Checked18 November 2026, public default branches
MethodNew and changed projects were cloned fresh and their tests run. Unchanged repositories were compared with the commit each page was last reviewed at, and their live URLs requested.
Flagships
| Project | Status | Strongest current proof | Main open gap |
|---|---|---|---|
| MUTX | Shipped, v1.4.0 source, hosted service offline | Local API on SQLite with role checks and a working policy block. 58 policy, approval, RBAC and run tests pass | Hosted service down, published install paths stale, CI workflows disabled |
| Archon | Experiment, public verifier, desk trades on Hyperliquid testnet | 45 tests and a tamper demo pass on Python 3.10, 3.12 and 3.13 in CI, refusing both edited and re-chained records | No production record or verdict files are public, so no backtest or live result can be checked |
| Bollard | Shipped, notarised Homebrew cask, CLI and MCP server | Notarised and stapled v3.0.0 DMG accepted by spctl. BollardChecks --policy-only passes 16 policy checks in CI on 8da1282 | Verified only on macOS 27.2. Live and lsof-parity checks not rerun independently. Steady scan reran at 33.5 ms against the repository’s 9 ms |
Agent controls
| Project | Status | Strongest current proof | Main open gap |
|---|---|---|---|
| BoundaryBench | Verified, run from a clone, not on npm | 38 committed Claude Code traces re-score to the published table, including 5 of 5 budget-cap breaches | 3 to 6 runs per scenario, one model and one client. The trace is an unsigned file |
| Switchyard | Verified, Python package from source, console demo on Vercel | 5 stop and budget guarantees with property tests. 54 guard, 20 backend and 28 frontend tests pass, CI green on 34e9dbc | Not on PyPI, and the README’s one-venv install fails on conflicting pytest pins. The console replays one invented drill |
| Ambient Relay | Shipped, browser lab on Vercel, synthetic events only | 22 Node tests pass, the checked-in terminal proof matches the replay, and the live site serves a build of 456bcb9 | The Discord adapter has never connected to a real account, and the stations are keyword templates with no measured accuracy |
Local-first tools
| Project | Status | Strongest current proof | Main open gap |
|---|---|---|---|
| Tablebeam | Shipped, browser demo and Streamlit app | Pinned evaluation at f141ee7 reproduces byte for byte. 0.4% silent wrong against pandas 13.7%, SQLite 17.2%, DuckDB 11.7% | The low wrong rate comes mostly from refusing. The latest release, v2.0.0, predates the parser and the evaluation |
| Fieldwork plugins | Shipped, catalogue live on Vercel, not in the OpenAI plugin directory | 80 unit tests (2 skipped) and all eight catalogue demos pass against the real scripts at 16470b5 | Not listed in the OpenAI plugin directory, no GitHub releases, and no workflow deploys the Vercel site |
| World Watch | Experiment, local demo, hosted deployment paused | 314 of 315 offline tests pass (1 skipped) and the severity-gate evaluation reruns byte for byte | Gate measured on earthquakes only, with stand-in proposers. It does not check that a quote is about the same event |
Labs and benches
| Project | Status | Strongest current proof | Main open gap |
|---|---|---|---|
| Queueglass | Shipped, live on Vercel | BPI Challenge 2019 headline reproduces from the committed CSV. 19 tests and the browser smoke pass | Fixed capacity and every item through every step. No release tagged since v1.0.0 |
| Barter | Simulated, static lab on GitHub Pages | 22 engine tests pass, final journal hashes are pinned, and the SHA-256 chain checks with shasum | Three fixed journals only, the /lab deep link fails on Pages, and the gh-pages branch is published by hand |
| Silicon Shader | Shipped, GitHub Pages library and Python CLI | Per-frame intervals are published in every bundle, and recomputing them reproduces the README figures | One base M4, one flight per side, all bundles from one author |
Games
| Project | Status | Strongest current proof | Main open gap |
|---|---|---|---|
| Solar Drift | Shipped, GitHub Pages | 15 rule tests, including seed determinism. Pages serves the main bundle | The Vercel demo the README links is stale, and there are no browser tests |
| Loop Courier | Shipped, GitHub Pages and Vercel | 25 Node tests and 4 Playwright suites pass in Chromium and WebKit, with replay parity across frame rates | The economy test checks a simplified model, not the game’s own simulation |
| Terminal Starfield | Shipped, v4.0.0 release and web build | 128 unittest cases pass, and a seeded --snapshot is byte-identical across runs | Not on PyPI, so the README’s pipx install terminal-starfield fails |
| ckitty | Shipped, v1.1.0 binaries and web build | make check and make sanitize pass, and the rebuilt wasm is byte-identical to the committed file | No release or Homebrew formula includes the web build |