# Leave-GFOD2-out sensitivity analysis v0.1

## Result

Removing GFOD2 from the 20-realization v2.3 benchmark leaves the central model-comparison conclusion unchanged. The corrected two-stage 95% intervals remain below zero for RMSE versus both zero and train-mean comparators. The response-module block analysis also retains both RMSE conclusions, while the two cosine intervals remain mixed, as in the full analysis.

- RMSE versus zero, two-stage 95% CI: [-0.006389, -0.005905]
- RMSE versus train mean, two-stage 95% CI: [-0.000371, -0.000224]
- Raw-cosine module-block classification: mixed_crosses_zero
- Residual-cosine module-block classification: mixed_crosses_zero
- Mean RMSE difference versus train mean after exclusion: -0.000303225 (change versus all targets -0.000001368)

## Prespecified structure

GFOD2 belonged to response-profile module 49. That module contains 39 targets after exclusion; the other 53 modules retain 40 targets. The original module assignment is preserved, and modules are not re-clustered or rebalanced.

## Candidate interpretation

The current action queue contains 20 remaining active candidates after GFOD2 is omitted. Their action categories and action-then-gene display policy are unchanged. No replacement top gene and no new numeric rank are created.

## Claim boundary

This analysis shows that the resource-level benchmark and current action queue do not depend on GFOD2. It does not establish that GFOD2 is biologically unimportant. The author-reported cross-project assessment places GFOD2 on provisional hold pending auditable quantitative evidence import; it does not yet support formal deprioritization.

## Reproduction

Run `python3 scripts/build_gfod2_exclusion_sensitivity_v01.py`. The full-scope bootstrap is recomputed and must match the canonical v2.3 serialized interval values within an absolute tolerance of 1 × 10^-15 before the leave-out result is accepted. This tolerance covers only the final decimal-to-binary round trip; it is not a statistical tolerance. See `gfod2_exclusion_manifest.json` for seeds and SHA-256 provenance.