on the local-first deliberation app, and a .dmg for Apple Silicon with its SHA-256 in SHA256SUMS.txt — on top of the per-launch token and the cleared advisories described earlier in the week.
python-hosts-checker: the certificate check asks for TLS 1.2 at least
python-hosts-checker — the AWS Lambda endpoint and certificate monitor whose findings arrive as Telegram alerts — now sets an explicit minimum_version of TLS 1.2 on the SSLContext it uses for the expiry handshake, so TLSv1 and TLSv1.1 can never be negotiated whatever the interpreter and OpenSSL defaults happen to be (e853782, #6). Python 3.10 and later already default create_default_context() to TLS 1.2, so nothing changes in practice; the floor becomes explicit instead of inherited, and it clears the CodeQL finding code-scanning #1 (py/insecure-protocol). A test captures the context handed to the handshake and asserts the floor rather than trusting the default, and the suite stands at 15 passing.
LeanFrontier: the corpus moves to Mathlib v4.35.0, and the threat model names what stays unguarded
LeanFrontier is on Lean/Mathlib v4.35.0 (#502, fdba09a). The weekly upgrade merged unattended, as the v4.34 one did in September: the required mathlib-upgrade check rebuilt the fingerprint index and blocked any newly exact duplicate in the accepted corpus, the audit accepted all 305 entrypoints with no collisions and no corpus deprecations, and the kernel recheck and downstream smoke passed. The follow-up is the separation of statement from proof the corpus insists on — ChungErdos keeps only the v4.35 proof, since the v4.34 alternative added for compatibility in #500 can no longer elaborate on main and had drawn an unused-tactic lint (#503, 4a6eaa2). That the change is proof-only is evidence rather than assertion: lake exe frontier-audit reports the same 1,184 declarations with identical canonical types, kinds and axioms before and after, and leanchecker passes on the module.
LeanFrontier: Field Note 25, and the kernel audit becomes a job
LeanFrontier’s Field Note 25, Fixed Before It Broke., tells the v4.35.0 upgrade as one dry run: the audit against v4.35.0-rc4 on 9 October failed on a single line of ChungErdos (MemLp.indicator now takes a NullMeasurableSet), a two-version proof was written a week early so the release could land unattended once Lean v4.35.0 and Mathlib appeared on 12–13 October, and #503 then kept the v4.35 proof alone (#508, fe01993).
contributed to arminbiere/kissat: clause-section parse errors all point at line 1
Opened issue #83 on arminbiere/kissat, reporting that most parse errors in the clause section report line 1 whatever line they are actually on: an x on line 7 of a seven-line DIMACS file comes back as kissat: error: x.cnf:1: parse error: expected digit or '-'. The cause is in parse_dimacs, which counts lines in a local lineno and sets *lineno_ptr = 1 once at the start, copying the local value back only in NONL(...) and on success — every other error path returns its message directly, so the caller prints the initial 1. The report proposes a sibling macro of NONL that assigns the counter before returning, notes that CaDiCaL 3.0.1 reports the right line on the same input, and records that it was reproduced by hand against kissat 4.0.4 and current master by an LLM-assisted analysis. It was found while testing a DIMACS parser against the SATLIB uf* files that end with %, which is also how cdclkit closed the same roadmap item without reporting the % upstream.
forgepulse: a restart no longer leaves a sync run marked as still running
forgepulse’s sync-run history could not survive its own deployment: a sync interrupted by a restart stayed running for good, so the table advertised a job that had neither finished nor failed. A new fail_interrupted_sync_runs() runs at startup, marks every such row failed with interrupted by restart and a finish time, and the test that covers it also asserts that finished runs are left alone and that a second call is a no-op (13e081c). The same round added browser tests for the parts unit tests cannot see — the award badge, the fleet-wide stars feed and the sidebar breakpoint (1f67aea) — a one-command local deploy script with a README line (272af74), and a transitive devalue bumped to 5.9.4 to clear six Dependabot alerts (cc8b94a).
dratify: the Rust checker's memory bound is written down, and Python 3.15 is declared
Two gaps closed in dratify, the in-process DRAT/DRUP proof checker. The Python layer has bounded its literal widths since 0.1.7, but a Rust caller handed an untrusted literal near u32::MAX makes the checker ask for roughly 200 GB and abort — a real property of the API that appeared nowhere in its documentation, so Checker now says so and names the bound the Python layer applies (85ddd58). The package also declares the Python 3.15 classifier, 3.15.0 being final and CI having tested it since the release candidates; it stays on main until the next release rather than warranting a version of its own (f4f1ccb).
diaz-modulus-lean and prove2me-logs: note v1.27 ranks chapter 14, and v1.28 adds the cyclic case
The Diaz companion note went out twice on 15 October. v1.27 adds no result and corrects three attributions found by two blind referee reports: the step “a 2×3 configuration in a four-dimensional space is a geometric progression” inside Theorem 5.5 is Bachoc–Serra–Zémor (2017), Lemma 4; Fischler’s Lemma 6.1 excludes four-term progressions from the logarithms of algebraic numbers with the six exponentials theorem, while Diaz’s Théorème 7(2) is its form for ℒ̃, two passages having equated them; and the u³ limit holds for single powers through their hulls, where “nothing beyond” overstated it (e3bb5c0, which also ranked chapter 14 tier 1 on a fixed rubric and credited proofs by amorphic and curiyu).
diaz-modulus-lean: note v1.26 answers the two questions Proposition 5.11 left open
The Diaz companion note is at v1.26, which answers the two questions Proposition 5.11 left open: whether every conjugation-stable space invisible to Theorem 5.5 has a configuration factor of degree one, and which of the real pencil spaces W(λ, f) Theorem 5.5 can see. diaz-modulus-lean ports the eleven new results (milestones 127–133) and the mirrored project now holds all 414 proved results on Lean’s three standard axioms (d749dbb, 46 files and about 4,700 lines).
cdclkit 0.1.6: CDCLKIT_ENGINE now does what the documentation said it did
cdclkit is tagged v0.1.6, and the fix is a variable that never existed as documentation claimed (35e9b14). AGENTS.md and the cdclkit.native docstring both said CDCLKIT_ENGINE selected the engine and that the native engine was never chosen implicitly; nothing read the variable, and solve --adaptive and solve -j N used native whenever the build was installed. Both paths now honour CDCLKIT_ENGINE=python|native, print c engine: … so a result says which engine produced it, and exit 1 with a message on an unknown value or on native without the build, where the silent fallback used to be. Defaults are unchanged — native when installed for --adaptive and -j, Python for plain solve — and the engines are bit-exact, so the setting moves the speed and not the answer.
hrodrig stars roundstorm
hrodrig — Hermes Rodríguez, who works on infrastructure (Kubernetes, Go, Rust, Postgres) from Caracas and writes at hermesrodriguez.com — put the first star on roundstorm. The login is not a stranger to this account: he maintains gghstats, the self-hosted GitHub traffic-stats tool that two pull requests went into in August, selectable clone statistics first among them. A star from someone who reads traffic and sizing tools closely is the useful kind of approval.
aristowrap 0.1.1: the returned archive is treated as hostile, and the tests cover everything
aristowrap — the Docker-first wrapper that submits a Lean 4 project to Harmonic’s Aristotle and typechecks the archive it returns — shipped 0.1.1, a maintenance round that starts from the archive being untrusted input: verify applies the tar data filter on every interpreter that has it (3.11.4+, not only 3.12+) and refuses to extract at all on one that does not, reporting refused members as “unsafe archive rejected”; submit no longer uploads symlinked .lean files, which could point outside the project; and a prompt starting with - is no longer parsed as an aristotle option, because a -- separator now precedes it (9b0cd7c). A missing aristotle or lake gives a one-line error and exit 127 instead of a traceback, Ctrl-C stops the child and exits 130, --destination with --no-wait is rejected rather than silently ignored, and a warning goes to stderr when ARISTOTLE_API_KEY is unset, where aristotle itself still exits 0.
prove2me-logs: thirty-seven paper theorems accepted first time, and four submitted twice
The 12 October round on the prove2me-logs
board: 171 missions had appeared since the last survey, 84 roots were untouched
once their creators’ own work is set aside, four reviewers rated them — counting
the in-house kits built the day before as available alongside Mathlib — at 8
easy, 29 medium, 47 hard and none suspect, and nine provers took the 37 easy and
medium ones. All 37 were accepted on first submission, from online learning
(twelve theorems off one textbook) and Nesterov’s ODE and accelerated methods to
robust MDPs, stochastic dual dynamic integer programming, regularised learning in
games, customer assortment, phase retrieval, reaction-network combinatorics,
ring-LWE ideals and approximate Nash equilibria; several used the previous day’s
kits, the semidefinite reformulation of robust MDPs needing the S-lemma with
Slater’s condition and the SDDiP envelope theorem needing LP duality
(a14f5be). The 47 left
open wait mostly on Brouwer’s fixed-point theorem, spherical harmonics,
Golden–Thompson, Kolmogorov–Chentsov, the Moreau envelope or long probabilistic
arguments, and one is equivalent to the Riemann Hypothesis by its own account. A
fifth kit closed the two distributionally robust chance-constraint theorems —
DRJointCC.Individual.theorem_2_2 and DRJointCC.Joint.theorem_3_6 — by moment
duality, taking worst-case expectation over all distributions with a given mean
and covariance, dual attainment under Slater’s condition shown with explicit
discrete measures, plus CVaR against VaR and an affine S-lemma. The round also
carries a mistake: four of the 37 went out twice, because a replaced submission
watcher survived its kill and for an hour two ran side by side, each seeing the
node still open while the other’s verdict was pending — the duplicates were
accepted and change nothing, and the submission script now refuses whenever the
account already has an accepted or pending submission on the node.
jmikedupont2 stars aristowrap
jmikedupont2 — Mike DuPont, a full-stack developer working on peer-to-peer, ML and AI tooling, and a member of the meta-introspector organisation whose fork of aristowrap appeared the same day — starred aristowrap. The star and the fork landing together is the point: the org that keeps its own Aristotle and Lean-worker tooling put a copy of this wrapper on its list and a star on it in one day.
aristowrap was forked by meta-introspector
carlok/aristowrap — the Docker-first wrapper that submits a Lean 4 project to Harmonic’s Aristotle and typechecks the returned archive, so no elan, Lake or Python is needed on the host — was forked by meta-introspector, whose fork is here. meta-introspector builds its own Aristotle and Lean tooling around the same problem — an Aristotle CLI in Rust, lean-worker and a registry for Lean workers — so the fork arrives from someone already working in the space. No changes have come back from it.
roundstorm: a room's log dies with the room, and unknown cost stays unknown
roundstorm closed five issues in a day
of follow-up work on the local-first deliberation app. The events table was
never pruned, so a cleared or deleted room’s turns outlived it — the log is a
debugging aid, not an audit trail past the room’s life, so a delete now takes it
with it, a decision written down in db.ts and docs/design.md rather than left
implicit (#43). Cost had the
opposite problem: only the Claude adapter reports one, other brains stored
NULL, and the persona lab and the headless output summed that as 0, so a
comparison between a priced and an unpriced arm read as a saving; cost is now a
null-aware sum with a priced-turn count, rendered n/a or “at least” when only
part of the total is known (#44).
prove2me-logs: Opial's lemma, a duality kit, and all six QCQP tightness theorems
The day’s pattern on the prove2me-logs
board was to build a missing piece of library once and then use it as a black
box. Weak convergence of splitting methods always ends with a Fejér-type energy
estimate and Opial’s lemma, and Mathlib has neither Opial’s lemma nor sequential
weak compactness of bounded sets in a general real Hilbert space: one file now
proves bounded sequences have weakly convergent subsequences (sequential
Banach–Alaoglu on the separable closed span, then Riesz), weak limits are
unique, Opial’s lemma, and the easy direction of Minty’s theorem, and four of
the five waiting roots were accepted on first submission —
FRBSplitting.Weak.theorem_2_5, FRBSplitting.Inertial.theorem_4_3,
FRBSplitting.ThreeOp.theorem_5_2, ProjReflGrad.Weak.theorem_3_2. The fifth
stays open for a reason in its statement rather than in the library: one branch
of the formalised adaptive run may shrink the step by any factor in (0, 1]
infinitely often, and nothing then forces weak cluster points into the solution
set (5d4a758).
hermes-podman: the porting guide for a host that already ran Hermes natively
hermes-podman now carries the path it
was missing: the rest of the repository assumes a fresh VM, and
docs/PORTING-native-install.md
covers the other arrival — a host that already ran Hermes as a git install, with
~/.hermes and a systemd unit, moving into the pinned container
(#2,
6e4bba1). Everything
in it comes from a port actually done against this deployment rather than from
reading the code: what travels (the secrets, state.db moved with SQLite’s
backup API, and repositories as a bind mount at the same path so absolute cron
workdirs keep resolving, where about 5 of the 2.7 GB home is host-only and must
not be copied), the HOME layout traps — the gateway runs with HOME=/opt/data
while interactive sessions get /opt/data/home, both real, so scripts and
prompts silently read the wrong ~/.hermes or ~/projects until two symlink
shims fix it without editing them — and the tooling the image lacks (gh, jq,
tirith, himalaya, sqlite3) with a /opt/data/.local/bin drop-in that keeps
gh- and jq-based jobs working without touching the image.
diaz-modulus-lean: every involution of the u-line, and a conjugation-stable space with no progression
Two research sprints landed in
diaz-modulus-lean, each
formalised the same day it was answered and each published as a companion-note
version. R9 extends the two-pole family to every Möbius involution of the
u-line: for u ∉ Q̄ and the orbit pencil q₁, q₂ of an involution, the space
H₀ + Σᵢ Q̄ q₁(u)ⁱq₂(u)³⁻ⁱ/(u l₁(u) l₂(u)) carries a 2×3 configuration — the
progression with ratio q₁(u)/q₂(u) — and contains none of u², u⁻² or
1/(u − b), so no four-dimensional space of Theorem B inside it carries one.
It is Theorem 5.9 of
note v1.23,
four nodes and 393 results
(8336e66).
diaz-modulus-lean: note v1.25 classifies the progression spaces, and the library reaches 403
The Diaz companion note is at
v1.25,
which makes the classification of Remark 5.10 formal for a ratio rational in
u: a five-dimensional space spanned by 1, u, u⁻¹ and the four powers
c·hᵏ of a progression ratio h = p(u)/q(u) either contains u², contains
u⁻², contains some (u − b)⁻¹ with b algebraic and non-zero, or is an
involution-pencil space
(8341107). Two
steps carry it. The ratio’s polynomials have degree at most two — an
embedding Q̄(X) → ℂ at X ↦ u, coprimality forcing p ∣ num f and
q ∣ den f, and Mathlib’s Lüroth degree formula [K(X) : K(f)] = max(deg num f,
deg den f) — and a degree of exactly two makes the space an involution-pencil
space, c hᵏ = κ pᵏ q³⁻ᵏ / (u l₁ l₂) for two pencil members
lᵢ = aᵢp + bᵢq. Three new formal results, milestone 126, and the mirrored
project now holds all 403 proved results on Lean’s three standard axioms; the
progression step (Bachoc–Serra–Zémor, Lemma 4) and the rationality of h stay
in the text rather than in Lean, and the classification was not found in the
sources read
(f8102b4).
prove2me-logs: forty-two paper theorems in one night, all accepted first time
A day after the first round, the prove2me-logs board recorded a second: 200 more single-theorem paper missions had appeared, mikedeng1 had published 150 that nobody had touched and Mazecto two more, and over a field of 152 roots with 265 definitions behind them six reviewers read every statement together with its definitions — 5 easy, 37 medium, 110 hard, none suspect — before six provers took the 42 easy and medium ones, grouped by paper so that one descent lemma or one cone lemma served a whole family, and submitted them one at a time with a written explanation of the argument (c4422cd). All 42 were accepted, each on its first submission, from accelerated methods and their ODEs through splitting and constraint qualifications to pricing, online learning, matrix relaxations and integer programming.
lspace-det-sigma: u = g for the Baker–Kegel and Himeno families, and a determinant that does not fix the signature
New note k32 in lspace-det-sigma
proves u = g₄ = (m n² − n + 2)/2 for the closed 2n-braids H(n,m) = X_n^m T_n,
which contain the Baker–Kegel knots at n = 2 and Himeno’s knots at m = 3: the
lower bound is slice-Bennequin, the upper bound an explicit unknotting —
palindromic copies of X_n cancel in pairs, half of the last copy and one band
of T_n are flipped, and the leftover strands lie under everything — checked by
computer for n ≤ 12, m ≤ 7, and the same argument covers the non-positive
L-space (2,q)-cables of T(2,2k+1)
(162c38f). The
mirror of 12n_642 is fibered and strongly quasipositive with g = 2 and
u ≥ 3, so the L-space hypothesis is doing work; the literature check that
found it credited the example to Bode and Truol and the surrounding bounds to
Stoimenow, Kegel et al., Lidman–Moore, Baker–Moore, Hedden,
Borodzik–Liu–Zemke and Teragaito, and showed g = g₄ for the whole family, not
only its L-space members
(71f08cb).
hermes-podman is public: the headless Hermes Agent gateway on rootless Podman
hermes-podman is now public: the
deployment behind this site’s own agent gateway, written down as a repository
instead of shell history. A Hermes Agent container
(nousresearch/hermes-agent:v0.21.6,
pinned by digest as well as tag) runs rootless under Podman and podman-compose
on an Ubuntu 26.04 host — as the normal user, with userns keep-id so data/
stays owned by that user, in gateway run mode only: no dashboard, no published
ports, no browser tools, everything mutable in one BASE_DIR/data mounted at
/opt/data (698f8af).
A Makefile and scripts cover setup, up/down, verify, logs, chat, backup,
restore, upgrade and uninstall, a Quadlet unit is offered as the compose
alternative, CI exercises the whole lifecycle, and
docs/DECISIONS.md
records why each of those choices was made.
deck-lovers: the CI token is scoped to read-only contents
deck-lovers’ CI workflow declared no
permissions block, so every run inherited the repository’s default token
scope — read/write on the repository’s contents, for a job that only checks out
the code and runs the tests. It now asks for contents: read
(9c30dae), the same
scoping forgepulse got after a CodeQL
alert on its workflow.
prove2me-logs: the false statement is notified, and the node now reads Disproved
Yesterday’s round
ended with a kernel-checked negation of SymPolyOpt.PowerSumUB.theorem_6_7
recorded in
prove2me-logs
and nothing said on the platform about it: the
node was
still Open, with no votes, no submissions, an empty mission discussion, and
no trace of the theorem’s name anywhere on GitHub. The two accounts on its
audit trail had vouched for the statement when it was approved, which is not
the same as knowing it is false.
roundstorm: a per-launch token, four cleared advisories, and a Tauri shell that compiles in CI
roundstorm’s desktop app now generates
32 random bytes on every launch, hands them to the daemon as ROUNDSTORM_TOKEN
and injects them into its own page, and the daemon refuses /api and /ws
without it — unset, the headless CLI and every script behave exactly as before.
It is narrower than “auth”, and SECURITY.md now says so: it stops another
process on the loopback socket and another user on a shared machine, and makes
ROUNDSTORM_HOST=0.0.0.0 defensible, but not malware running as the user
(17618e7).
prove2me-logs: nineteen paper theorems in one evening, and one that is false
The prove2me-logs board recorded a round on the single-theorem paper missions published on 9 October: 78 statements read, 20 attempted, 19 accepted on first submission — and one statement shown false as published, with the negation checked in Lean (289fae1). The false one is the interesting entry: the pipeline’s job is to machine-check what a paper claims, and a paper claim refuted by a kernel-checked negation is the outcome that says the checking is real rather than a rubber stamp.
LeanFrontier: a weekly screen that every accepted statement can be stated as written
Four silent bugs in how LeanFrontier’s
receiver reads a theorem out of its source were found in one week — two of them
by accident, while writing field notes — so the one-off corpus screen is now a
guard: tools/screen_statements.py states every accepted entrypoint with
sorry in exactly the probe’s context and classifies each result as stated,
own definition (the statement mentions a name its own module defines, which
the probe deliberately cannot see) or other, and any other fails the job.
.github/workflows/screen-statements.yml builds the corpus in the validator
image, screens it offline with --network none --read-only --cap-drop ALL, and
runs weekly as well as on the relevant pull requests
(#498).
diaz-modulus-lean: note v1.22, two poles, and Kirby's weak Schanuel conjecture at a candidate
The Diaz companion note is now at
v1.22:
diaz-modulus-lean ports the seven
nodes published on Prove2Me on 9 October and the mirrored library stands at all
389 proved results on Lean’s three standard axioms
(3450aec). For
u ∉ ℚ̄ with uū algebraic and distinct non-zero algebraic a₁, a₂, the space
H₀ + ℚ̄·u/(u²−a₁) + ℚ̄·u/(u²−a₂) carries a rank-one 2×3 configuration — the
progression b, bu², bu⁴, bu⁶ — that no four-dimensional H₀ + ℚ̄w inside it
carries, with no conjugation condition; at a candidate, under Roy’s theorem, the
two terms are not both in ℒ̃, and in the family aᵢāᵢ = |u|⁴ their sum is not
(substitutions in Diaz 2007, Théorème 7(2) = Fischler 2001, Lemma 6.1, not
claimed). Under the case n = 2 of Kirby’s weak Schanuel conjecture every
candidate has Im u ∈ πℚ, so Diaz’s conjecture is equivalent to the single
relation t² + π² transcendental for real t ≠ 0 with eᵗ algebraic — the
remark is now formal rather than prose. Chapter 13 of the blueprint gained the
three results from the manuscript, the pair dichotomy, two candidates on an
axis-parallel line and mixed rigidity, with the four nodes their proofs import
(68a87a7). The same
day’s prove2me-logs
entry carries the mission from the research side.
diaz-modulus-lean: note v1.21 mirrors eighteen nodes on separation, the normal form and Laurent hulls
The Diaz companion note is now at v1.21: diaz-modulus-lean ports the eighteen nodes of the polar-degree working note, published on Prove2Me on 8 October, and the mirrored library stands at all 382 proved results on Lean’s three standard axioms (99c0a11). The batch proves separation over any field — numbers algebraically independent over F never enter a p × q configuration with p, q ≥ 2 inside V₀ + Kw₁ + ⋯ + Kw_m with V₀ ⊆ F — alongside linear Cauchy–Davenport in K[X] and the bound p + q ≤ dim V₀ + 1 in K(u); the normal form of 2×2 configurations near a point of a circle, with the invertible matrix of constant terms and, at a candidate, an entry outside the ℚ̄-span of the logarithms in every row and column (this one uses Baker); and configurations inside Laurent hulls, where the sumset criterion, the pairs {0, ±1, ±k, ±l} and the powers u^(4ʲ) mark what the hulls can and cannot exclude — the power pairs excluded at a candidate are substitutions in Diaz 2007. The blueprint gained its chapters 10 and 14 and now lists 157 results, and the statement thm:novan assumes only u ∉ K, as the Lean does. The same day’s prove2me-logs entry carries the mission from the research side (7ca8c55).
deck-lovers: the projector disconnect survives cancellation, and the deploy script reads .env
deck-lovers’ /ws disconnect cleanup
ran unshielded, so a cancelled task could skip telling the audience the
projector had left and leave stale clients behind — Starlette’s TestClient
cancels the app task right after sending the disconnect, which made the
websocket tests hang about four runs in ten. The cleanup is now wrapped in a
shielded anyio cancel scope, two test races are gone, and a regression test
cancels the handler the way the client does
(3c4e938).
cdclkit: the trail copy stops going quadratic on decision-only searches
cdclkit’s target-phase search copied the whole trail into the target at every new deepest trail. The comment said improvements become rare quickly, which is true once conflicts start and false before: a search that makes many decisions without one improves on every decision, so n decisions copied 1, 2, …, n entries — 65,536 declared variables and one unit clause took 144 s in Python, 131,072 took 4.2 s natively, and each 4× in variables cost about 13–19× (e43a4ac). A counter now records how much of the trail the target already holds, only the new suffix is copied, and the target still ends up byte-for-byte what the full copy produced.
caciarabot: telegram-shaped photo validation, enforced rule priority, and a lint-found dangling task
caciarabot’s validator already caught a
file over Telegram’s upload ceiling, but size is not the only way Telegram
refuses a photo: width and height above 10000 px, or a longer side more than 20×
the shorter, fails with PHOTO_INVALID_DIMENSIONS even at a few hundred KB —
and a panorama or tall screenshot is exactly the file people drop in. A new
telegram/imagesize.py reads the dimensions out of the header with the standard
library alone — PNG (IHDR), JPEG (skip the APPn segments to the
start-of-frame marker) and all three WebP variants — checked against macOS
sips on all 142 real images in media/ with no false positives
(b06e315).
LeanFrontier: the Furstenberg integers sit densely inside the profinite completion
The step Field Note 24 named as its next target has landed: qazW12345 added a module proving that Mathlib’s canonical map from the integers into their additive profinite completion induces exactly the Furstenberg topology and is a dense embedding (#491) — furstenbergTopology_eq_induced_profiniteCompletion and isDenseEmbedding_furstenbergProfiniteMap — so Furstenberg’s evenly spaced topology is the integers as a dense subset of a compact group. The note was revised in place later the same day rather than given a second entry, and now stands at 110 modules and 90 internal import edges (#495). That revision also supersedes the morning’s account of the triviality probes: they were written into a scratch file without their module’s namespace, opens or variables and with Lean’s automatic variables on, so an unknown name silently became a universally quantified one and most statements were probed as a more general claim than the one submitted (#489). Each statement is now probed the way its module writes it, a statement that cannot be stated is reported as “not elaborated” with Lean’s reason, and re-reading all 303 accepted entrypoints leaves every historical rejection sound.
curve-symmetry-lean: the note's Lean port is published as a Prove2Me mission
The formalization behind curve-symmetry-lean is now published on Prove2Me as the mission Sharp symmetry bounds for real algebraic curves — 12 definition files and 208 proved theorems on Lean v4.33.1, with the port, the checks and the upload recorded in verification/PROVE2ME.md and the generator, the comparison and the uploader under scripts/prove2me/ (743870b). The mission was then approved for Prove2Me’s public catalog, its goal and all 41 milestones proved, and the README, STATUS and the verification record carry the approval (4bccf12). The same two commits also rewrote six docstrings that still described results as pending or “not yet” when later modules prove them; no statement changed.
lspace-det-sigma: counting families by knot, and a counterexample that must be hyperbolic of genus at least 3
Three corrections landed in lspace-det-sigma, and all three narrow claims rather than add results. The generated families had been deduplicated by diagram, so their counts were counts of diagrams: recounted by knot type, the 549 twisted torus records are 87 knots and the 14,452 random records are 46 — and every hyperbolic knot in those two families is already a SnapPy census knot, while the 274 one-bridge triples are 195 knots with at least 37 hyperbolic ones outside the census, so that family does extend the evidence (40cd61a). The claim that a counterexample in the sharp case could not be braid positive had no proof behind it and is withdrawn — a non-positive braid word does not make a knot non-braid-positive, and what has content is a set of twelve named knots, three of them provably not braid positive. The note now cites DeYeso’s thin L-space conjecture and Baldwin–Sivek’s Proposition 6.8 — an L-space knot with the Alexander polynomial of T(2,2g+1) is not a satellite — so a counterexample to the sharp case must be hyperbolic of genus at least 3, which is also what replaces the withdrawn property in PROMPTS.md (32f05ff, bd95a55). No violation appears and no conclusion reverses; the evidence beyond the census is thinner than the earlier version said.
LeanFrontier Field Notes 23 and 24: the receiver misreads a case split, and the Furstenberg topology becomes the profinite topology
Field Note 23 — The Receiver Misread It. — covers a submission that carries the Stern–Brocot runs into Mathlib’s own continued fractions, the corpus’s first module to use them, and the bug reading it exposed: the receiver that extracts each theorem’s statement out of its source stopped at the first bar of a match … with case split, read the claim as ending there, and recorded the resulting probe as inconclusive — so the reading was repaired along Lean’s own rule, bars after with belong to the case split and a named argument such as (n := 3) no longer ends a statement (#482), which over the corpus’s 548 statements changes exactly one reading. The note also records the receiver’s rejection collector running its first scheduled week on its own, adding 28 September with 22 accepted validation runs, five over a resource limit, two that did not build and one malformed claim.
diaz-modulus-lean: note v1.20 adds the five-dimensional case Theorem 5.5 left open
Version 1.20 of the Diaz companion note adds the five-dimensional case left open by Theorem 5.5, and with it five new formal results (milestones 108–110), taking the mirrored project to all 364 proved results on Lean’s three standard axioms (c5a6007). Proposition 5.6 settles where the 2×3 configuration lives once a fifth coordinate is allowed: for u ∉ ℚ̄ with ρ = uū algebraic and a ∈ ℚ̄ non-zero, put z = u/(u² − a) when aā ≠ ρ² and z = u/(u² − a)² when aā = ρ²; then W = ℚ̄ + ℚ̄u + ℚ̄ū + ℚ̄z + ℚ̄z̄ carries a configuration but no four-dimensional space ℚ̄ + ℚ̄u + ℚ̄ū + ℚ̄w inside it does, and in the aā = ρ² case the conjugation-stable space ℚ̄ + ℚ̄u + ℚ̄ū + ℚ̄·u/(u² − a) carries none. Remark 5.7 prints the configuration’s shape, attributes its consequences at a candidate — u/(u² − a) ∉ ℒ̃ is Diaz 2004, Théorème 2, and u/(u² − a)² ∉ ℒ̃ is Diaz 2007, Théorème 6(3) — and records that the configuration’s invisibility to the four-dimensional spaces of Theorem 5.5 was not found in the sources read. The blueprint gained a chapter holding the results the earlier ones listed as library-only, drawn as 18 nodes with their Prove2Me pages cached (14b4337, c5b4a67), and the chapter of results not found in the sources read moved to the end (90c96da). The same day’s prove2me-logs entry carries the 6 October mission, five-dimensional spaces carrying configurations Theorem B cannot see (007ec16).
roundstorm: the test suite stops opening the real database and running real brains
A Dependabot pull request that failed one Ubuntu cell with “database is locked” was not the bump’s fault. Importing nearly anything under server/src imports db.ts, which opens a SQLite file at import, and ten test files never set ROUNDSTORM_DATA first — so they opened the default data directory, the real database on a developer’s machine, and on CI one file shared by every parallel test process racing to create and migrate it (44a5bd9). A second bug surfaced in the same empty-HOME experiment: gate.test.ts called /api/bootstrap, which probes every brain, so npm test executed the real agy and cursor-agent CLIs on any machine that had them. scripts/test-setup.mjs is now preloaded into every test process and hands each its own empty data directory rather than relying on the next author to remember, gate.test.ts hits /api/memory instead, and a re-run under an empty HOME passes 196 of 196 while creating nothing there.
diaz-modulus-lean: note v1.18 classifies the four-dimensional extensions with a 2×3 configuration, and v1.19 adds Kirby's weak Schanuel
Version 1.18 of the Diaz companion note records what the 5 October batch machine-checks, keeping the mirrored library at all 359 proved results on Lean’s three standard axioms. The new Theorem 5.5 classifies the four-dimensional extensions that carry a rank-one 2×3 configuration: for u ∉ ℚ̄ with uū algebraic and z outside H₀ = ℚ̄ + ℚ̄u + ℚ̄ū, the space H₀ + ℚ̄z carries one exactly when z ∈ H₀ + ℚ̄w for w = u², w = ū², or w = 1/(u − a) with a algebraic and non-zero (b46b618, milestone 107; 43 theorems audited). With Roy’s strong six exponentials theorem these are Diaz’s own exclusions (2007, Corollaire 5(1) and 5(4)); every such configuration is the geometric progression b, bh, bh², bh³ of Fischler’s Lemma 6.1 and Diaz’s Théorème 7(2); and the classification was not found in the sources read.
consilean: token-free transport across the seed bridge, and a preregistered H3 precision at k
Sprint 6 of consilean measured token-free transport across the seed bridge and merged it: the H4 amendment freezes the match rule before the run, and every table stays on the pinned corpora (38eb8c7, PR #7). A second sprint recorded a preregistered H3 precision at k on the Mathlib v4.28.0 snapshot — the candidate set fixed before the run, with the neighbor-recall table kept in place (2a751b5, PR #8).
roundstorm 0.1.0: a local-first deliberation app, released publicly
roundstorm is now public at v0.1.0: a local-first app where persistent, heterogeneous AI researchers deliberate for a set number of rounds on a hard question — you write the question once, they do the arguing. The download is a macOS Apple Silicon image, with the daemon runnable on Linux and Windows; the app bundles no models and no keys, needs Node 22.16 or newer and at least one installed, signed-in CLI brain (claude, codex, agy or cursor-agent), and the release notes state the limits rather than let them be discovered — not notarised, no authentication (the daemon listens on loopback and refuses website origins, which is not auth), advisory capability tiers, and Windows never actually run. The release-day work was a pre-public tune: CI, a real CSP, and a Node floor corrected to a 22.16 that CI measured rather than guessed (4c98425, 73568f8), the daemon moved to Express 5, CSS imports declared so the project type-checks under TypeScript 7, dependency updates taken where safe with Tauri held at 2.11 on purpose, and one open advisory recorded with why it does not reach the shipped app.
prove2me-logs: the 4 October entry, and a correction on Baker's theorem
The prove2me-logs mission journal carries 4 October from the research side — the barrier on generic data, the dilogarithm dichotomy and the three manuscript results that the same day’s companion note v1.17 machine-checks — and the eight nodes contributed by nickrobbins95, mirrored with credit (fc34d3c). The 2 October Baker entry was corrected rather than extended: M. Karatarakis’s public Lean branch baker (github.com/mkaratarakis/mathlib4, since 25 September) formalises the same DALAG Ch. 4 route with the same step-5 repair, so the proof the entry had recorded as a first is not one — the 1 October survey missed it because code search skips forks (90fedab).
lspace-det-sigma: an L-space knot inequality, a data policy, and a Lean 4 lattice lemma
lspace-det-sigma asks whether det(K) ≤ 1 + |σ(K)| for every L-space knot — a conjecture found by a program that fits linear inequalities to a table of knot invariants and discards what its own checks refute, not a theorem except where stated. The note proves it for every iterated torus knot, hence every algebraic knot, by induction over cabling with Litherland’s formula and the bound |σ(T(p,q))| ≥ g(T(p,q)), proved here from the lattice count for torus knot signatures; the combinatorial core of that lemma — N_< ≤ (p−1)(q−1)/8 for 2 ≤ p < q — is formalised in Lean 4 with no sorry, on propext, Classical.choice and Quot.sound alone. The hyperbolic case, where the inequality is sharp, stays open. KnotInfo is cited as its maintainers ask — cited, not copied, since a copy of the database goes out of date — at the two places the note relies on it, and DATA.md and the README give that reason for its absence (dfec446); a later commit sharpens the same file without quoting the correspondence (cc42290).
LeanFrontier Field Note 22: the Stern–Brocot runs become Euclid's algorithm
Field Note 22 — Written Two Hours Apart — covers three submissions in four days, two of which iterate a result the corpus already had: an accepted module showed that the first run of a Stern–Brocot path records the first quotient of Euclid’s algorithm on the fraction the path reaches, and on Saturday night qazW12345 encoded a whole path as its list of runs, provably expandable back (#466), then proved the theorem that first step pointed at — the full list of quotients Euclid’s algorithm produces is the list of run lengths with one added to the last (#467). Both start from the same commit and neither could import the other, so each writes its own run-peeling recursion with the same termination argument, and the duplication can only be closed by a later submission proving the two agree; the note also records that #466’s module built in the fork but failed in the receiver — a missing proof that each path shortens, and an empty-path equation that did not hold by computation — and was fixed without changing any statement. The corpus stands at 107 modules and 86 internal import edges, and the note’s other section, circles given centres, is the curvature-centre tangency bridge already covered. The same day’s maintenance: #470 moved every action off the Node 20 runtime, #474 names the runner image instead of following ubuntu-latest ahead of October’s Ubuntu 26 migration, #475 fixed the regex the receiver and the catalogue share so pattern-matching, dotted and primed theorem shapes are read — the catalogue now cards all 299 entrypoints, including the open Markov uniqueness conjecture — and #477 mints App tokens with client-id.
dratify 0.1.7 fixes a soundness hole, and says why PySAT's CaDiCaL proofs fail
dratify 0.1.6 fixed the README’s PySAT example — it loaded formulas with PySAT’s own parser, which stops at the % line every SATLIB file ends with — and closed a release path that let workflow_dispatch publish from a branch, with the pypi and crates environments now restricted to v* tags (108dd50, 5d0bb47). 0.1.7 is the security fix: three checker bugs found while tracing why PySAT’s CaDiCaL proofs do not verify, one of them a soundness hole where the pure-Python checker accepted refutations of satisfiable formulas when a list of steps held a negative literal — reachable through check_proof, though not from cdclkit’s own solver, and text proofs were never affected — alongside an unbounded literal that could size arrays for 10¹¹ variables and two false rejections, both checkers refusing valid proofs in which a lemma arrived unit at the root (61165f9). The PySAT failure itself is not the checker’s: the binding reads every CaDiCaL proof before flushing it, and with the documented workaround all four CaDiCaL versions verify 50 of 50, where glucose and Lingeling already did (2823854).
diaz-modulus-lean: note v1.17 — a dilogarithm dichotomy, a barrier on generic data, three manuscript results
Version 1.17 of the Diaz companion note records what the 4 October batch machine-checks, taking the mirrored library to all 346 proved results on Lean’s three standard axioms (f91a531). Proposition 3.9 and Corollary 3.10 turn a rational relation a·t² + b·π² into a dichotomy: Li₂(1/2) = π²/12 − (log 2)²/2 is irrational, or e^(iγ/π) is transcendental for every rational γ ≠ 0 — both alternatives open, and the proposition is Brownawell’s Corollary 5 (1974) in general form. Theorem 5.4 is a barrier on generic data: no rank-one 2×3 configuration has its products in ℚ̄ + ℚ̄u + ℚ̄ū + Σℚ̄w_j, so Roy’s strong six exponentials theorem cannot refute a candidate on generic data, whatever logarithms are added. Corollaries 6.11 and 6.12 and Theorem 6.13 are Theorems 2.5, 2.3 and 3.9 of the manuscript, as direct instances of Waldschmidt’s 1973 Corollaire 4; Appendix A adds six rows (121 identifiers checked against the platform, 0 mismatches) and the README now credits M. Karatarakis’s Lean formalisation of Baker’s theorem, which the 1 October survey missed (5f940ac). The same day the library mirrored eight Diaz nodes contributed on Prove2Me by nickrobbins95, each module header naming the author (354 of 354, f0fde04).
cdclkit 0.1.4 requires the dratify that checks proofs correctly
cdclkit 0.1.4 moves both halves onto dratify>=0.1.7, whose three fixed checker bugs include two that reach this package: --self-check could reject a valid proof, and the pure-Python checker could accept a refutation through check_proof (d33f84a). Publishing is now tag-only — the environment restricted to v* tags, the workflow_dispatch trigger gone, and the version check failing outright on any other ref — and make smoke, broken since the dratify split, now ties its module list to the source tree, with a packaging CI job that builds the wheel, installs it cleanly and runs it (b7b5426). Python 3.15 rc3 joins the test matrix (including a native-engine leg) and the README example is now executed by a test; the runbook’s recorded Sigstore gap was already covered, since every file on PyPI has carried a PEP 740 attestation since the first Trusted Publishing release.
caciarabot: the weekend Wikipedia draws get their own prompt pool
The weekend digest swapped the tech feeds for a random Wikipedia article but kept the tech prompts, which tell the model it reads “computer-science-adjacent feeds” and must say something “technically true and not obvious” — handed a 150-character stub, that forces either an invented fact or a bolted-on code joke, and a county sheriff’s office had drawn a package.json comparison. A separate config/prompts/digest_weekend/ pool now holds three tones mirroring the tech one, confined to the excerpt and forbidden from asserting facts from memory or reaching for code-and-servers comparisons; the prompt is chosen by candidate.source rather than by re-checking the weekday, so it always matches its content, an empty weekend pool falls back to the tech pool instead of losing the day, and the validator requires the pool whenever the digest is enabled (932b0a6). A first pass still leaked — the Nagano article drew “visti i risultati complessivi”, implying an outcome the excerpt never stated — so the prompts now also forbid implying outcomes or quality the excerpt does not support.
forgepulse: an award badge for repositories in the top 10 of both rankings
forgepulse publishes two rankings of the
same fleet — human attention, built from unique views, external referrers and
new stars and forks, and raw clone volume — and they rarely agree, so a
repository that reaches the top ten of both now carries an award badge beside
its rank, in either view
(55028f2). The condition
lives in web/src/lib/rank.ts as inBothTopN, which is true only when the
attention rank exists and both ranks are 10 or better; the badge is styled in
styles.css and the rule is covered by rank.test.ts.
diaz-modulus-lean: note v1.16 proves Baker's theorem, and Diaz's (Qr2) in transcendence degree one
The Diaz companion note reached
v1.16, where
Baker’s theorem is
proved rather than assumed: ℚ-independent logarithms of algebraic numbers are
linearly independent over ℚ̄, along the Bertrand–Masser route through the
Schneider–Lang criterion for ℂ^{d₀} × (ℂ^×)^{d₁} with d₀ <= 1 and a Schwarz
lemma for Cartesian products — so the four appendix rows that had read “Proved,
assuming Baker” now name unconditional forms. A new subsection, “Products on the
axes”, adds Diaz’s conjecture (Qr2) of 2007 in transcendence degree one
(Proposition 6.11, Theorem 6.12, Corollary 6.13), by running Diaz’s own argument
with the four exponentials theorem in degree one in place of the conjecture, and
the title of Brownawell’s 1974 paper is corrected. The appendix was checked
against the live board (115 identifiers, no mismatch) and the library now holds
all 340 proved results on Lean’s three standard axioms;
prove2me-logs carries the same day from
the research side, down to Baker’s theorem as a hypothesis discharged
(8c8d14b). The same
batches published the
blueprint site, which states each
classical theorem with a dependency graph, links to the declaring line of the
Lean source and a PDF
(af4f944), and
refuses to build an incomplete site
(e807733).
caciarabot: weekends take a non-technical turn in the digest
Weekend days — in the bot’s own timezone, default Europe/Rome — now pull the
daily digest’s candidate from Wikipedia under a non-technical topic filter
instead of the configured tech sources, and the daily-thought Wikipedia rabbit
hole uses the same filter on those days; weekdays are unchanged, and weekday
Wikipedia draws stay unfiltered (00e467f,
PR #1). is_weekend_in_bot_timezone()
in llm/scheduler.py is the shared check, looks_technical() in
llm/wikipedia.py the heuristic, and weekend Wikipedia picks skip the
English-only page fetch so Italian reads come through. The same day’s second
change fixes how updates land: deploy/update.sh now rebuilds, then does
podman compose down and up -d, because up -d followed by restart could
leave the old container running the previous image; the script deliberately
keeps -v off down, so the bind-mounted config, media and data/ trees
survive (bbaa82c,
PR #2). The test suite stands at
174 passing.
LeanFrontier: a curvature-center tangency bridge from an outside contributor
The fork of
qazW12345 — whose fork was the first thing I
wrote about in September — landed another slice of the corpus’s Direction E
roadmap: a new CurvatureCenter.Circle carrying a curvature plus a Euclidean
centre in ℂ, a bendCenter, lossless conversions to and from
EuclideanGeometry.Sphere ℂ, an IsExternallyTangent equation in
reciprocal-curvature coordinates equivalent to Mathlib’s Sphere.IsExtTangent,
and a Ford-circle specialization that recovers the accepted Farey-neighbour
cross-determinant criterion. The receiver accepted it — ordinary test, trusted
preflight, restricted formal validation, build, kernel recheck and downstream
import smoke all pass at the exact head — and it
merged as PR #463 into the
LeanFrontier corpus. No new
mathematics is claimed: the module deliberately stops before any claim that the
algebraic Descartes Vieta reflection equals an inversive reflection, which the
roadmap requires a geometric configuration for first.
diaz-modulus-lean: notes 1.13 to 1.15 — Roy's strong six exponentials, and the route that stops at u³
Three more companion-note versions landed on 1 October, taking the mirrored
library from 283 to 307 results. The
companion note reached v1.13
(293 results) with the consequences of Roy’s strong six exponentials theorem
that are now machine-checked — Diaz’s Corollaires 1, 2, 4 and 5 of the 2007
paper in general, and at a candidate the facts that u³ and axis multiples
leave ℒ̃, so e^(βπu) is transcendental
(f3c56f5).
v1.14
(304) then takes all of Section 2 of Diaz 2007 under the same hypothesis, and
pins down where that route stops: a strong six exponentials configuration fed
with a candidate’s own data exists only for k = 2 and 3, so the theorem
excludes u² and u³ from ℒ̃ and no higher power
(2c75c4b).
The day closes at
v1.15
(307), where the strong four exponentials conjecture plus Baker’s theorem
implies the sharp four and that implies Waldschmidt’s strong five, and
Waldschmidt’s 1988 remark is corrected to compare the strong five with the
sharp four rather than with Conjecture 1.2
(631d51f). No new
mathematics is claimed, and
prove2me-logs carries the same day
from the research side, down to a note that Brownawell 1974 is “related by the
exponential function”
(d89c37f).
LeanFrontier Field Notes 20 and 21: the diagnostic was the fix, and two measures of building on the corpus
Field Note 20 — The Message Was the Fix — follows the Varignon perimeter submission that had been rejected six times: after the diagnostic was rewritten to say which term it measures, how large the statement is once expanded, and to suggest naming a large subexpression with a definition, the next attempt stated the perimeter as a named definition and was accepted. The note also corrects two published numbers (the pre-registered “imports something recent” column was computed over every module first seen that day, so 32 of 96 rows flip from true to false; the new weekly rejection count had mixed seven weeks of runs at once) and explains Monday’s stalled kernel re-check: asked for one module, the checker checks every module whose name begins with that name, and that one shares its name with a folder of nineteen others.
diaz-modulus-lean: Waldschmidt's 1973 theorem is machine-checked, and Schneider's eighth problem with it
The Diaz companion note reached
v1.12:
Waldschmidt’s 1973 theorem in its algebraic-independence form is now
machine-checked — if x₁, x₂ and y₁, y₂ are linearly independent over ℚ and
e^(x₁y₂), e^(x₂y₂) are algebraic, then two of the eight numbers x_i, y_j,
e^(x_iy_j) are algebraically independent
(8b015b3) — formalised
as a one-column extension of the four exponentials development rather than a new
development of its own. That closes what the
previous version had
to carry as a hypothesis: the consequence added in v1.11 is unconditional,
Section 4 no longer says the theorem is not formalised, and the status “Proved,
assuming Waldschmidt 1973” is gone from Appendix A. It brings Schneider’s
eighth problem — at least one of e^e and e^(e²) is transcendental — and the
e^(π²) statement with it. No new mathematics is claimed, the mirrored library
stands at 283 results, and
prove2me-logs carries the same day
from the research side
(22c5195).
curve-symmetry-lean: the coding agents' handoff file leaves the public tree
curve-symmetry-lean no longer carries ONBOARDING.md, the coding agents’ handoff and tracker, removed at the author’s request (d99bc17); it stays in the git history, last version at f618c4a, and the README says so for the older verification notes that cite it. The README now lists the repository’s tags through sprint-4-complete and sprint-7-complete, and STATUS records the Sprint 7 commit’s two CI runs. This is the follow-up to the 30 September publication, not a new result.
Registered at the fourth attempt
On 29 November 2024 Terence Tao conjectured on the Lean Zulip that the only nontrivial one-generated magma satisfying law 1518 of the Equational Theories Project, together with four target laws, is the three-element cyclic shift. I proved it in core Lean this month, in magma-1518-obstruction-lean. Since 30 September it is a registry record: PALOMAR-2026-09-30-000005.
LeanFrontier v0.2.0, and a day of cyclotomic-eight submissions
LeanFrontier shipped
v0.2.0, its first
library release since v0.1.1’s two modules: 102 accepted modules on Lean and
Mathlib v4.34.1 across 15 top-level areas, nine modules deep at its longest
import chain, with the Markov uniqueness conjecture stated formally and one
classical coprimality hypothesis resolved by a formal computation of two rings
of integers. Admission is still mechanical and grew with what went wrong in
practice — kernel re-check by leanchecker, no build- or import-time code,
add-only submissions, and conjectures as a first-class quota-limited kind. The
same day a batch of outside submissions from
@qazW12345 landed: the eighth cyclotomic
field’s Galois group identified as a Klein four group
(#445), the missing Gaussian
quadratic direction added
(#446), and exactly three
quadratic intermediate fields proved, which completes that thread
(#456); topological
transitivity for the full tent map, transported through the accepted Ulam
homeomorphism to the parameter-four logistic map
(#444); the first cofinality
bridge between the corpus’s Furstenberg topology and Mathlib’s generic
profinite-completion indexing category
(#443); and Descartes
curvature reflections bundled into an algebraic action layer
(#457). The evidence ships
beside the code, including the citable
corpus-v1
dataset snapshot; the running record is at the
field notes.
diaz-modulus-lean and prove2me-logs: note 1.11 on rational squared moduli, and the fifth size wave
The Diaz companion note reached
v1.11
with three corollaries on rational squared moduli, each machine-checked: if
t^2 + pi^2 is rational for a real t != 0 with e^t algebraic, then
e^(i*gamma/pi) is transcendental for every rational gamma != 0
(Corollary 3.8), so the two remaining open statements cannot both fail at
rational data; logarithms of algebraic numbers that are algebraic over Q(pi)
with rational squared moduli are rational multiples of one another up to
conjugation (Corollary 6.9); and at most one pair ±t makes t^2 + pi^2
rational, so (log 2)^2 + pi^2 and (log 3)^2 + pi^2 are not both rational
(Corollary 6.10). Section 4 adds a specialisation of Waldschmidt’s 1973 theorem,
checked with that theorem as a hypothesis. The mirrored library stands at 268
results: the fifth size wave
adds six second proofs and one node, closing at 268 of 268, and
prove2me-logs carries the same day
from the research side
(9bed59e,
1c5714c).
curve-symmetry-lean is public: the note and its Lean port
curve-symmetry-lean went public
on 30 September: the note Sharp symmetry bounds for real algebraic curves, as
PDF and TeX,
together with the Lean 4 port that checks every theorem, lemma and remark of it
— 108 modules and about 900 theorems and lemmas, using only propext,
Classical.choice and Quot.sound, with no sorry and no native_decide
(f0b97b7).
COVERAGE.md
maps each claim to its declarations and to the reading it is checked at: for
d >= 5 the note classifies the curves attaining the maximum max(d, 2d-4)
rotations as Re(z^(d-2)(|z|^2 + a)) = 0 up to similarity, with their exact
ambient Möbius groups, and the port reaches the genus from an explicit basis of
holomorphic differentials rather than Riemann–Hurwitz. Theorem 1 alone is also
published in
sharp-symmetry-bounds-lean
and registered with Palomar as PALOMAR-2026-09-18-000007. The last sprints
before publication put the machine-checked halves into the prose — an
Appendix A to the note
saying where each statement is checked, by what, and who checked it — and added
negative controls
that make three mutated copies of the registry entry fail their own pre-checks
for the intended reason. An e-mail to Alcázar, Lávička and Vršek went out on 30
September reporting that the note’s irreducible quintic is a counterexample to
Lemma 9 of their
arXiv:1801.09962v1; no novelty is claimed
anywhere, and
CHECKS.md
records exactly what has and has not been checked.
topshift-trend: a transient failure no longer wipes out what was already delivered
A scheduled notify pass in
topshift-trend can now fail halfway
without repeating itself: per-chat deliveries are recorded as they succeed, so
the next run skips the links a chat already received and retries only the ones
that never went out, and the global cooldown baseline advances only after a
clean pass
(0ce1688,
PR #9). Before this, one
transient Telegram error in the middle of a batch could resend everything that
had already arrived. A ruff UP035 fix — Mapping imported from
collections.abc — rode along.
sharp-symmetry-bounds-lean shows the extremal quintic
sharp-symmetry-bounds-lean
now carries the equality case in the README instead of only in the statement:
a figure and its drawing script
for Re(z^3(|z|^2 + i)) = 0, the d = 5 member of the extremal family, which
has six rotations — the maximum 2d - 4 — plus a view of its complex points.
The script pins numpy and matplotlib inline and runs with
uv run figures/quintic.py. No Lean file, challenge, solution, comparator or
registry configuration changed, so the registered commit is unaffected.
LeanFrontier: an edge aimed at on purpose, and a corpus of 96 modules (Field Note 19)
Field Note 19 —
Aimed at the Number — opens with a submission that arrived ten hours after the
previous note’s day
saying, in writing, that it had been chosen to add an import edge: the thing
the accumulation series counts had become something a producer aims at. The two
facts are about Varignon’s parallelogram’s perimeter — adjacent sides half the
diagonals, so the perimeter is their sum — and the import is genuinely used,
calls the corpus’s Varignon theorem inside the proof. It was then rejected for
RESOURCE_LIMIT_EXCEEDED: the statement, fully expanded, is larger than what the
receiver will fingerprint and probe, and the mathematics was never in question.
The agent answered with five pushes in thirty-five minutes, each smaller, each
refused in the same words — because the diagnostic said the theorem “exceeds the
normalized-term limit”, and term reads as the proof while the receiver only
ever measured the statement
(4c4210f now says which
term it measures, how large it is, and that the proof is not counted). The same
contributor closed the day with the first bridge between the corpus’s
Stern–Brocot tree and the Euclidean algorithm — k identical turns along a path
record a quotient k of the algorithm, or k + 1 if the path ends there — both
theorems stated in terms of SternBrocot.pair, a corpus constant that now
appears in the statements of six submissions, which is the catalogue measure that
cannot be moved cheaply. The corpus stands at 96 modules and 74 internal import
edges.
curve-symmetry-lean closes R01: the quartic's genus three, from an explicit basis of differentials
Roadmap item R01 in curve-symmetry-lean
is closed: the quartic Re(z⁴) = 1 is now proved to have genus three, and the
proof reaches it from an explicit basis of holomorphic differentials rather than
from Riemann–Hurwitz. Three steps, each a clean check.sh run against
mathlib-v4.34.0-reuse: first the chart at infinity for y⁴ = 2 − x⁴, where
f = 2 − x⁴ and g = 2s⁴ − 1 are squarefree of degree four so R01c-2 applies to
both Kummer fields, and F·dx is regular at the place pulled back from (0, ζ)
with ζ⁴ = −1 exactly when φ(F)/s² lies in the local ring
(fdc7192); then the
isotypic split of the holomorphic differentials under y ↦ i·y, which recovers
each aⱼ(x)·yʲ·dx/y³ from F, σF, σ²F, σ³F with coefficients ±1, ±i
(2024297); then the
degree bound deg aⱼ + j <= 1 at infinity, leaving dx/y³, x·dx/y³ and dx/y²
to span the holomorphic space and forcing genus three
(b8c207e). The same
commit settles Remark 5 of the note as printed — Re(z⁴) = 1 has four rotations
and genus three, every curve of the m = 2 family has genus two, and no direct or
opposite similarity carries the quartic onto one of them — and the axiom audit
climbed 1,991 → 2,034 → 2,058 across the three steps. The note and its Lean port
went public in the same week, with the launch described
here.
pickmydegree stops tracking its own generated coverage reports
pickmydegree had been committing the
HTML output of its test-coverage run, and CodeQL flagged the Istanbul report’s
sorter.js assets as js/xss-through-dom inside that generated output
(11e72fc). Those files
are test artifacts rather than application code, so coverage/ is now ignored
and the committed reports left the tree — roughly 22,800 lines of generated HTML
and CSS removed and three added, merged as
PR #2. The app source is
untouched; only the repository stops carrying a copy of its own coverage run.
Contributed to PalomarRegistry/PalomarSubmission: registration stopped creating archive copies
Contributed PalomarRegistry/PalomarSubmission#156:
registration of an accepted submission had been “under way” for more than seven
hours, the status page’s last event being 2026-09-28T08:09:00Z — The submitter
asked for this result to be registered, with no palomar/… tag on the
repository and no PalomarArchive copy created. The issue reports that this is
not local to one submission: the newest repository in the PalomarArchive
organisation was created at 06:50Z and none since, for anyone, while the rest
of the pipeline looked healthy — 24 workflow runs started in PalomarSubmission
since 09:00Z, 20 of them successful.
pacenotch fingerprints credentials by expiry instead of hashing the token
pacenotch notices that Claude Code has
refreshed its OAuth token by fingerprinting the credentials it reads, and that
fingerprint used to be a SHA-256 of the access token — which CodeQL flagged as
go/weak-sensitive-data-hashing
(a63ee1f). The hash never
left memory, but hashing the secret was never needed to detect a refresh: the
fingerprint is now claudeAiOauth.expiresAt, a value every refresh moves and
which is not secret at all, while a PACENOTCH_TOKEN cannot change while the
program runs and gets a fixed fingerprint. It is the second CodeQL-driven
correction in pacenotch’s credential layer after the recovery path stopped
re-reading the token.
magma-1518-obstruction-lean narrows its Palomar package to Theorem A after the registry's automated review
Palomar’s automated review declined registration of
magma-1518-obstruction-lean
on two grounds, and both are now answered. The first was an overstatement: the
abstract, the source record and the Challenge account presented “no finiteness”
and “one target law instead of four” as a strengthening of the cited conjecture,
while the repository’s own prior-work account says the conjecture carried no
finiteness hypothesis — the README and notes had been corrected on 13 September,
the metadata and the Challenge docstring were missed, so the repository
contradicted itself in the two places a reviewer reads first
(5d521ea).
The second was scope: the compared declarations included the F5 and F13 members
of Theorem F, two concrete finite examples rather than the family, and the review
did not find research interest established for that group. The package now
compares Theorem A and Corollary A′ alone, Challenge.lean falls from 178 to
84 lines, the family facts stay audited under their library names, and
gen_palomar.py grew a --with-family flag that reproduces the earlier
ten-declaration package
(af908bb).
magma-1518-obstruction-lean ports every Lean file to the module system for its third Palomar attempt
The second automated Palomar review found no mathematical blocking issue in the
selected statements of
magma-1518-obstruction-lean
and one presentation problem: the README’s Palomar section still described the
ten-declaration package, F5 and F13 members of Theorem F included, while the
submitted Comparator configuration checks only the four Theorem A declarations —
and still said nothing had been submitted. The README now separates what the
configuration checks from what else the repository holds and how each part is
checked, and PALOMAR.md splits its check record by package so the earlier
ten-declaration runs, including the 33-theorem audit and the rc3 Comparator run
made before the narrowing, are no longer read as evidence about the narrowed one
(1794778).
LeanFrontier: the same fact twice, and a lemma of its own (Field Notes 17 and 18)
Two Field Notes landed in two
days. Field Note 17
— “The Same Fact, Twice” — has the Srinivasan collision identity coming back
after the previous note recorded it lost, and then two submissions opened
forty-six seconds apart from the same fork, both proving that −1 is a
square modulo every Markov number: the first in the integers modulo m via
Mathlib’s primitive sums of two squares, the second as the existence of r
with r² ≡ −1 (mod m) via a Bézout lemma the same contributor had landed that
evening. The receiver accepted both, correctly by its own rules — it compares
each statement with Mathlib and the corpus exactly, and two notations for one
fact are two statements to it — so the corpus now holds a result twice, and the
only reason anyone knows is that a person read both. The same note extends the
deepest chain in the corpus to nine modules.
forgepulse: a fleet-wide feed of every star, newest first
forgepulse can now answer “who starred
what, and when” across the whole fleet: a new /stars page lists every star
across every tracked repository in one chronological feed — avatar, who, which
repository, when — which GitHub only offers per repository and never across an
account (af4a8ea). The
backend adds a star_events table populated from each repository’s
/stargazers endpoint requested with the star+json media type, which carries
the true starred_at per user and so needs no backfill wait, unlike the
human-attention baselines; it syncs alongside the existing star count.
diaz-modulus-lean and prove2me-logs: the fourth size wave, and companion note 1.10 withdraws a priority claim
The companion note to diaz-modulus-lean
reached version 1.10
to withdraw a claim the note should not have made: Section 6 called the
transcendence degree two of two unrelated candidates of commensurable moduli
“the first constraint of any kind on such families”, and Diaz 2007, Corollaire
4(2) — a consequence of Roy’s strong six exponentials theorem — already puts one
of u/v, v/u outside the algebraic span of 1 and the logarithms for such a
pair. The sentence now cites it, all 81 identifiers match the platform, and the
mirror holds 262 results
(c04cf7a).
deck-lovers: the projector password stops living in the browser cookie
deck-lovers’ /login handler set the
proj_auth cookie to the projector password itself, so the plaintext password
lived in every presenter’s browser — which CodeQL reports as
py/clear-text-storage-sensitive-data. The server now issues a random
per-process session token and compares password and cookie in constant time,
which also means a server restart invalidates existing projector cookies; two
converter test assertions that matched the bare fonts.googleapis.com hostname
(py/incomplete-url-substring-sanitization) now check the exact stylesheet URL
built by google_fonts_css_url()
(fa4dfce).
consilean is public: scoring Lean statements by similarity and dependency distance
consilean is a new public Python repository that scores pairs of formal statements in Lean corpora on two axes — how similar they are, and how far apart they sit in the dependency graph. Similar statements that sit near each other are duplicate candidates; similar statements that sit far apart are candidate hidden connections, and Lean’s kernel stays the authority on whether a connection holds: a similarity score is a pattern, not a claim.
LeanFrontier: a Markov slope-scale gcd identity from the outside contributor
LeanFrontier merged
#385, a 160-line
LeanFrontier/NumberTheory/MarkovEquation/SlopeScaleGCD.lean
carrying three new declarations, submitted through the fork of
qazW12345 — the outside contributor whose fork
was the intake route for the
Nesbitt submission, and whose resolution of the corpus’s only conjecture
closed Field Note 16 the day
before.
consilean: the second H2 ranking, a read-only H3 replay, and why the frozen deprecations cannot be H3 links
Sprint 2 of consilean — the scorer of Lean
statement pairs by similarity and dependency-graph distance, with the kernel left
as the authority on whether a candidate connection holds — recorded a second H2
ranking and a read-only replay of the H3 axis against Mathlib v4.28.0, with the
run reports generated into docs/
(1592971). One negative
result is recorded rather than buried: the frozen Mathlib deprecation pairs cannot
serve as later H3 links, because every since date is already earlier than the
Mathlib revision that was scored
(39a5a75). The wider
statement reader therefore stays on the frozen deprecation set, and the duplicate
list stays empty — none of the recovered pairs closed on the later revision.
LeanFrontier: the corpus's only open conjecture is closed (Field Note 16)
Field Note 16 — “The Only Open Question Closed” — reports that the corpus’s single conjecture, stated on 22 September as #243 claiming the coprime-differents hypothesis in the discriminant formula for linearly disjoint number fields is load-bearing, was resolved three days later by the project’s external contributor. The witness is the eighth cyclotomic field: its two quadratic subfields generated by ζ + ζ⁷ and ζ − ζ⁷ are the square roots of 2 and of −2, each of discriminant of absolute value 8 and ramified only at 2, so their different ideals share a prime and the formula predicts 8² · 8² = 4096 where the actual discriminant is 256. The resolution arrived twice — a first pull request carrying the field theory alone was closed as superseded by the second, whose 512-line module is mostly the two rings of integers — and six further submissions came with it, each importing an accepted module: Kochen–Stone from Chung–Erdős, two layers of the Furstenberg topology, Stedman’s plain changes closing into a cyclic Gray code, and two on the Markov tree, the Stern–Brocot embedding shown injective on whole triples and the branch that always turns the same way identified with the odd-indexed Fibonacci numbers.
diaz-modulus-lean: companion note 1.9, and three size waves that turn 2,865 lines into 856
The companion note to
diaz-modulus-lean reached
version 1.9,
which checks Appendix A statement by statement against the Lean nodes: four rows
whose proofs take Baker’s theorem on linear forms in logarithms as a hypothesis
are now marked “Proved, assuming Baker” and Section 1 lists that theorem as a
third input used without proof, Corollary 3.4 is marked Not formalised because
no node states it, Proposition 3.2 is restated as
Diaz.quantisation_orbit_iff_re_ne_zero proves it, and all 81 identifiers match
the platform, the mirror holding 254 results.
portcullis: Phase 1 stops accepting any SSH host key
portcullis replaced paramiko’s
AutoAddPolicy — which accepts whatever key a host presents — with an explicit
trust-on-first-use policy (25da71b).
Phase 1 is the first contact with a VM created seconds earlier by the same
process, so there is no prior key to compare against and Hetzner does not
publish the fingerprint through its API: the new policy accepts that first key
once, logs its type and fingerprint, and Phase 2 still connects with that exact
key pinned, so a later substitution is detected. The change addresses CodeQL’s
py/paramiko-missing-host-key-validation alert, and the test that asserted the
blanket policy now asserts the new one. The secret-handling pass of 21
September had already pinned the key Phase
2 connects with; the first contact was still an unconditional accept.
LeanFrontier: seven modules deep, and a five-week-old semiconjugacy finished (Field Note 15)
Field Note 15 is
out, and both submissions of its day were merged without anyone touching a
button. #307 shows that the
three oriented Markov branches the coverage theorem split out are one branch up
to a cyclic permutation of the coordinates, and packages the permutation that
carries the canonical branch onto each of the others — which puts the module at
the end of the corpus’s deepest chain: Markov equation, descent step, path
layer, orientation, Stern–Brocot bridge, coverage, symmetry, seven modules and
six import hops, all six built since Monday.
#308 finishes a result left
five weeks earlier: the Ulam–von Neumann semiconjugacy between the tent map and
the logistic map at parameter four, submitted on 18 August through the change of
variables x ↦ sin (π x / 2)², was missing the structural fact that that change
of variables is a homeomorphism on the unit interval, so supplying it upgrades
the semiconjugacy to a topological conjugacy — the two maps are the same
dynamical system seen through a change of coordinates.
forgepulse: the 1d/7d/30d windows now count back from the newest day GitHub reported
The window columns in forgepulse were
mostly empty at their recent end: GitHub’s traffic API lags a day or more, so
counting back from the wall clock left the newest days without rows and the 1d
column read zero for nearly every repository
(7f5acb4). The windows now
count back from the newest day that at least half of all repositories have a row
for, and each window is capped at that day, so a stray early same-day row cannot
leak in as a partial day or move the anchor — a plain MAX(day) would let one
early reporter blank out everyone else. It also makes 1d genuinely one day: the
old >= now − 1 day spanned up to two. It is the same lag the per-repository
clone baselines were added for, now
fixed at the window itself rather than at the comparison.
diaz-modulus-lean: the companion note reaches 1.8, and Gelfond–Schneider is rebuilt as a tree of eleven nodes
Five more versions of the companion note to
diaz-modulus-lean landed on 24
September, and the mirrored library grew with each.
Version 1.4
machine-checks the three barrier statements version 1.3 had proved on paper:
the period never enters (Theorem 5.6 — for u algebraically independent of
π the only quadratic relation with algebraic coefficients among 1, u, ū, iπ
is the norm X₁X₂ − ρX₀², so a configuration that could detect a candidate uses
the constant term, u and ū and never iπ), and Proposition 5.7 gives the
relations a non-generic candidate can carry. Version
1.5 then
corrected 1.4’s own claim that nothing known excludes a relation such as
Re(u²) = π²: Théorème 0.2 of Roy and Waldschmidt (1997), the quadric version of
the 1973–74 theorem, excludes every rational quadratic relation among u, ū, iπ
for a candidate algebraic over ℚ(π) with Im u ∉ ℚπ, and two more barrier
results followed (Theorems 5.4(c) and 5.6(c)).
caciarabot: the secret feature logs how many people, not which ones
The segreto feature in caciarabot keeps
a roster of who has posted — display names only, no message content, because the
Bot API cannot list a group’s membership — but its dry-run log line wrote those
names out in full. It now logs a count instead
(db905db), and the
README’s Privacy section says so explicitly: the events say how many people a
secret was about, never which ones. The change also cleared CodeQL alert
py/clear-text-logging-sensitive-data, whose “sensitive data” label was firing
on the word “secret” rather than on a credential — the dataflow was real even
where the rule’s name was not, and the count is all the operator needs to see the
feature working.
LeanFrontier: nine submissions, every one built on something the corpus already accepted (Field Note 14)
Field Note 14 is out, and
its day is nine submissions that each import an accepted result rather than starting beside
it: most build on results accepted the day before, one on a module accepted three hours
earlier, and one on the Furstenberg topology another contributor landed on 18 August — so
the note’s own answer does not need a second producer, only an accepted one. The named
work includes explicit Thue–Morse cube sums via Prouhet and Nicomachus, the Markov tree
path layer that makes the descent step iterable, the classical Turán bound derived from
the Caro–Wei bound accepted the previous evening, the Horadam addition formula built on the
companion matrix, Sylvester’s reciprocal series summed to infinity, and Markov root
reachability — every positive Markov triple reached from (1,1,1) by Vieta moves, the
third of three consecutive mornings whose results each became the next one’s substrate.
LeanFrontier: documentation proposals from anyone, a merge queue, and the launcher A/B suspended
The receiver’s own day in LeanFrontier was
governance rather than mathematics. A pull request whose every changed path is a Markdown
file under docs/ is now a documentation proposal
(#291): no claim, no Lean, no build, and
the three stages that build or run candidate code are gated off, so a prose change costs one
short job instead of a Docker build and a Mathlib fetch — decided from the changed paths and
never from the branch name, with docs/catalogue/ and docs/website/ excluded because one
is generated from the corpus and the other publishes under the project’s name. The friction
it removes was real and twice theirs:
#290 carries a contributor’s rewritten
docs/CONTRIBUTION-DIRECTIONS.md with their authorship after the OWNER gate had refused it,
leaving six open directions and marking the ones that have landed.
diaz-modulus-lean: the companion note reaches 1.3, and the mirror closes at 200 of 200
Three versions of the companion note to diaz-modulus-lean
landed in twenty-four hours, and the mirrored library grew with each: 182 of 182
proved results at note-v1.1,
195 of 195 at note-v1.2,
and 200 of 200 at note-v1.3,
with lake build Diaz clean and nothing behind the headline theorems but propext,
Classical.choice and Quot.sound.
LeanFrontier: nineteen submissions in a day, and what trusting a contributor would have skipped (Field Note 13)
Field Note 13 is out,
and its day was nineteen submissions — eighteen of them from the outside contributor
of the last two notes — which took the corpus from 33 to 53 modules at a speed where
the only step the contributor could not perform was the maintainer’s merge click. Most
of the new work formed two clusters: the rational trees (both path enumerations of the
positive rationals, the Stern–Brocot interval invariants, and a bridge proving that a
Calkin–Wilf path and a Stern–Brocot path reach the same pair exactly when one is the
reverse of the other), and a new Probability/ directory opened with three versions of
the Paley–Zygmund inequality plus Cantelli’s and the Chung–Erdős inequality. Six of the
eleven targets in the contributor’s own
contribution directions
are now marked landed there. The last arrival sat outside both clusters: the Caro–Wei
bound, an independent set at least as large as the sum of the reciprocals of one plus
each degree, merged by hand as
#259.
diaz-modulus-lean: thirteen warnings the v4.34 bump left in the mirror, and a header that now says what changed
The Lean and Mathlib v4.34.0 bump in diaz-modulus-lean
(a4f0779, the same
weekly bump that moved the rest of the Lean portfolio, most of it written up
on 21 September)
left thirteen warnings behind, all of them in ported mirror modules: eleven tactic
steps the unused-tactic linter now reports as doing nothing — four push_cast, five
field_simp <;> ring whose ring never runs, a gcongr <;> positivity — and two
haveI the linter asks to be have. Each is removed or respelled
(0c7ef1b), and because the
edits also live in the porter’s patch table, regenerating the mirror reproduces them
instead of reintroducing them: the porter’s --verify reads 117 identical and 0
differing against this commit.
LeanFrontier: the stranger came back, and their harness found the receiver's bug (Field Note 12)
Field Note 12 is
out, and its subject is the contributor from Field Note 11 — the account behind
the fork covered earlier this month —
returning with two theorems built on modules already in the corpus rather than two
isolated results. #179 proves
the local descent step of the Markov tree: for an ordered positive Markov triple
other than (1,1,1), the Vieta jump in the largest coordinate is positive, at most
the middle coordinate and so strictly smaller than the largest.
#180 bridges the accepted
Ford-circle criterion to Mathlib’s own geometry: two Ford circles with non-zero
denominators are externally tangent in the sense of
EuclideanGeometry.Sphere.IsExtTangent exactly when their cross determinant
squares to one. The receiver accepted them in 111 and 79 seconds, with zero exact
Mathlib matches, kernel replay, downstream import and all 118 and 119 entrypoints
already in the corpus still passing.
deck-lovers: the deploy script checks host ports before it builds
Podman only reports a host-port clash after the images are built, and only as an opaque “proxy already running” error that never names the port. The deploy script in deck-lovers now runs a pre-flight check in local serve mode before the build, naming the container or host process that is actually holding the port — 80 and 443 included when Caddy is started — while skipping this compose project’s own containers, so re-running over a live deck-lovers server still attaches instead of failing (5c07394).
portcullis stops printing its own secrets and pins the host key across phases
portcullis hardened its own handling of
the secrets it moves around during provisioning
(f41693f): it no longer
logs the first and last four characters of HCLOUD_TOKEN, keys/id_rsa is
created 0600 from the start rather than being written and then narrowed, and
the smtp.env values are shell-quoted before Phase 2 sources them as root —
passwords containing $, spaces, quotes or backticks had been mangled or
executed, and the file is now chmod 0600 on the VM before credentials are
written into it. Phase 1’s SSH host key is pinned and Phase 2 rejects a different
one immediately, so a swapped host cannot receive the second phase’s credentials,
and preflight validation now runs before any Hetzner resource exists.
Seven Lean repositories move to Mathlib v4.34.0
The Lean toolchain and Mathlib moved to v4.34.0 across the portfolio in one pass.
erdos-straus-offset-lean,
magma-1518-obstruction-lean,
moebius-transcendental-lean and
sharp-symmetry-bounds-lean
took the bump together with a new CI job that audits the axioms their headline
theorems actually depend on, so a silently new dependency on Classical.choice
or worse would fail the build rather than sit in the proof term.
inversive-geometry-lean
took the same bump and its README’s pinned version followed
(d512c1f).
Licences: MIT for the small tools, CC BY 4.0 for the prose
An MIT licence was added to repositories that had been public without one:
cold-path-server-podman,
llm-source-security-review,
spriter,
lifechess,
eclipse-3d-sim — whose
README was rewritten in the same commit — plus
book and
daily-drift-run.
solids-hunter had a
licence file whose text had drifted from canonical MIT; the canonical text is
restored and the third-party carve-out it had absorbed lives in THIRD-PARTY.md,
where it belonged.
lean-corpus-density pins the corpora its committed data was built from
The committed density data in lean-corpus-density now names the corpora and revisions it was built from (550a97d), so a figure in the report can be traced back to a specific corpus state instead of to the tool that read it, and the reproduction check of 21 September is recorded alongside it (31558d4). The repository’s claim is that machine mathematics accumulates; the data has to be pinned before that claim is checkable by anyone else.
forgepulse: the echarts 6 bump moved the legend onto the axis labels
The echarts 5 → 6 bump in forgepulse
changed the default legend position from top to bottom, so on both charts the
legend collided with the x-axis date labels while the 48px the grid reserves at
the top sat empty — legend.top is now set explicitly, so the layout stops
depending on a library default
(94fb8c1). The same release
draws a filled marker on every data point, which on a hundred-day series is
clutter; the dots are hidden on all eight series through one shared base while the
axis tooltip still marks the hovered point
(1f70b65). Both rode in with
the vitest 4.1.11 / echarts 6.1.0 security bump.
fleetlens is public: agentless health reports for a small VM fleet
fleetlens is now public: for a handful of Debian/Ubuntu VMs it logs in over SSH with Ansible, collects a fixed set of read-only facts and turns them into a JSON report, a Markdown report, a terminal summary and an optional email — nothing installed on the targets and nothing on them changed. It sits between “I SSH in and look around every few weeks” and a full monitoring stack: disk usage, pending updates, reboot-required, failed units and journal errors, with each host marked OK, WARNING or CRITICAL and the fleet taking the worst of them.
diaz-modulus-lean: the companion note reaches 1.0, where the case analysis ends
The companion note to diaz-modulus-lean
is now fixed at stable version 1.0
(f4cca8b) — the point
at which the formal case analysis stops rather than a point along it. Every branch
of Diaz’s conjecture is either closed by a machine-checked proof or reduced, by a
machine-checked reduction, to one of three statements: that e^{-iγ/π} is
transcendental for real algebraic γ ≠ 0, that e^{β/π} is transcendental for
real algebraic β ≠ 0, and that |u| is transcendental for a generic conjugate
pair of logarithms, which is the conjecture itself. All three follow from the
strong four exponentials conjecture, and the note says so, with the section the
abstract had always promised and the body never contained; section 5 gains the two
interpolation obstructions proved on 19 September.
A dependency-advisory sweep across the JavaScript, Python and Go repositories
A pass over open dependency advisories bumped the affected packages in every
repository that carried them. On the JavaScript side:
pickmydegree (vitest,
happy-dom, vite, surge),
storygen (vitest 4, vite
6, react-router 7), solids-hunter
(vitest 4, vite 6.4.3), eclipse-3d-sim,
lifechess,
agility-trainer and
they-live-agent
(vitest 4.1.11), platosdf
(vitest and its coverage package, 5.0.1), with lockfile refreshes in
deck-lovers and express
updates in neon-bumper-cars
and collective-starship-game.
CI workflows land across seven repositories
Seven repositories that had no continuous integration now run their own tests on
every push: topshift-trend
(running ruff and pytest), caciarabot
and python-hosts-checker
(pytest), euclean (pytest
together with its Lean build), and quadratula,
unstated-conclusions
and prove2me-logs with
smoke tests — in prove2me-logs’ case covering the p2m helper, the piece the
mission journal is actually read through.
200 to 181: the sweep I had already written and left out
The graded results for the SAIR Stage 2 challenge are in. My solver scored 181 of 200, the same figure on both tracks from the same file. Earlier I wrote about how it reached 200 on the organizer’s published sample set by removing things: six per-problem gates, a borrowed reference solver, two lookup tables. That post was right about what it described and incomplete about what it implied.
vitthorio follows the work
vitthorio started following this account — a
follow says someone is watching, the same way a fork says someone intends to try,
so it is worth naming rather than counting. The login is not a stranger to the
repositories here: their own
fork of jalhund/cold-path-server
is the other copy of the Cold Path game server that
cold-path-server-podman containerises.
diaz-modulus-lean and prove2me-logs: the two analytic obstructions fall, and the library says the 1973-74 theorem is proved
The two items diaz-modulus-lean
still listed under “What is not proved” — real analysis and interpolation
determinants, it said — are proved, and neither needed them
(2bc0828).
DiazModulus.no_first_order_arithmetic_operator shows why Schneider–Lang has no
input here: for a candidate u and F(z,w) = exp(uz + conj u w), a first-order
operator a ∂/∂z + b ∂/∂w with polynomial coefficients whose values on Z² are
all algebraic must have a = b = 0, because the bracket a(m,n)u + b(m,n)conj u
is algebraic — the exponential factor is a non-zero algebraic number — Baker
(carried as an explicit hypothesis, since Mathlib has no form of it) kills it, and
a polynomial vanishing on a product of infinite sets is zero; ∂/∂z ∂/∂w, by
contrast, does take algebraic values and is not a derivation.
DiazModulus.kronecker_factorisation closes the second: the matrix of values of
exp(auz + b conj u w) on the lattice is the Kronecker product of two Vandermonde
matrices, so its determinant is a product of powers of theirs and is non-zero —
the nodes are distinct because |exp u| = exp(Re u) ≠ 1, and the candidate’s
arithmetic does not appear in it. Both hypothesis classes are conjecturally empty,
which the nodes say on their face, and neither claims novelty; the first is stated
for polynomial coefficients where the note states it for rational functions
regular on Z², since clearing denominators is not formalised. The library now
holds 166 of 166 proved
nodes,
axioms clean, and the dead-branch table for the four superseded FourExp
statements moved into refresh_prove2me_archive.py, because the file carrying it
is generated and the refresh had been overwriting it.
sharp-symmetry-bounds-lean registered in the Palomar registry: PALOMAR-2026-09-18-000007
carlok/sharp-symmetry-bounds-lean
is now a registered entry in the Palomar registry —
PALOMAR-2026-09-18-000007, version 1,
status registered, trust level high, on the source of
commit ced9fe2
(the record’s own copy of the entry is
here).
This is the first third-party verification of one of these formalizations, and
what the registry actually did is the news: it rebuilt the project in a sandbox
from the pinned dependencies (Lean v4.32.0), exported the proof terms with
lean4export and replayed them on nanoda, an independent kernel
implementation rather than the author’s own build, then used the
Comparator
to check that Solution.lean proves the Challenge.lean statements — five
theorems of SharpSymmetryBounds — using only propext, Quot.sound and
Classical.choice, verified at 2026-09-18T13:50:47Z.
diaz-modulus-lean and prove2me-logs: the four exponentials theorem in transcendence degree one is proved
The 1973 construction whose four children
began to close in
diaz-modulus-lean is finished, and
with it the four exponentials theorem in transcendence degree one. The norm child
went first: FourExp.norm_to_polynomial_alg is proved
(a38d9db) by linear
algebra rather than the paper’s conjugates — the value is presented as
Π ∈ ℤ[X][Y] reduced modulo the monic Q, P = det M is the determinant of
multiplication by Π in the basis 1, Y, …, Y^(d−1), non-vanishing comes from a
kernel vector that would contradict the minimality of Q, and smallness from the
adjugate identity M·adj M = (det M)I. That made the core,
FourExp.construction_core_1973, a Proved node by cascade.
caciarabot: the digest stops posting links it cannot read
The digest in caciarabot now skips
candidate links whose target page is in another language
(7fd451b), because
GitHub trending routinely surfaces repositories documented entirely in Chinese —
two of the fifty candidates in that day’s live fetch were exactly that, and
posting one of them as the link of the day wastes the slot. The check runs in two
stages, cheapest first: the pool is filtered on the title and description the
source already returned, no extra request; only the picked candidate is fetched
and checked, because verifying fifty to post one would be fifty requests a day,
and a rejection drops that link and draws again, up to five times. For a GitHub
repo the check reads the raw README rather than the repo page, since github.com
serves <html lang="en"> on every page it renders, including for repos written
entirely in Chinese; elsewhere the declared lang decides, falling back to an
English-stopword ratio over the visible text. Two deliberate non-rejections:
Latin-script languages pass the metadata stage (ten words of French cannot be told
from English reliably), and a page yielding no usable evidence is accepted rather
than quietly thinning the pool — and Greek is left out of the non-Latin script set
on purpose, since a lone alpha here is more likely to be mathematics than prose.
slop-audit: the brush-up pass, and the rename to slop_audit
slop-audit spent the rest of the week
being made presentable rather than more capable. A
brush-up design split
the work into Phase 1 (docs and packaging) and Phase 2 (hygiene), and Phase 1
followed: an MIT license, a changelog and pyproject.toml metadata
(e7bc0a3) with the
README polished and GitHub publication marked done
(9b6a418), merged as
PR #1.
sharp-symmetry-bounds-lean is public: sharp symmetry bounds for real plane curves
sharp-symmetry-bounds-lean
is a new public Lean 4 repository formalizing
Theorem 1 of the working note “Sharp symmetry bounds for real algebraic curves”:
if an infinite real plane curve of degree d ≥ 2 is irreducible over the complex
numbers and is not a circle, its Euclidean symmetry group is finite — the
orientation-preserving part cyclic of order at most max(d, 2d − 4) and the full
group of order at most 2d — both bounds are attained in every degree, and for
d ≥ 5 the curves attaining the rotation bound are classified, up to
orientation-preserving similarity, as Re(z^(d−2)(|z|² + a)) = 0 with |a| = 1
and a not real.
sharp-symmetry-bounds-lean gains a reviewer's account of Theorem 1
sharp-symmetry-bounds-lean now
carries a document written for someone checking the formalization rather than reading
it: docs/THEOREM1.md gives the informal statement of Theorem 1, a proof outline
mapped step by step to the Lean declarations that carry it, fidelity notes on where the
informal and formal statements diverge, and reproduction instructions
(85ddda8). The
addition makes the informal argument the declared source, since formalization.yaml now
points its original-proof source at this document and records the passing mechanical
preflight. A stale comment about sharpness in lean/DirectBound.lean was corrected in
the same commit.
portcullis: a two-phase hardened Ubuntu VM for Hetzner Cloud
portcullis is a new public Python
repository: one command creates a hardened Ubuntu 26.04 VM on Hetzner Cloud, and
it drops the gate first — an unprivileged user with key-only SSH on a random high
port, root locked, UFW default-deny and sysctl hardening, all inside about 30
seconds and using only what the stock image already ships — then switches the
Hetzner firewall to the new port and deletes the temporary API key before the
full CIS-style pass (package upgrades, AppArmor, auditd, AIDE, PAM policy,
fail2ban, rkhunter, msmtp alerts, Docker and Podman) runs behind both firewalls.
verify.sh then runs
58 checks on the finished host. Everything runs inside a Podman container so
nothing is installed locally, and teardown deletes a half-provisioned server
together with the keys and firewalls it created. The
README was rebranded as
the repository went public, with a
GitHub Actions test workflow
on actions/checkout@v5 and a
logo.
forgepulse: the rank column stops stretching to the row height
The rank-trend arrows in
forgepulse’s Repository signal table
cost every row its bottom border alignment:
display: flex on .rank-share, added to stack the arrow, pulled that <td>
out of table layout, so it stopped growing to the row’s real height and stayed
clipped to its own content while the Name column stretched normally for longer
descriptions. Dropping the flex display and keeping the existing block-span
stacking restores the alignment — two lines of CSS, which is what a change made
to fit a two-character indicator should have cost in the first place.
forgepulse: per-repository clone baselines, and a sidebar that folds on a 14-inch screen
The clone-trend arrows added to forgepulse
two days ago were reading
“stable” almost everywhere, and the cause was GitHub’s traffic API rather than the
repositories: it reports a day’s row inconsistently per repository, so at any moment
only a handful have a same-day daily_traffic entry while most lag behind, and a single
fleet-wide “yesterday” cutoff therefore compares most repositories against data already
included in their current total
(762ea7b). The fix is to compare
each repository’s clone total against one day before that repository’s own most recent
row, computed in one query through a per-repository MAX(day), so the comparison never
depends on a shared calendar date — verified live, the trends now show real up and down
movement instead of universal stability.
diaz-modulus-lean and prove2me-logs: the 1973 construction's four children start closing
The restated 1973 construction
split in four the day before has
begun to close in diaz-modulus-lean,
one archive-and-port step per child. The field came first: FourExp.trdeg_one_presentation
makes ω = x₁y₁ transcendental by Hermite–Lindemann, everything else algebraic over
Q(ω) from transcendence degree one, then a primitive element scaled to an integral
generator with a monic minimal relation and a common denominator for the eight numbers
(5f3ee4e). Two of the
remaining children needed restating first, because auxiliary_function and
norm_to_polynomial had omitted that the four exponentials are algebraic — without it
the ω-degree of the powers (e^{x_i y_2})^{jb} is unbounded — so they came back as
auxiliary_function_alg and norm_to_polynomial_alg with the old pair left as a dead
branch (fced437).
diaz-modulus-lean and prove2me-logs: a misread exponent, and the 1973 construction split in four
With the transcendence criterion and the zero count closed, the last thing between
the four exponentials theorem in transcendence degree one and a full proof is
Waldschmidt’s 1973 construction — and re-deriving it against the published
statement before splitting its core turned up
a sign lost when the formula was read off a noisy text layer:
the order of differentiation is S = ⌊N²(log N)^(−1/2)⌋, not the published
S = ⌊N²√log N⌋. It matters — the construction’s polynomial has degree about
r·S and log-height about S·log S, both of which must stay under a fixed
multiple of N²/√log N and N²√log N; with the paper’s S the ratios stay
bounded (1.00 and about 1.9 at every size checked), with the published one they
grow to 27.6 and 56.9 at N = 10¹².
prove2me-logs: four exponentials gets a tree, its first leaf closes, and the zero count's source has a hole
The Sep 14–15 entries in prove2me-logs
open a subtree under four_exponentials_trdeg_one, the one Open leaf of the
Diaz mission that is a theorem in the literature: six new nodes joined by three
accepted reductions
(b4002e2), routed
through Waldschmidt 1973 with two tools from his 1971 paper after reading and
rejecting the two Roy–Waldschmidt proofs. Transcribing the statements from page
images rather than the text layer caught a missing hypothesis (σ₂ ≤ σ₁) and a
wrong exponent, and the lattice step forced one classical fact out into a node
of its own — an exponential polynomial with distinct frequencies does not vanish
identically — which closed as the branch’s first leaf
(f8f57d7).
prove2me-logs: the zero count closes, and Gel'fond's criterion with it
The Sep 15 entries in prove2me-logs pick the Diaz mission up where the previous posting left it — the zero count split, with a hole in its source — and carry it to the end of the branch. Every FourExp leaf was given a reduction rather than a citation (e14a87b), the transcendence criterion was written out as a Lean argument (75ed372), and two of the classical leaves were proved from what Mathlib already has: Gel’fond’s height bound for a divisor, which is the Mahler-measure file plus the last inequality, and the resultant step that Mathlib reaches through the Bézout identity but has to finish with the Leibniz expansion instead of Hadamard (bed556a).
pacenotch: a Settings window, start at login, and one instance at a time
pacenotch’s GUI grew the controls a
.app cannot be handed on the command line: a Settings window (footer
button, tray menu, pacenotch → Settings… on macOS) for the on-pace band,
the refresh interval, notifications, compact mode, appearance and Dock
behaviour, applied and saved as they change
(13dc75f). Because a
MacBook’s camera notch can hide the menu bar icon and with it the only Quit,
the app can now switch to the Regular activation policy and live in the Dock
and Cmd-Tab with a Quit in the window and a Settings window that is not hidden
by the notch (b6f4031).
forgepulse: rank-trend arrows in the repository signal table
forgepulse’s Repository signal table now
carries a Billboard-chart-style up / down / stable indicator on every row, in
both ranking modes, showing whether a repository moved since yesterday
(9306fec). It tracks
rank rather than the raw figure on purpose: clone volume ranks by
total_clones, a cumulative counter that only ever grows, so diffing the value
would read “up” almost every day regardless of real trend, while rank stays
relative and keeps meaning.
diaz-modulus-lean: the four exponentials subtree archived and mirrored, and the zero count reduced
diaz-modulus-lean now holds the
four-exponentials branch instead of citing it: the FourExp nodes are archived
alongside the Diaz ones
(3d5d9ce) with the
reduction sketch and its pieces
(57f3ec7,
6714dcb), the
refresh script searches the FourExp namespace too
(ab6aa90), and
each Open node’s formal statement and write-up is now mirrored into
archive/prove2me/open — rewritten when the board changes, deleted once the
node stops being Open, and covered by --check
(3eb78a7).
diaz-modulus-lean: the four-exponentials leaves close and the transcendence criterion is machine-checked
The four-exponentials branch of
diaz-modulus-lean went from a
tree of Open nodes to a closed criterion, in a sequence of archive-and-port
steps: the transcendence criterion reductions and their children were mirrored
(707c4d2), then
the 1973 construction reduction and its four children
(e79b003), and
then the leaves were proved one at a time, each archived, ported and dropped
from open/:
slop-audit: a local harness for AI-slop signals and formulaic prose
slop-audit is now public: a Python
harness that scores a piece of prose for AI-slop and formulaic style, general
writing issues and descriptive stats, and — kept strictly separate from all of
it — local AI-authorship detector signals. Nothing leaves the machine: there
are no hosted detector APIs and no remote LLM touching the target text, with
the network used only for installs and optional model-weight downloads. A run
produces a Slop Index (0–100 plus band) from the slop tools and linters
(05e1527), deterministic
edit suggestions quoted from the tools that would lower it — no LLM rewrite of
the prose (7bcdcc3) —
and a per-tool matrix reporting OK / NOT_RUN / ERROR with the exact
reason, so a tool that failed never invents a score.
quadratula: how much of Schröder's 990 quasigroup laws small quasigroups already witness
quadratula is a new public repository that asks how much of the implication structure of Schröder’s 990 quasigroup equational laws is already visible in small quasigroups. It enumerates every quasigroup of order 1 to 6 up to isomorphism — 1,131,984 classes, in Rust on a pinned toolchain — computes which laws each one satisfies, and measures that exhaustive floor against Bruno Le Floch’s arXiv:2603.29909: quasigroups of order at most 4 already witness 95.34% of the 726,207 law-level non-implications (91.68% of the 1,958 between the 47 classes), and going to order 6 reaches 96.70%, realising 94 of the 114 varieties and separating 42 of the 47 classes.
prove2me-logs: two cloud runs on the Diaz mission, and a helper that stops mangling LaTeX
prove2me-logs records the Diaz
mission being worked by a cloud agent with nothing but a Prove2Me key, a GitHub
token and the public brief: the
first run proved
DiazModulus.candidate_no_real_algebraic_line, the
second added
candidate_distance_transcendental by polarization on top of the first run’s
line exclusion — both accepted, both mirrored into
diaz-modulus-lean. Offline the
same day, the
multiplier module
determined what an extension of the three-dimensional hull could be
({z : u·z ∈ L̃} = Q̄ + Q̄/u, so no shifted reciprocal survives) and the one-log
saturation node ruled out a candidate inside Q̄ + Q̄·l. The session’s most
useful artifact is a
curl-free helper:
passing a JSON body through a double-quoted shell string let bash expand $u,
$v and $$ inside the LaTeX before the request went out, which wrote mangled
mathematics into 52 published nodes — repaired from a pre-damage snapshot — and
tools/p2m.py now builds the multipart request itself and accepts any body as
@file.
magma-1518-obstruction-lean: the November 2024 thread gets the credit, and the l2_note catches up with its theorems
The credit corrections in
magma-1518-obstruction-lean
went a step further: every equivalence among the four 1518 targets was posted on the
Lean Zulip in November 2024 — Tao relating 47 and 614 through law 359, Tencer tying 817
to S³x = x, Bolan noting with Prover9 that 3862 implies the other three, and Nielsen
verifying with Vampire that all four agree under left cancellation — so
3862 is no longer presented as an improvement over Tao’s conjecture,
which the earlier correction still called one, and what the repository adds is now
stated as the proof rather than the reduction.
LeanFrontier: a stranger ran the same machine (Field Note 11)
Field Note 11
— “A Stranger Ran the Same Machine” — records the corpus’s first contribution
from a first-time outside author: qazW12345,
whose fork was the previous day’s news,
sent PR #175 formalizing
Nesbitt’s inequality, 3/2 ≤ a/(b+c) + b/(c+a) + c/(a+b), by clearing a
positive common denominator onto
2N - 3D = (a-b)²(a+b) + (b-c)²(b+c) + (c-a)²(c+a). The only human
intervention was approving the paused first-time-fork workflow; the trusted
receiver then accepted the submission in 107 seconds — 61 added lines, no exact
Mathlib fingerprint match, a kernel replay pass and an allowed axiom closure of
propext, Classical.choice and Quot.sound — and the theorem is now
importable as LeanFrontier.Analysis.Nesbitt, with the report and observation
persisted by the usual generated changes
(PR #178). Nobody judged the
proof’s elegance or importance, which is the point: an unfamiliar contributor
reached the same boundary through the ordinary public route. The corpus is at
carlok.github.io/LeanFrontier.
forgepulse: manual refresh and an on-demand sync, behind a single-flight guard
forgepulse grew the two controls its
dashboard was missing: Refresh re-fetches the stored dashboard and detail
data without touching GitHub at all, while Sync now performs a real pull
through the existing POST /api/v1/sync endpoint outside the hourly schedule
and now surfaces the server’s error inline — an expired token says so
immediately instead of failing silently until the next scheduled attempt
(87cfe54). On-demand
sync then exposed a contention bug worth fixing properly: the scheduled tick and
the new endpoint shared one GitHub collector with no coordination, so a click
landing during the hourly run let multiple full 84-repo syncs proceed at once,
each with its own bounded per-repo fetch — enough load to trip secondary rate
limiting and queue on SQLite’s single writer, turning overlap into a crawl
rather than an error. A single-flight guard now rejects the second sync instead
of letting the two contend
(47b2e73).
diaz-modulus-lean mirrors the Prove2Me Diaz nodes, and archives all 167 accepted proofs
diaz-modulus-lean absorbed the
Prove2Me work rather than citing it: diaz_of_sfe — strong four exponentials
implies the conjecture — was ported first
(fcf8130), then
the rest: Hermite–Lindemann is now proved here instead of assumed, so
#print axioms Diaz.diaz_of_sfe returns only propext, Classical.choice and
Quot.sound, alongside the fibre bound, the quantisation batch, the six
exponentials node and the candidate statements mirrored with their platform
submission ids
(0144e57,
fcf8130,
e4880ab,
8d02fd4,
7cb8b3a). The
repository also gained a complete copy of the mission:
archive/prove2me/
holds all 167 accepted submissions for its 132 Proved nodes plus a manifest, so
no proof of this mission exists only on the platform
(1f05ae7) — kept
deliberately unbuilt, since the files target the platform’s Mathlib revision and
nothing in CI reads them, with
scripts/refresh_prove2me_archive.py
refreshing the archive and --check failing when it goes stale
(735f5b4). The
companion note is published as
tex/diaz_prove2me.tex
and the CI build was fixed by dropping a says-verified simp list
(e8265ae).
pacenotch is public: Claude usage limits with a vertical pace notch
pacenotch is a new public Go repository: a CLI, tray icon and window for macOS, Windows and Linux that reports Claude’s usage windows (5-hour session, 7-day all models, and the 7-day Sonnet/Opus windows when the plan reports them) with a thin vertical notch drawn at the elapsed fraction of each window — the fill is past the notch when you are ahead of an even pace, short of it when there is room left, and the tray icon turns red on that side of the line (d7d6da6, 371aefc). The first day of commits also landed the usage/pace packages and terminal renderer with a compact mode below 60 columns, parity and coverage tooling, and a CI matrix with packaging and the README (b8f822a, b6c6151). The macOS build learned to raise its window in front of other apps and to reopen it when the running app is launched again (9f9aa5e, 6ca1263). MIT-licensed and unofficial — not affiliated with Anthropic.
carlok/LeanFrontier was forked by qazW12345
carlok/LeanFrontier was forked by qazW12345 on 2026-09-11. Forks are the corpus’s intake path rather than a copy: a submission arrives as a branch on somebody else’s fork and is judged there, which is what Field Note 10 — PR #167, the Tribonacci submission — was about. A new fork is worth recording for the same reason a new submission is: it is the only outward sign that someone intends to try. The corpus itself is at carlok.github.io/LeanFrontier.
contributed to mathlib4: #43660 merged — unused hypotheses weakened in Data and Order
Contributed to
leanprover-community/mathlib4:
PR #43660 was
merged into master on Sep 10, weakening unused hypotheses in
Mathlib/Data/Finset/Pairwise.lean, Mathlib/Order/Antichain.lean and
Mathlib/Order/Monotone/Extension.lean — five one-line binder changes, five
insertions and five deletions. It answers
a request on the mathlib Zulip
to slice the machine-found weakenings into smaller, reviewable PRs, and the
unused-assumptions pipeline
supplied the edits: replace one binder with a weaker class, keep the proof
unchanged, let the compiler decide. The larger
#43503 — 32
typeclass weakenings across 29 files — is still open.
unused-assumptions: the published repository learns what to keep out
The unused-assumptions repository
now separates what is published from what is merely being worked on: outreach/
workshop text for upstream pull requests and messages — including other people’s
contact details and prose that is not its author’s to publish — is ignored at the
root rather than parked in a sibling repository
(43e092a), since the
drafts are about this project’s results and belong beside them. The same instinct
ran through the leak-detector test, which had been asserting that no shipped tool
names the private sibling project while spelling out exactly those names as
literals — now assembled from fragments so the repository stops carrying the one
string it exists to withhold
(3e1cb9a,
78ce2c0). One
weakening was also withdrawn as unwanted despite compiling
(b0e1f4f): a patch
surviving the compiler is not the same as a patch worth publishing.
topshift-trend: a cooldown so a chat is not notified twice about the same repo
topshift-trend, the Telegram bot
that watches GitHub’s monthly trending list, learned to
suppress repositories it recently notified:
a repo that drops out of the top-N and comes back used to announce itself to
every subscriber again. The bot now keeps a notification_history.json
alongside its state and subscribers, and the scheduled check filters out keys
seen inside a configurable window — NOTIFICATION_COOLDOWN_DAYS, default 30
days, 0 disables the suppression. The commit touches the config, store and
main loop plus three test files, and the check’s log line now reports how many
entries were suppressed next to how many notifications actually went out.
prove2me-logs: the Diaz root falls to strong four exponentials, and half the tree goes nowhere
The Sep 9 session in prove2me-logs
closed DiazModulus.diaz_of_sfe:
with x = (1, u) and y = (1, conj u) the four products are 1, conj u, u and the
squared norm, all in the tilde space, and Hermite–Lindemann discharges from a
node the mission had already proved — so strong four exponentials is the only
hypothesis left standing. It came out of a
design review that rejected its own design:
three scoped reviewers returned REJECT on the alternative route via algebraic
independence of logarithms, nothing was published from it, and the argument
turned out to already exist verbatim in the accepted diaz_of_schanuel
submission.
magma-1518-obstruction-lean: a Palomar-ready package, and two credit corrections
magma-1518-obstruction-lean grew a
Palomar-ready package:
Challenge.lean states Theorem A and the F_5 / F_13 members of Theorem F as
coefficient matrices without imports, Solution.lean proves them from the core
development and kernel decide, and a comparator pinned to Lean v4.33.0 accepts
the pair locally — formalization.yaml, comparator.json and a PALOMAR.md
readiness record are generated by scripts/gen_palomar.py.
forgepulse: the dashboard now shows when its numbers were last refreshed
forgepulse shows the age of its own data:
the homepage now reads the most recent successful sync run — already recorded in
the sync_runs table but never surfaced — and renders it next to the git-ref
pill as Updated <UTC> · <local> (Xh Ym ago), ticking every minute
(d4c5e85). Traffic
history is only as useful as the moment it was collected, and until now the
dashboard gave no way to tell a stale snapshot from a fresh one. The change
touches two new frontend modules plus tests for the relative-time formatter and
the API field behind it.
prove2me-logs: the Diaz tree's free half is closed — frontier at four leaves
A second Sep 8 session on the Diaz tree closed the free half: no admissible matrix exists there, and that is now proved, after which norm_free closes onto the published statement (S) and the frontier stands at four leaves. The log also publishes an SFE route the session’s other model proved and withheld, records the lesson that the relation you are using is the one that breaks your argument, and scopes the next target with a brief for the four-exponentials theorem at transcendence degree one.
prove2me-logs is public: the Diaz main theorem reduces to a single named leaf
carlok/prove2me-logs is the working log of my Prove2Me formalization activity — per-mission entries with theorem uuids, Lean environments, and what remains open — public since Sep 7 as a record of the work rather than an archive of the proofs. The first entries record dead ends beside progress: the free-ring no-go formalized with its prose proof intact, a six-exponentials no-go note with a Waldschmidt erratum, and a Diaz bridge node that turned out to attach to nothing. Day two records the Diaz tree and writes up the decompose–link–iterate rule behind it: by the fourth generation the main theorem’s difficulty sits in a single named leaf — pi transcendence — with both leaves of the second branch named.
unused-assumptions: Mathlib theorems whose typeclass setting is stronger than their proof
carlok/unused-assumptions is a
new public repository (visible since Sep 4; v1.0–v1.2 tagged Sep 6, archived at
doi:10.5281/zenodo.22549525):
theorems in Mathlib whose stated algebraic setting is stronger than their own
proof requires. The method is one sentence — take a theorem, replace one binder
with a weaker class, keep the proof byte for byte, compile it alone — and every
row of data/survivors.jsonl carries what is needed to put the claim back in
front of the compiler. A verifier rechecks each row at the Mathlib revision the
manifest names, refuses to run against a different one, and requires
#print axioms to rest on nothing beyond propext, Classical.choice and
Quot.sound. Re-verifying the 36 candidate patches against a later Mathlib kept
the 33 that still hold,
and the README now answers the prior-art question explicitly rather than leaving
it to the reader.
unstated-conclusions: theorems whose proofs deliver more than they state
unstated-conclusions (public
since Sep 6) is the dual of
unused-assumptions: instead of
weakening hypotheses, it asks which theorems in Mathlib prove a stronger
conclusion than they state. It reads the root of the proof term — if the last
step is a weakening lemma (le_of_lt, Or.inl, Exists.intro w _,
And.left, …), the stronger statement is already there as a subterm with its
own proof, so the finding typechecks by construction. A hand-written table of
22 weakening lemmas is the only judgement. The project runs in two parts with a
statistical wall between them: Part 1 is a pilot at the fifty-candidate gate on
unused-assumptions’ own survivors (a rate there is a rate among those theorems,
and nothing more), and Part 2, the library-wide measurement, starts only after
Part 1’s table is frozen.
Prove2Me week one: the EML ladder's size-7 step, an Ash–Stevens cusp sum, and Spencer's trivial range
First check-in on my Prove2Me profile:
joined this month, rank Master, trust 33, six missions, 35 statements solved
and 32 posted. Three proofs landed today with my name on them, all against
Mathlib 0df444a (Lean v4.33.1).
portcullis: Vim stops stealing the terminal's mouse selection
portcullis hardens a fresh Ubuntu host,
and hardening has to leave the machine pleasant to actually use: Ubuntu’s Vim
defaults enable xterm mouse reporting, so on a remote terminal Vim captures the
selection and ordinary copy/paste stops working. A new Phase 2 section,
1.6a — Terminal editor defaults, now
ships /etc/vim/vimrc.local
with set mouse= (and a note in the README’s Phase 2 list), leaving selection to
the terminal emulator while anyone who wants Vim’s mouse support can opt back in.
contributed to mathlib4: #43503 bundles 32 mechanically found typeclass weakenings
Contributed to
leanprover-community/mathlib4:
PR #43503
weakens unused typeclass assumptions on section variables across 29 files —
32 one-line changes, 32 insertions and 32 deletions. Each was found by the
unused-assumptions pipeline
(LLM + Lean 4 propose, the compiler decides): replace one binder with a weaker
class, keep the proof unchanged, compile it alone. Every change built against
its unmodified file, with the whole library compiling locally at 633b366493.
It is one PR rather than many by deliberate choice, citing mathlib’s own
#42214 (813
files of the same kind of change) as precedent; the background is on the
mathlib Zulip,
where a reader spotted a further simplification in one refactored file.
unused-assumptions: the patches are not a set, and one written off was still true
Building a branch from all 36 weakenings in
unused-assumptions
applied only 28, and the
eight that did not apply are not all failures: two patches for Indicator.lean and three
for Untop0.lean rewrite the same variable line in different directions — one weakening
the algebraic class, another the order — so they are alternatives rather than additions,
and applied in sequence the first moves the context and the rest silently do not apply.
patches/README.md now says so, with the further caveat that a file carrying two
weakenings at once is a combination nothing here has compiled, since
REPORT-master.json is per-patch and not per-combination, and three tests hold the line:
alternatives must be documented wherever a pair rewrites one line, every patch touches
exactly one file, and every changed line is a variable line.
parsimagma: the 820 infinite-only pairs, enumerated
parsimagma now
enumerates, pair by pair,
every implication that is false in general but holds for every finite magma:
820 ordered pairs — 610 from the unresolved hard core and 210 from the
saturation-refuted set — decoded from the ETP’s closed implication graphs into
data/etp/infinite-only.tsv. The exact split answers the question left open
in the project’s
#1474, and it
makes the corpus’s coverage readable honestly: 610 of the 1062 hard-core pairs
can never be refuted by a finite construction, so the real figure is 411 of
the 450 finitely refutable (91%), with the remaining 39 listed. A differential
test against the ETP’s Lean-verified finite graph agrees on all 797 finite
claims.
magma-1518-obstruction-lean: Theorem F — explicit finite models refuting 1518 ⇒ 47/614/817/3862
magma-1518-obstruction-lean
grew a third result overnight:
Theorem F
builds an explicit parametric family of finite magmas satisfying law 1518 that
violate the four targets 47, 614, 817 and 3862 — the ETP’s known 15-element
countermodels are the family’s two smallest members, which then continues with
27, 39, 51, 75, … elements. A companion classification shows the
base-dependent extensions of the Z/3 shift come in exactly 8 classes for
p ≡ 1 mod 4 and 4 for p ≡ 3 mod 4, with only the family and its conjugate
refuting. The 15- and 39-element members are checked by the Lean kernel via
decide with no axioms, and the classification rests on Gröbner bases over
Q and over every F_p with p ≤ 257.
parsimagma: where the certificates' three axioms actually come from
The certificates README in parsimagma had
the origin of its three permitted axioms wrong — they enter through the
finOpTable encoding, not through what decide invokes — and the
correction credits Wenlin
Zhang, who settled it by re-emitting 44 affine models as plain arithmetic
operations: same goal, same tactic, no axioms, 44/44 at carriers 2–9. The same
README now also records that the judge’s default policy admits no axioms at all,
so the false column of the table fails under it while the true column does not —
which is the difference between a certificate that is checked and one that is
merely re-runnable.
magma-1518-obstruction-lean: the announcement lives on the Zulip
The results of magma-1518-obstruction-lean went out on the Lean Zulip, and the repository briefly carried a draft of that announcement before it was removed the same morning: an announcement lives in the thread it was posted to, not as a copy in the repository that announces it. The same pass repointed the markdown notes at repository paths, so a reader followed a link into the repository rather than into a working copy.
magma-1518-obstruction-lean: one-generated 1518-magmas, and a cohomology wall
magma-1518-obstruction-lean
is a new public Lean 4 repository about the Equational Theories Project law
1518, x = (y ◇ y) ◇ (x ◇ (y ◇ x)). It proves that every one-generated magma
satisfying 1518 and 3862 is trivial or the Z/3 shift — Terence Tao’s
conjecture from the November 2024 Lean Zulip, now without a finiteness
assumption and with a single target law — where the core table-and-closure
argument carries no axioms. A second result shows constant-coefficient magma
cohomology cannot refute 1518 ⇒ 47, 614, 817, 3862 from any finite base:
H² vanishes over the shift, so every such extension is a direct product. The
README states the full theorem set (A–E) with a per-statement tally of
confirming tools — core Lean, Vampire, Mace4, z3, brute force — and notes the
repository claims no new implication, since the ETP has settled them all.
magma-1518-obstruction-lean: Proposition A″ re-derives an observation of Bruno Le Floch
The write-up in magma-1518-obstruction-lean
now says where Proposition A″ comes from: its equivalence of the four single-variable
targets 47, 614, 817 and 3862 under 1518 together with left injectivity and left
surjectivity — which hold in every finite 1518-magma — re-derives
an observation of Bruno Le Floch
posted on the Lean Zulip’s Equational stream in October 2025, that in a finite 1518-magma
the squaring map and all left multiplications are bijective and the single-variable
targets are then equivalent to each other and to S S S x = x. Naming the observation
that was being re-derived is the difference between a proposition that stands on its own
and one that stands on someone else’s step.
dratify v0.1.4/v0.1.5: published speed figures replaced with numbers anyone can reproduce
dratify tagged v0.1.4 and v0.1.5 after an audit of its published numbers: the README’s speed table came from CNF files that are not in the repository, and its “~18x” was the top of a range quoted as the typical value. A committed benchmark script (bench/repro.py) now produces figures anyone can reproduce — a second run held the geometric mean at ~15x while individual ratios moved between 11x and 22x, so the release quotes the mean and states the spread as a spread. The same pass (c34c31b) added a fuzz workflow and stopped error messages from reporting version skew in the vocabulary of an unrelated project’s module.
cdclkit v0.1.3: a self-check that verified nothing, honest benchmark notes, and docs that describe the program
cdclkit tagged
v0.1.3: the release leads
with the bug that mattered — --adaptive --self-check printed
s UNSATISFIABLE while verifying nothing. Both engines now require
dratify 0.1.4, with a test that fails if
they ever ask for different versions again. The same pass rewrote the docs to
describe the program (a fresh make test failed with 24 import errors while
the Makefile still claimed no dependencies), and
BENCHMARKS.md now
says what the published 20.1x figure is a sample of: one run per instance,
with individual ratios seen moving by a third between runs while the geometric
mean held.
179 to 172 to 200: a Stage 2 solver that improved by subtraction
The release itself is already noted: parsimagma v1.0-stage2 settles all 200 problems of the SAIR Stage 2 sample set in one standard-library-only Python file, archived at doi:10.5281/zenodo.22237100 with a write-up at doi:10.5281/zenodo.22214743.
dratify v0.1.3: tests behind the differential claim, and untrusted input that can no longer ask for gigabytes
dratify is tagged
v0.1.3: the docs said the
Python checker and the Rust checker were compared against DRATChecker on
acceptances and rejections, but the only two tests that did so had silently
skipped on every run — an import-time guard — so that claim had never once been
exercised. The new differential suite
(2889df3) sweeps 150 random
instances plus hand-written accept/reject cases, and CI installs the native
checker and fails if the tests skip. The same release hardens the
untrusted-input paths: header counts like p cnf 99999999999 0 that would have
asked for hundreds of gigabytes are rejected outright, register_native
validates the checker it is handed instead of unpacking a fixed tuple, and
proof parsing reports the offending line and token. Tests go from 87 to 96 and
coverage 84% to 87%, with each new guard confirmed to fail when its fix is
reverted.
An external check: fifty pairs agreed, two numbers did not
After parsimagma went public, someone read it properly.
parsimagma v1.0-stage2: 200/200 on the Stage 2 sample set
parsimagma is tagged v1.0-stage2: one standard-library-only Python file — no database, no lookup table, no LLM call — reaches 200/200 on the organizer’s 200-problem sample set, confirmed by two independent runs agreeing problem for problem and explicitly not the private graded set. Three search mechanisms split the work: finite model search refutes false implications cell by cell, critical-pair completion proves most of the true ones, and ordered superposition under a Knuth–Bendix ordering takes the rest. Every derived equation carries its replayable derivation, paths are double-checked before any Lean is written, and certificates are emitted from the inference DAG rather than flattened — 13,728 bytes against 3,634,949 on the hardest problem.
moebius-transcendental-lean v0.2.0: the conjugation-degree spectrum is classified
moebius-transcendental-lean
tagged v0.2.0,
which classifies the conjugation-degree spectrum on the transcendental locus:
conjDegree attains every value in ℕ∞ except 0, which it never attains. Every
finite degree gets the same explicit witness, zₙ = sⁿ + i·s with
s = liouvilleNumber 2, while the ⊤ stratum comes from an algebraically
independent real pair. It compiles against Mathlib v4.32.0 with no sorry,
and the permanent axiom-verification module confirms the ten new declarations
close over exactly {propext, Classical.choice, Quot.sound}.
forgepulse: human attention ranking lands on main
forgepulse merged its first pull request, PR #1: human attention is now the default explainable ranking alongside clone volume. The dashboard persists fork history, collects release/tag and Actions-checkout signals, and shows each traffic diagnosis with its evidence inline. Clone-volume mode also gained sortable Stars/Views/Clones column headers (c69739b).
parsimagma: concept DOIs in the badge, version DOIs in the paper, no working directory in the proof logs
parsimagma’s citation surface was settled deliberately: the README badge and posts cite the concept DOI 10.5281/zenodo.22237100, which always resolves to the newest release, while the paper’s bibliography keeps the version DOI, because a report of measurements should pin the artifact it measured — and the write-up DOI was repointed at the current version rather than the one carrying a superseded-version banner (16d60e0). The same pass took the working directory out of the published Vampire logs: the proofs are worth publishing, the absolute temp path they were produced under is not (2f9bf2d).
LeanFrontier: the fork was the test (Field Note 10)
Field Note 10 — “The Fork Was the Test” — records the first submission to arrive from another person’s fork: PR #167 defines the Tribonacci sequence and proves that every term past zero is positive and that finite partial sums satisfy a telescoping identity — a genuine recurrence-sequence result, not a re-labelled Fibonacci or Padovan one. A Mathlib base bump moved underneath the PR mid-flight, a clean rebase put it back at the boundary, and the receiver — not a mathematical reviewer — decided whether it could cross. The corpus is at carlok.github.io/LeanFrontier.
kiwifarmit/cra-lab is public: CI/CD pipelines that produce CRA evidence
kiwifarmit/cra-lab is now public — a
small lab in the Kiwifarm org that builds the pipeline half of a Cyber
Resilience Act story instead of describing it. Pull requests go through Semgrep
SAST plus SCA and IaC scanning (Semgrep rather than CodeQL, which needs GitHub
Advanced Security on private repos, and it runs entirely on the runner so the
code never leaves it); a v* tag generates a CycloneDX SBOM with Syft, hashes
it, hands the digest to the SLSA generic generator for provenance, and runs
Trivy over the release SBOM at CRITICAL/HIGH before attaching it to the release
(pr-security.yml,
release-security.yml).
A weekly Trivy and Semgrep sweep covers vulnerability, misconfiguration and
licence scanning on a timer, and Dependabot keeps the actions the workflows
themselves depend on pinned, with each workflow’s comments citing the CRA
Annex I Part II clause it answers.
dratify: the release workflow now tests the tag it publishes
dratify’s release pipeline could report a
successful release that published nothing: ci.yml triggers on branches and
pull requests and never on tags, so no test ever ran against the ref being
published, and combined with skip-existing a tag that forgot to bump the
version built the old one, had PyPI skip it as already present, and exited
green. The suite now
runs on the tag, and the build fails when the tag does not name the version it
produced — skip-existing
should absorb a re-run of the same release, not disguise a forgotten bump.
caciarabot: videos and GIFs alongside images
caciarabot word-trigger responses now ship videos and GIFs alongside images, so the bot can reply with moving media instead of only stills. The update script also always restarts the bot now, so config-only changes actually land on the next deploy.
forgepulse v0.1.0: the baseline release
forgepulse hit its first tagged
release, v0.1.0:
the self-hosted GitHub traffic dashboard with retained clone and view history,
repository analytics, charts, referrer and path snapshots, and JSONL export.
The release note frames it explicitly as the baseline before the next feature
— human attention ranking, which is already landing on the develop branch.
LeanFrontier: the test was a submission (Field Note 09)
Field Note 09
— “The Test Was a Submission” — records a deliberately ordinary theorem
contribution exercising the freshly upgraded receiver end to end: local
report, trusted validation, observation, catalogue, and automatic merge. The
submission behind it is
PR #169, adding
natDegree_eq_zero_of_comp_X_add_C_eq_self, the natural-degree consequence of
the existing polynomial translation-rigidity theorem, with the receiver report
returning accepted (the note itself landed as
PR #171). The corpus is at
carlok.github.io/LeanFrontier.
forgepulse: repos above the fleet's median clone count
forgepulse now highlights repositories whose clone counts sit above the fleet’s median, and the comparison was fixed to use each repo’s own daily median against the fleet’s rather than its raw total, so a busy repo can no longer skew its own marker.
teorth/equational_theories: a PR correcting the paper's section 5.1 counts
Contributed to teorth/equational_theories a pull request correcting the paper’s section 5.1 brute-force counts: the two stated addends, 13,345,053 and 415,293, sum to 13,760,346 rather than the printed 13,632,566, and the “96.3% of the false ones” figure describes the first addend, not the total. The corrected numbers come from the project’s own All4x4Tables README, fixed earlier in #1335; the analysis is disclosed as AI-produced under human direction and reproducible from parsimagma.
dratify: check DRAT/DRUP proofs of unsatisfiability in-process
dratify is a new public DRAT/DRUP proof
checker that runs inside your Python process: zero dependencies, no subprocess,
no compiler. It replays a solver’s derived-clause log to confirm the empty
clause really follows, sharing no code with the solver that wrote the proof,
and the same checker is published as a Rust crate. A register_native() seam
lets the Rust engine (~18x faster) be supplied without shipping a toolchain,
publishing is tokenless via Trusted Publishing, and the checker’s own tests
went from 12 to 71 with a coverage gate.
cdclkit: a readable CDCL SAT solver where every answer comes with a certificate
cdclkit is a new public CDCL SAT solver, preprocessor and encoding library, written from scratch in readable Python with an optional Rust engine (roughly 18x faster) shipped as cdclkit-native. Every answer comes with a certificate: a satisfying model is re-evaluated clause by clause, and an UNSAT answer emits a DRAT proof that an independent checker replays. It publishes on PyPI with no third-party code, and the tutorial walks one problem end to end, from DIMACS through the files between stages.
parsimagma-greedy-cover: a refuter published on the SAIR Contributor Network
parsimagma-greedy-cover
is now public on the SAIR Contributor Network, entered into the Mathematics
Distillation Challenge Stage 2. It is a deterministic refuter with no LLM calls:
given two equational laws it finds a finite magma satisfying the first and
violating the second, and emits a Lean 4 certificate the organisers’ judge
accepts under plain decide. The point of publishing is the ordering rather
than the score — four instances of Z/2 refute 11,871,871 of the ETP’s
13,855,357 false implications, twenty-five refute 97.2%, and x ◇ y = 7x + 7y
over Z/13 alone refutes 268 pairs that Vampire’s fmb, Mace4 and z3 on a
ground encoding all fail to find. The honest number is on the item: 55 of 200 on
the graded distribution, zero on the true half by design, and zero on the
extra_hard category, which is selected against exactly this method.
parsimagma: a coverage engine over the Equational Theories Project law set
parsimagma is a new public signature and coverage engine over the 4,694 equational laws of the Equational Theories Project, asking which ETP constructions cover which separations. Its central finding is that the residual is not uniformly hard: at least 411 of the 1,062 Vampire-unresolved implications have finite countermodels on 9 to 32 elements, found in seconds by a structured algebraic sweep where three SAT/SMT-style solvers all fail from carrier 11 upward. The repo also argues the published count of unresolved implications sits below a provable floor and independently reproduces corrected figures for the paper’s section 5.1.
parsimagma: the hard core splits exactly, and a coverage number that was understated by its own denominator
The Equational Theories Project’s completed implication graph turns out to be
fetchable: finite_graph.json and graph.json are build artifacts the site
serves and the repository does not track, and decoding them reproduces the
project dashboard exactly, recovering its two remaining open cells,
(677, 255) and its dual, without being told. That splits the 1,062
Vampire-unresolved implications precisely — 610 require an infinite model,
450 have a finite counterexample, 2 are still open — which settles the
question parsimagma had filed as
unanswerable in issue #1474,
and corrects its own headline: 610 of those 1,062 admit no finite counterexample
at all, so the corpus reaches 411 of 450, not 411 of 1,062. Same measurement,
39% or 91% depending on which denominator you bother to compute. The earlier
claim that the circulating figure of 310 sits below a provable floor is
withdrawn: counted up to duality the same set is 316, so 310 looks like a
dual-class count against an earlier snapshot rather than an error. The graph is
also Lean-verified, so it can contradict the engine, and does not: 790 of 790
finite witnesses agree, and 19,392 order-5 laws extracted from a fork’s branch
agree too.
moebius-transcendental-lean: the conjugation degree on the transcendental locus
moebius-transcendental-lean is a new Lean 4 + Mathlib formalization, now public, of the conjugation degree δ(z) = [Q̄(z, conj z) : Q̄(z)] on the transcendental locus ℂ ∖ Q̄, following the companion paper p19.tex. It is archived with a Zenodo DOI and shipped as v0.1.0 and v0.1.1 with CI and a permanent axiom-verification module. The δ = 1 stratum is the subject of the companion diaz-modulus-lean.
teorth/equational_theories: 411 Vampire-unresolved implications have finite models
Contributed to teorth/equational_theories an issue reporting that the 1,062 Vampire-unresolved implications are not uniformly hard: at least 411 of them have finite countermodels on 9 to 32 elements. The analysis is reproducible from parsimagma and disclosed in the issue as AI-produced under human direction.
diaz-modulus-lean: a Palomar submission surface on its palomar branch
diaz-modulus-lean gained a
Palomar submission surface
on its palomar branch, parallel to the erdos-straus-offset-lean one: an
axiom-free core, a Challenge.lean that imports full Mathlib so the statements
match on the nose, and a companion note with attribution and search record. The
repo also picked up an Apache-2.0 licence at the root.
erdos-straus-offset-lean: a Palomar submission surface, checked by two kernels
erdos-straus-offset-lean now
carries a Palomar submission surface on its palomar branch: a
Challenge.lean statement surface, a Solution.lean with the proved
counterparts, and a comparator that audits the pair, pinned to Lean/Mathlib
v4.32.0. The Comparator result recorded in formalization.yaml passed on both
kernels — Lean’s default kernel and the independent NanoDa kernel accept the
four compared theorems, with the axiom closure limited to propext,
Classical.choice, and Quot.sound. The metadata file also declares the
automation method as an agent (Claude Opus 5 High via Claude Code).
caciarabot: config example file and mood-range daily thoughts
caciarabot stopped tracking its live
config: the working bot.jsonc is no longer in the repo, and a
bot.jsonc.example ships instead, so local state can’t be clobbered by updates
or leak into history. The daily thought generator also gained a real mood range
and occasionally follows a Wikipedia rabbit hole instead of picking a templated
topic.
forgepulse: routing and link-click fixes
forgepulse fixed two web-UI bugs: opening a repository detail page directly, or in a new tab, no longer 404s, and modifier-clicking a link now opens it in a new tab instead of swallowing the click. CI was retriggered after a GitHub Actions outage stalled the pipeline.
hrodrig/gghstats: selectable clone statistics merged upstream
The pull request adding selectable daily clone statistics to hrodrig/gghstats was merged upstream: the “Unique” line on the clones-over-time chart and the compact daily clone panel are now part of the project.
forgepulse: GitHub links on detail pages, gghstats issues checked
forgepulse’s repository detail page now links its title, top referrers, and popular paths straight to GitHub (paths always, referrers only when they look like real hostnames), and the JSONL export filename embeds a UTC timestamp so re-downloads stop silently overwriting the previous file. A cross-check against the issues opened on hrodrig/gghstats found that three of them applied to forgepulse too — the export filename, the stats-panel jargon, and the rank column header — and were fixed here; the locale-formatting issue does not apply in the same shape yet.
LeanFrontier: nobody chose the version (Field Note 08)
LeanFrontier’s Field Note 08 records the corpus moving to Mathlib v4.33.1 on schedule, with the version chosen by nobody: the upgrade pipeline rebuilt the fingerprint index, re-audited all 114 entrypoints, replayed the kernel, and opened the pull request itself — a person only merged it. Getting there took three stacked defects, each invisible until the one before it was fixed, the best being a validator that refused the upgrade over bytecode it had compiled into its own working tree while running. A fourth outside contributor also sent in a submission, still open pending a rebase onto the new toolchain, and no acceptance was weakened by the upgrade.
forgepulse: launch-week fixes and features
forgepulse got its launch-week pass: the repository table is paginated, a favicon landed, and the static “LOCAL” badge now shows the running git ref. On the ops side, a scheduled sync that was silently failing (the /data volume was never chowned to the app user) and a repository detail page that replayed every historical referrer day are fixed, and the production container now restarts itself after a crash or reboot.
solids-hunter: pre-release polish on the boolean-rule hunt
solids-hunter, the first-person Babylon.js game where you hunt solids defined by boolean rules, is in pre-release polish: full gamepad menu control, a working HUD speaker, spawn facing derived from the room, and rule complexity that ramps with each clear. A pre-release review surfaced five launch blockers, now fixed, gameplay screenshots landed in the README, and the shadow-pass watchdog now halves its cadence and can shed shadows entirely when quality is at risk.
carlok/LeanFrontier was forked by giacomoleonzi
carlok/LeanFrontier was forked by giacomoleonzi on 2026-08-24. In this corpus a fork is not a copy but the way in: a submission arrives as a branch on somebody else’s fork and is judged there, which is what Field Note 10 was about. This fork is the one the Tribonacci sequence submission later arrived through (PR #167). The corpus is at carlok.github.io/LeanFrontier.
forgepulse: self-hosted GitHub traffic-history analytics
forgepulse is now public: a self-hosted GitHub traffic-history analytics app built with Rust, a Svelte frontend, SQLite for storage, and Podman for deployment. It keeps a running history of GitHub traffic metrics so trends accumulate instead of vanishing after GitHub’s rolling window.
caciarabot Phase 3: admin commands and reply fixes
caciarabot grew a Phase 3 admin surface: /sleep, /wake, /categories, /stats, and /reload let moderators pause the bot, inspect its categories, and reload state without touching the container. Two reply bugs were also fixed — a cited reply and a word-trigger image can now fire together, and the LLM prompt no longer drifts into recurring ants/insects imagery.
contributed to TauCetiProject/TauCeti: an import-density measurement
Opened issue #3954 on TauCetiProject/TauCeti, the machine-generated Lean 4 library, to thank its maintainer and share a measurement: replaying its public history shows internal import density rising from 0.40 to 1.55 per module. The report is the courtesy side of the lean-corpus-density analysis, which used TauCeti’s history as evidence that machine mathematics accumulates.
LeanFrontier: nobody clicked merge (Field Note 07)
LeanFrontier’s Field Note 07 documents the day the maintainer left the merge path: receiver-accepted submissions now merge themselves, with an allowlist read from the default branch so no submission can approve itself, and four have landed with leanfrontier-receiver[bot] as the merging identity — two of them four minutes after opening. Two new theorems entered the corpus under the unattended gate: translation rigidity for characteristic-zero polynomials (over a characteristic-zero integral domain, a polynomial invariant under translation by one nonzero element is constant) and the finite pairwise squared-difference identity (in any commutative ring, the sum of squared differences over all ordered pairs is twice the cardinality times the sum of squares, minus twice the square of the sum). The note also pre-registers the launcher A/B experiment — two launchers differing in exactly one paragraph, eighteen accepted submissions per arm, and a stopping rule fixed in advance — with the first three arm-carrying submissions already in.
contributed to hrodrig/gghstats: selectable clone statistics
Contributed to hrodrig/gghstats with a pull request that adds selectable daily clone statistics to the repository index: a second “Unique” line on the clones-over-time chart with its legend enabled, plus a compact daily clone panel. The change makes unique cloners visible alongside total clone events.
dash: a serverless DMARC aggregate-report parser for Gmail
dash parses DMARC aggregate reports in a serverless setup for Gmail: it extracts and parses incoming reports, enriches the failing sources, and emails a summary of which senders are failing DMARC and why. The repo is public and actively developed.
LeanFrontier: a neighbour answers half the question (Field Note 06)
LeanFrontier’s Field Note 06 measures the corpus against Tau Ceti, a second machine-generated Lean 4 library, and finds that machine mathematics does accumulate: Tau Ceti’s internal import density rose monotonically across its history, narrowing the open question to whether accumulation survives without a human-written roadmap. The day also took the human out of the merge path — receiver-accepted submissions now merge themselves, and the first unattended submission opened at 15:39 and merged at 15:56:56 — and admitted conjectures as Prop-valued definitions fingerprinted by value, so a conjecture restating known mathematics is rejected as a duplicate. Three new submissions landed (thue-morse-prouhet-power-sums, stern-brocot-coprime-enumeration, padovan-sequence-sum), taking the corpus to 27 modules, and running the pipeline unattended surfaced four defects that reading had missed. The measurement itself lives in the new lean-corpus-density repository.
lean-corpus-density: does machine mathematics accumulate?
lean-corpus-density is a new, reproducible measurement of dependency density in Lean 4 corpora, human and machine-generated. Replaying Tau Ceti’s commit history shows its internal import density rising monotonically as it grew — 0.40 edges per module at ten modules up to 1.55 at 2,314 — evidence that machine-generated mathematics builds on itself rather than merely piling up. The whole analysis is reproducible from file headers; no build is required.
diaz-modulus-lean: a formalized negative result on Diaz's modulus conjecture
diaz-modulus-lean is a new Lean 4
formalization of a negative result on Diaz’s 2004 modulus conjecture: for a
candidate u with e^u and |u| both algebraic, the conjugate ū is a
rational function of u with algebraic coefficients, so no statement about
vanishing matrix coefficients over the algebraic hull can separate a candidate
from an ordinary complex number. It machine-checks the conjecture’s question of
method — how non-holomorphic maps like conjugation and modulus could enter a
transcendence proof at all — in the negative direction.
cross-tetris: a shared-queue four-well Tetris
cross-tetris is a new game: four standard Tetris wells arranged in a cross, sharing a single piece stream. You — or a greedy rule-based AI — pick which well each piece falls into, the wells play out as ordinary real-time Tetris, and the game ends when any well tops out. The engine is Rust compiled to WASM with a React UI, and it is playable live.
LeanFrontier: first theorem submissions land, verified through the kernel
LeanFrontier accepted its first batch of machine-generated theorem submissions: the Furstenberg topology on the integers, the tent map semiconjugate to the logistic map, Lucas numbers and their Fibonacci bridges, the Fibonacci Q-matrix, and a closed form for the Josephus problem’s survivor. Submitted modules are now replayed through the kernel before acceptance, and the whole corpus is rechecked whenever Mathlib is upgraded.
LeanFrontier: five new submissions in a day, from sequences to transcendence
LeanFrontier accepted five more machine-generated submissions in a single day: a greatest sequence below a ceiling with bounded steps, how the greatest-step bounded minorant responds to its constraints, Stedman’s plain changes ringing a full extent, Hermite-Lindemann implying that exp is injective on algebraic numbers, and a generalization of finite-group character sum vanishing to noncommutative rings. The receiver was hardened along the way — submissions now authenticate as the LeanFrontier Receiver app, and the corpus dependency graph is published in the catalogue. The whole day is documented in Field Note 05.
cold-path-server-podman: the Cold Path game server in a container
cold-path-server-podman is a
new repository that runs the Cold Path
multiplayer game server inside a Podman container, so no LuaJIT, LuaSocket, or
LuaSec install is needed on the host. The upstream sources aren’t vendored —
the image build pulls them from GitHub, and make build / make run / make logs
handle the rest.
spriter: raster images to pixel art, entirely in the browser
spriter is a new browser-only tool that turns a raster image into a pixel-art sprite. There’s no upload, no build step, and no dependencies — the whole conversion happens locally in the browser.
LeanFrontier: machine-generated mathematics, verified by the Lean kernel
LeanFrontier is a new Lean 4 library of machine-generated, kernel-verified mathematics. Every theorem it contains has been checked by Lean’s proof kernel rather than taken on faith, which means the machine-generated results carry the same formal guarantees as hand-written proofs.
carlok/LeanFrontier was forked by matteodesimone and rlacroce-rdi
carlok/LeanFrontier was forked by matteodesimone and rlacroce-rdi on 2026-08-18. A fork is this corpus’s intake path rather than a copy: a submission arrives as a branch on somebody else’s fork and is judged there, so a new fork is the outward sign that someone intends to try. Both used it that day — rlacroce-rdi for the Josephus survivor closed form for step two (PR #55, merged the same afternoon) and Matteo De Simone for a run of submissions from the Fibonacci Q-matrix to the Furstenberg topology on the integers (PR #56 onward). The corpus is at carlok.github.io/LeanFrontier.
inversive-geometry-lean: circles and lines as one object
inversive-geometry-lean is a new Lean 4 library for generalized circles — “circlines” — that treats circles and lines as a single object. Each circline is cut out by a Hermitian equation, giving a uniform treatment of inversive geometry inside a proof assistant.
euclean: can a machine recover structure from an anonymized theory?
euclean is a new research project asking whether a machine can recover mathematical structure from an anonymized formal theory armed only with a proof checker. The experiment strips away all the human-readable names and intuition, leaving just a formal theory and the kernel’s verdicts to work from.
caciarabot: an Italian-first Telegram group bot
caciarabot is a new self-hosted, reactive Telegram group bot designed Italian-first. It runs on your own infrastructure rather than a hosted service, and it reacts to group activity rather than only to explicit commands.
A public activity log
This blog is a running, public log of activity across my GitHub repositories.