# Reproduction-tolerance debug log

Date: 2026-07-17  
Scope: seed `20260717` reproduction gate only; no new sensitivity seed was run.

## Trigger

The first execution stopped at the prespecified frozen-primary hard gate before
seed `20260718`. Fold assignments reproduced exactly, but an unrealistically
strict `5e-12` row-level tolerance rejected float32/BLAS numerical differences.

## Observed differences

- Outer-fold manifest: exact, including all 2,160 gene assignments and 5×432
  fold counts.
- Inner-CV summaries: maximum absolute difference `3.69e-08`.
- Model-summary quantities: maximum absolute difference `4.98e-08` across the
  inspected summary fields.
- Target-score `model_rmse`: maximum absolute difference `1.38e-06`.
- Target-score cosine fields: maximum absolute difference `7.03e-06`.
- Target-score vector-norm fields: maximum absolute difference `8.13e-05`.

Differences were quantized at float32-scale increments, while categorical
fields, target identities, folds, alphas and denominators matched. The largest
summary difference was more than four orders of magnitude below the primary
RMSE effect versus train mean and did not change any displayed result.

## Outcome-blind correction

Before any sensitivity seed was executed, the reproduction tolerances were
changed to:

- fold manifest: exact (`0` absolute tolerance);
- target-level float32 metrics: `1e-4` absolute tolerance;
- inner-CV and model summaries: `1e-6` absolute tolerance;
- all text, model, gene, seed, fold and denominator fields: exact.

The target tolerance is below 0.001% of the primary mean RMSE. The runner now
records the maximum observed absolute difference for each reproduced file.

No seed, input, model, feature, fold rule, alpha grid, endpoint, aggregation or
interpretation rule was changed. The failed seed-20260717 output is retained and
will be reused only after it passes the amended numerical gate; no result was
deleted or silently overwritten.