Big O Speedrun data

Checked 2026-10-08. Local static source bundle. See NOTES.md for source links, tier limits, date rules, and open questions.

Read registry.json, then each problems[].data_file. Use each record's exact rational score, its tier, and record_metadata[id]. eligible_for_record governs record-holder labels. No site proof badge is available. Keep separate assumption tracks separate.

Run:

python3 validate.py

The installed jsonschema package is required. validation-summary.txt has the dated result.

submissions.json preserves all observed CrocSwap PRs, including nonnumeric work. chart-cross-check.json records the earlier 42-point chart comparison. sources/ retains evidence and source notices. manifests/ contains the read-only listing manifests. schemas/ contains the design schemas and data-wrapper schemas.

Scheduled source refresh

Added 2026-10-09. python3 refresh.py reads both integer-multiplication repos with gh api --method GET. It never runs their code. The current GitHub CLI login needs read access. The script resolves each default branch to a commit, fetches the README and selected record/certificate, and reads all CrocSwap PRs. It keeps reviewed main-branch checkpoints separate from PR merge status. Unknown headline formats remain unscored submissions for review. Swapnil stays on its separate comparison track. Source checks do not grant a proof badge.

The script locks against overlapping refreshes, validates a staging copy, checks for concurrent edits, and replaces changed files only after validation. It stops on API, format, certificate hash, or validation errors. A source fingerprint makes an unchanged snapshot a no-op. The JSON diff summary is printed and saved in .git/refresh-summary.json. build_data.py is the historical snapshot builder; do not use it for scheduled refreshes.

python3 pivot_watch.py searches a rolling 72-hour window by creation and push time, plus all-time CrocSwap forks and README citations. It fetches the OpenAI preprint directory each time and searches the algorithmic problem families found there. The ranked results, timestamps, stars, query coverage, errors, and source signals are in pivot-watch.json. These are discovery guesses, not verified new races. GitHub search indexes can lag; capped queries are listed explicitly.

The active three-hour watch follows WATCH_RUNBOOK.md. It runs the people and pivot watches, checks site refresh health, and commits only watch reports. The scheduler supplies the checkout directory and local notification routing; no previous Git commits are needed. poll.py is a legacy adapter and must not run alongside the cronjob. Its local routing configuration is .git/watch-office.json; this file contains no credential values.

Offline checks: python3 -m unittest test_refresh.py and python3 validate.py.

Production uses bigospeedrun/site, with this bundle tracked under data/. Run python3 data/refresh.py and python3 data/validate.py from that checkout. The refresh finds the containing Git repository and keeps its lock, staging files, and summary in that repository's Git directory. It checks edits only inside the data folder. The validator checks receipt paths within the bundle; historical provenance paths are not opened. poll.py is not needed in GitHub Actions.

registry.updated_at and each problem's updated_at are required UTC date-times ending in Z. They record the last normalized data update, not a proof check or a fresh check of every source. An unchanged refresh preserves these fields. The integer-multiplication refresh updates that problem and the registry only. Other problems retain their own update times. The initial timestamp migration uses the retained intake time for integer multiplication and the last data commit time for the other problem files.

Exact DFT research entries were added on 2026-10-09. score_relation is optional: absent or attained means the scored value; strict-upper-bound means every fixed saving below the stored endpoint. Render the latter with <, including in tier summaries. The endpoint itself is not an attained pure-log saving. OpenAI's verified frontier comes from its smaller Lean tensor network and the polylog absorption lemmas, while its selected Comparator headline is 10^-13. The rerun remains pending. The two community drafts are claimed, with separate source tracks, credited notes and their exact model statements. They are not open submission races. Shea's unlicensed manuscript is linked; the bundle keeps a derived evidence summary and source hashes.

pivot-watchlist.json names repos that every watch run checks directly, even when search indexing or the 72-hour window omits them. watched_repositories in the report records pinned heads and detects new commits. The scheduled watch reviews new heads for likely pivots. python3 pivot_watch.py --watch-only checks those heads and rechecks retained candidates without repeating the broad search. A new head triggers review; it does not silently replace a mathematical claim.

The three-hour watch runs python3 people_watch.py. Its configuration is people-watchlist.json; people-watch.json stores checked identities, seen public event IDs, ranked leads, and coverage warnings. Named people and the ten most active CrocSwap PR authors are checked each run. The first run covers the last 72 hours; later runs deduplicate IDs. New repos and forks, pushes outside the registry/direct watch list, and newly opened bound PRs are leads only. Profile/README evidence identifies GitHub accounts. X handles supplied by Nil remain separate from confirmed profile links. This watch does not access X or change mathematical records automatically.

X watching is disabled in the scheduled watch. The optional x_watch.py reads x-watchlist.local.json, an ignored local file. Start from x-watchlist.example.json, which has an empty accounts list, if a separate operator authorizes X access. The script uses only tabs, goto, and text; it restores the original URL and keeps a local restoration journal.

Private state and reports are in the containing Git directory as x-watch-state.json and x-watch-report.json; they are not site data. The watch reads at most 12 rendered posts per profile. Text-only reads cannot expose hidden URLs, exact hidden timestamps, or unloaded posts. The watch retains visible full links, time text and date ranges; it marks uncertain newness rather than inventing an exact timestamp. It preserves the last complete check time when a profile read is partial or no posts are readable. A missing article ends that profile read and the watch continues with the next account. Unexpected profile navigation stops the watch and restores the original tab. There is no X API access.

Only GitHub API sources and arXiv metadata are checked automatically. Other PDF links remain for agent review. Alerts identify changed research sources for review; an X post never supplies a normalized score. The separate site GitHub Actions importer validates data before publication. New unsupported sources require the normal source review before any data change.

Browser safety and parsing checks: python3 -m unittest test_x_watch.py.

Pivot alerts suppress repositories already in the board registry unless their configured README headline saving changes. pivot-watchlist.json contains the bound patterns. The watch stores exact rational values in board_bound_states, using the latest imported score for the first comparison. Commit, star, prose, and verification changes alone do not alert. Missing headline patterns are logged as coverage errors; they do not count as changed bounds. Known board repositories without a configured bound pattern are suppressed. New repositories still use pivot discovery. This filter does not alter the separate people-watch event feed.

As of 2026-10-10, people discovery also handles first-branch CreateEvents and polls each watched user's owned repositories by last push. The inventory provides a second path when GitHub emits no repository-create or push event. It is capped at 300 repositories per user; caps are logged. The classifier migration rechecks previously ignored branch events once. threesum is an accepted problem spelling. Pivot discovery searches names/descriptions separately from README text and feeds relevant people-watch repositories directly into candidate review. Broad search caps remain visible. A watched board repository can configure bound_patterns for more than one problem; a change to either exact value triggers review.

Public snapshot

Prepared 2026-10-12 for bigospeedrun/watch. This tree has one initial public commit. The local private-history branch retains the earlier history and must never be pushed. Push only main; never use --all or --mirror. Source snapshots with removed email addresses or local provenance paths are listed in public-sanitization.json. Their public hashes refer to the scrubbed bytes. These copies are not byte-exact upstream receipts; original evidence remains in private history. The historical builder reads the vendored sources and retained chart digest; it does not require private research notes.