Four provisions you must be able to prove before you let it act.
governance · safety · provenance · continuity
The finding
1,301 of 1,312
cells have no measurement in any known benchmark.
99.2% blind · 349 enumerated provisions · 4 axes
The gap map
Which provisions of obligation-space have no measurement — by axis, by jurisdiction.
The refutation ledger
Seven of our own claims, published. Four killed our own bets. No precedent exists.
Verify a chain
Recompute the chain locally. Tamper-evidence, not authenticity — the label says what it does.
What the instrument has measured
| Claim | Measurement | Tag |
|---|---|---|
| Composed pipeline beats raw base | Δ +12.21 [+7.42, +17.00] | [MEASURED] n=195 |
| Deterministic gate carries most of it | +34.84 [+17.50, +52.18] | [MEASURED] n=31 |
| KB exact-match lookup helps where covered | +19.64 [+6.87, +32.41] | [MEASURED] n=14 lower bound |
| ProvBench: markings do not survive | 0 of 12 assets · one-sided 22.1% | [MEASURED] n=12 · asset |
| Care_cost (gpt-4o-mini) | 0.667 = 0.667 × (1 − 0.00) | [MEASURED] n=7 seed set |
| PQC signing is real | OpenSSL 3.6.3 · ML-DSA-65 · tamper→False · forgery→False | [MEASURED] |
Every figure above traces to a signed J-record. Every n<20 carries a lower-bound badge on the same line (audit criterion U4).
10 refutations, kept
Publishing what kills our own theses is the moat. A competitor can copy the corpus, the gate, the signing — and rebuild the harness — but not the discipline of publishing their own refutations.
Run the verify chain locally · Read the ledger · See the blind spots