Every release of ChipletOS and ChipletOS Photonic Signoff runs through an independent 22-check validation suite before it ships. 21 of 22 checks pass today; 1 is on the roadmap. We publish our retraction history, our cross-solver disagreement envelope, and our raw literature cross-check numbers. This page is the place to verify the engineering rigor before signing.
Each row below is an independent check. The runner is open to NDA-signed buyers under bash scripts/audit/verify_gates.sh. Witness files are signed with content hashes and surfaced in every release artifact.
| # | Check | Subject | Status |
|---|---|---|---|
| 01 | Surrogate accuracy ≥ 99% R² vs reference solver | Chiplet impedance + 5 photonic primitives | PASS |
| 02 | Conformal 95% confidence interval on every prediction | All ML models | PASS |
| 03 | Out-of-distribution detection per prediction | All ML models | PASS |
| 04 | Drift sentinel — no retracted citations slip through | Every commit | PASS |
| 05 | Model registry integrity check | Every release | PASS |
| 06 | Promoted-model audit (alias matches manifest) | Every model promotion | PASS |
| 07 | Route-backed end-to-end signoff harness | 60-case sweep, RF pass rate + mean / max worst-case | PASS |
| 08 | Cross-source consistency check | Every cited claim resolves to its source | PASS |
| 09 | Conformal coverage independent witness | Per-regime empirical coverage matches calibration record | PASS |
| 10 | Synthetic-data detector | 0 findings across production corpora | PASS |
| 11 | Phantom-citation detector | 0 stale citations beyond audited baseline | PASS |
| 12 | Golden-kit witness freshness | 5/5 reference geometries match live solver within 1% | PASS |
| 13 | API contract drift detector | Route surface vs frozen baseline | PASS |
| 14 | Thermal calibration head — mean abs err ≤ 5 K | Chiplet multi-physics | PASS |
| 15 | Crosstalk calibration head — mean abs err ≤ 3 dB | Chiplet multi-physics | PASS |
| 16 | Power-delivery calibration head — mean abs err ≤ 10 mΩ | Chiplet multi-physics | PASS |
| 17 | Mechanical / warpage calibration head — mean abs err ≤ 5 µm | Chiplet multi-physics | PASS |
| 18 | Photonic surrogate accuracy (≥ 99% R²) | 5 of 6 primitives clear the bar; waveguide on analytical | PASS |
| 19 | Photonic design-rule check (AIM-class rules) | 8 rule families | PASS |
| 20 | Cross-solver agreement (independent 3-way validation) | Cross-physics check on the roadmap | SKIP |
| 21 | Published-paper cross-check (5 silicon-photonic references) | Within analytical-model expected band | PASS |
| 22 | Adversarial robustness harness | 108-case sweep across all 6 primitives | PASS |
Every claim we have ever walked back is logged with the original number, the corrected number, the date of correction, and the audit-infrastructure check that caught it. We have logged 39 such retractions since the platform was founded. We believe this is a competitive moat: any competitor unwilling to publish theirs is hiding them. The full registry is on request under NDA as part of due-diligence packets.
Our analytical photonic models are independently cross-checked against five published-paper reference points (Bogaerts 2018, Pavanello 2020, Lim 2014, Selvaraja 2010, Xu 2017). The result is reported as a raw closed-form mean absolute error, not a post-calibrated number. We declined to ship per-material calibration scalars that would have collapsed the error to under 1% because that would have been test-fitting against the same five papers the check evaluates.
Our chiplet impedance signoff has been independently cross-validated against BEM, FastHenry2, OpenEMS, and a full-wave finite-element solver across 100 reference geometries. The disagreement envelope is published as a witness file. No competitor publishes a comparable cross-solver matrix.
When a method does not work, we publish what failed and why. Examples: the waveguide primitive runs on a closed-form analytical model because our AI-surrogate refresh has not yet cleared the 99% R² bar; an independent cross-solver agreement check is currently on the roadmap rather than live. These are the kinds of disclosures buyers should look for in any vendor's signoff stack — and demand if absent.
bash scripts/audit/buyer_verify.sh runs the validation suite + 110-geometry adversarial sweep in ~2 minutes against a checked-out source tree. NDA access via founder.