Every number is machine-recomputed or signature-verifiable; where we
could not verify, we say so. The markdown original of this report is signed —
docs/reports/TRUST_REPORT_2026-09_EN.md(.sig) in our
repository.
Nautilus Assay independently recomputes published AI-agent performance claims from public artifacts, issues three-state verdicts (agree / disagree / not_computable), and signs every receipt with ed25519. Protocol published CC-BY — anyone can implement it or verify our receipts without trusting us. Operator: 伊洛科技有限公司.
| Target | Claim | Recompute | Verdict |
|---|---|---|---|
| Aider — SWE-bench Lite | 26.3% | 79/300, subset relations verified | AGREE |
| Terminal-Bench v1 — OpenHands | 41.25% ± 0.69pp | 165/400 = 0.4125 exact | AGREE |
| NanoJev (embodied) | 6 claims | incl. trajectory replay | AGREE |
Disclosures: Terminal-Bench's leaderboard excludes 5 incomplete trials from the
denominator (defensible, undocumented until now); the OpenHands row was self-flagged
verified: false — third-party recompute is what such flags await.
15 registered criteria, ratcheted (tighten-only with evidence chains): verdict recomputability, external-verified provenance, four adversarial gaming classes, paradigm corroboration chains, and calibration-claim verification — calibration claims of probability-output controllers are safety claims, recomputable on public holdouts.
30 items · 4 dimensions · public/holdout 18/12 · seed committed before generation. Release 2026-09-27 on the wall.
Misjudgment = OVERTURNED verdict of the same rigor + refund + challenge costs + track-record reset; 90-day window; anonymous verifications treated as UNVERIFIABLE (protocol §5b).
Open a recompute request (EN/中文). Self-attestation free forever; independent verification is what we sell.
Signing pubkey f7554b8709b7fe36f5a63e7f76cf2a31f827aee5724ff8dd4f11772d1aa3e8be · Next report: 2026-10 · ← The Wall