# EXTERNAL BAR — `signoff-certificate-engine`

## The lab's own F2 line (verbatim, `CROWN_JEWELS_RESOLVED.md:159` at commit `cb6d9d7d`)

> **F2 — what would move the third-party axis to A.** Axis — (self-graded): self-graded: truth is a solver, corpus or kernel Genesis owns or authored (bar recorded: `n/a`). To reach axis A an external system Genesis does not own must produce this number against a bar pre-registered before the run, and its output must be cited. Nothing in this register does that yet — `PEER_REVIEW_2026-09.md:20-23` records axis A = 0.

## What an independent party would have to do to move this entry to axis A

1. **Pre-register the bar before the run.** The certificate's yield field must equal the exact Clopper-Pearson
   interval for its own k/n at its stated confidence, to 1e-12. The certified S-interval must contain every S-matrix
   in a Monte Carlo sweep of the stated perturbation ball (60 draws at seed 1030002026, or more). The v2 gate on the
   CI lower bound must be a subset of the v1 point gate. Seal the bar.
2. **Recompute it outside Genesis.** Take the certificate the witness writes
   (`benchmarks/certificate_v2/certificate_v2_gate_2026_07.json`). Recompute the interval with an implementation
   Genesis does not own (R `binom.test`, or statsmodels `proportion_confint(method="beta")`). Re-run the
   perturbation sweep with an independent MTL synthesis, or a circuit simulator, and cite that output.
3. **What they need.** A laptop; seconds. The data is the files listed in `INPUTS.sha256`. Tools: scipy (1.17.1
   here) and numpy.

## What would falsify the claim

- A certificate yield interval that differs from the textbook Clopper-Pearson bounds by more than 1e-12. This
  packet's `DEFECT.json` plants that, shifting the lower quantile by one success.
- A sweep draw whose S-matrix leaves the certified bound: an empirical worst case above the bound.
- A k where the v2 CI-lower gate passes but the v1 point gate fails.