# v2.2 repeated balanced cross-fitting status

Date: 2026-07-17  
Decision: **scientific execution and independent verification complete**

## Outcome

The fold-assignment sensitivity requested by the focused v2.1 peer review is
complete. The frozen seed `20260717` primary realization was retained, and 19
mechanically prespecified consecutive seeds were added. All 20 realizations had
five non-overlapping folds of exactly 432 targets, with one OOF record for each
of 2,160 targets and each of four models.

Independent verification passed **120/120** checks, including input and seed
locks, fold balance, target uniqueness, numerical summary recomputation, fresh
reproduction of the frozen primary run, hierarchical interval recomputation and
100 seed-output hashes.

## Prespecified interpretation

| Endpoint | Favorable realizations | Across-realization mean | Joint target + fold/training 95% sensitivity interval | Classification |
|---|---:|---:|---:|---|
| RMSE difference vs zero | 20/20 | −0.00629188 | −0.00735218 to −0.00524033 | `STABLE` |
| RMSE difference vs train mean | 20/20 | −0.00046627 | −0.00070221 to −0.00022488 | `STABLE` |
| Raw-cosine difference vs train mean | 20/20 | +0.00122360 | −0.00021601 to +0.00249840 | `MIXED` |
| Perturbation-specific residual cosine | 20/20 | +0.04528089 | +0.02964997 to +0.06117822 | `STABLE` |

The safe model had favorable point estimates for both RMSE contrasts in 20/20
realizations. No negative control had point RMSE support against both baselines
in any realization. All folds selected alpha 1000.

The raw-cosine point estimate was positive in 20/20 realizations, but its joint
sensitivity interval crossed zero. Therefore the frozen seed's conditional
Holm-adjusted P=0.00356 is not sufficient evidence of fold-robust raw-direction
support. The defensible result is stable, modest RMSE improvement with mixed raw
directional robustness.

## Fold-assignment behavior

Pairwise fold-assignment adjusted Rand index had median 0.0001 and range −0.0014
to 0.0030. Thus the seeds created effectively distinct balanced assignments,
rather than repeatedly evaluating nearly identical partitions. Stable RMSE
effects were observed despite this reassignment.

## Reproduction-gate amendment

The first run stopped before seed `20260718` because an initial `5e-12`
row-level tolerance rejected float32/BLAS summation differences. Gene/fold
assignments were exact; maximum differences were `8.13e-05` for target-level
vector norms and `4.98e-08` for model summaries. Before any sensitivity result
was generated, the tolerance was amended to `1e-4` for target metrics and
`1e-6` for CV/summary metrics, while all identifiers and fold assignments
remained exact. The stopped output and change log were retained.

## Evidence boundary

This analysis adds split/training sensitivity within one processed HepG2
dataset. It is not independent biological replication and does not establish
cross-cell-type, muscle/DMD, mechanistic, therapeutic, safety or clinical
validity.

## Authoritative objects

- Protocol: `submission_nar_database/evidence_upgrade_v22/00_repeated_balanced_crossfit_protocol_20260717.md`
- Seed manifest: `submission_nar_database/evidence_upgrade_v22/01_repeated_balanced_seed_manifest.tsv`
- Debug log: `submission_nar_database/evidence_upgrade_v22/02_reproduction_tolerance_debug_log_20260717.md`
- Results: `results/evidence_upgrade_v22/hepg2_repeated_balanced_crossfit/`
- Main report: `REPEATED_BALANCED_CROSSFIT_REPORT.md`
- Independent verification: `INDEPENDENT_VERIFICATION.md` and `independent_verification.tsv`

## Required integration before a new release lock

1. Add the compact v2.2 repeat summaries, joint intervals, fold counts, alpha
   selections, pairwise fold audit, target-effect stability and verification
   tables to the SQLite/TSV/JSON candidate resource.
2. Revise manuscript wording so the frozen primary P value is explicitly
   conditional and raw-direction robustness is `MIXED`.
3. Add a compact supplementary stability panel or table.
4. Rebuild the candidate archive and release lock only after these objects and
   manuscript wording pass cross-layer verification.