Skip to content

GenoLeWM-FX Borzoi residual model report

Generated by tools.research.fx_borzoi_residual at 2026-06-11T14:00:00Z.

Parent epic: #266. Model gate: #271. Final gate: #272.

Decision: residual_lift_candidate.

Final positive claim supported: False.

This report is an experimental residual-model gate only. It is not a final GenoLeWM-FX model-quality result, clinical result, deployment-readiness result, broad VEP superiority claim, broad Carbon comparison, useful-planning claim, ground-truth biology claim, or exact fipip overlap claim.

Reproduce

uv run python -m tools.research.fx_borzoi_residual \
  --cache-manifest docs/research/fx-borzoi-cache-manifest.json \
  --baseline-report docs/research/fx-borzoi-baseline-report.json \
  --output-predictions docs/research/fx-borzoi-residual-predictions.parquet \
  --output-json docs/research/fx-borzoi-residual-report.json \
  --output-md docs/research/fx-borzoi-residual-report.md

Metrics

Model AUROC AUPRC Balanced accuracy
Strongest simple baseline 0.6494 0.2225 0.5080
Residual ensemble 0.6594 0.2300 0.5026

Paired deltas versus strongest simple baseline:

Metric Delta 95% CI
AUROC 0.0099 [-0.0035, 0.0237]
AUPRC 0.0075 [-0.0029, 0.0198]

Prediction Artifact

Artifact Rows SHA-256 Size
docs/research/fx-borzoi-residual-predictions.parquet 3390 sha256:968ba05757c563b1d13dfcdaed0f0df7c9511a570c35f81da62515cc604fb978 144460

Interpretation

The residual model shows a small positive mean lift, but the paired confidence interval crosses zero. This is implementation evidence for #271, not a positive locked-result claim for #272.