Proof: how PYRION is checked
Claims are cheap. This page shows the checks behind PYRION and where to verify them yourself. Every number below is read from a file of the repository when this site is built; none is typed in. Each item says what it means in plain words first, then links to the file for experts.
Questions we threw at it
A scripted "judge battery" asks the app 42 questions, including 11 traps (a gauge pressure, an oxygen partial pressure given as total pressure, harmful requests) that it must refuse or handle correctly. The latest run recorded 0 failures, started 2026-11-15 08:39 UTC.
Result file: 20261115T084542Z.json. Questions: questions.json.
Every value is found in the NASA source, by code
PYRION extracted 10,303 records into the files (the engine uses the ones its validator does not reject; see /data). A validator then looks each one up in the source file: 8,529 table cells at their row and column and 1,121 quotations word for word, across 132 text files. A value it cannot find is rejected.
Report generated 2026-11-15 09:00 UTC: latest.md · latest.json · run_all.py
What the validator found wrong
Findings are graded. 0 blockers (would invalidate a result), 0 major and 134 minor (wording or notes). The minor ones are listed openly in the report rather than hidden.
Reviews by separate AI reviewer agents
Separate AI reviewer agents, run apart from the sessions that wrote the code, attacked the language-model layer in rounds. They are not NASA review or expert validation. Each round lists what they found and how it was closed. Sections of the audit report:
- Live tests
- Findings
- Action
- Round 2 (re-check of the corrected items only) — closed
- Round 3 (full re-audit after the open notes were resolved) — November 14, 2026
- Round 4 (deep audit, two independent reviewers at high effort) — November 14, 2026
- Phase 1B — NAVIGATOR (guiding dashboard) — closed November 14, 2026, 21:40
Automated tests
The engine has 344 declared test cases and the web app 207, counted from the test files. They cover units, regimes, the four verdicts, determinism and the safety checks. Counting declarations is not the same as running them: they run before every push. engine/test · web/test
Predictions sealed before the experiment
Before FM2 has flown (it is a planned experiment), PYRION wrote down what it will and will not say about it and recorded the fingerprint (SHA-256) of that text. The fingerprint is anchored in the Bitcoin blockchain at block 975,854, read from the proof file itself. Anyone can check it without trusting GitHub or us:
git clone https://github.com/Lucyyngceo/pyrion
cd pyrion/predictions
ots verify SHA256SUMS.otspredictions/README.md · SHA256SUMS · SHA256SUMS.ots · tool: opentimestamps.org
The code existed when we say it did
Git dates can be edited, so the repository state is also stamped with OpenTimestamps from time to time. Each stamp records a commit and a time:
- 2026-11-15T0500Z-head.txt: commit 34c5c78b47, 2026-11-15T04:43:56Z, Bitcoin confirmation pending
How AI was used
The full declaration of the AI tools used to build PYRION, and what each did, is in the project README under "Use of AI (SI) tools". Every AI-assisted commit names the model in a Co-Authored-By line, so the trail is in git. Read the declaration