golden: freeze the IMPROVE neoepitope example outputs

tests/golden/improve/GOLDEN.md

@@ -0,0 +1,45 @@
+# IMPROVE golden baseline
+
+Frozen regression baseline for the IMPROVE neoepitope-immunogenicity example. Generated by:
+
+```bash
+uv run python examples/improve/build_fixture.py --src 'improve_results/results copy'
+uv run python examples/improve/run_evaluation.py
+```
+
+(default config: `B=2000`, `seed=1337`, from `examples/improve/cohort.yaml`; 7 models, 17,520
+peptides, G=70 patient clusters, 467 immunogenic = 2.666% prevalence)
+
+| | |
+|---|---|
+| generated | 2026-10-07 |
+| predval commit | `4f9fbd0` (predval 0.1.0) |
+| source repo | `https://github.com/SRHgroup/IMPROVE_paper` at `c67094276b52358b5202ad9e5597acd87888ea1d` |
+| `results.zip` sha256 | `26897006104ef0e57aeedc003804c9a7e43d3735d9df157d651516f2739e00f9` |
+| wall time | ~12.5 min on a 4-core container |
+
+`findings.json` keeps its `provenance.git_commit` as generated; `tests/test_golden_improve.py`
+normalizes that one field out of the comparison, as the GUSTO and PCam goldens do.
+
+Reproducibility was verified at freeze time: a second independent run at the same commit was
+compared against these files by `tests/test_golden_improve.py`.
+
+## Headline numbers this baseline pins
+
+- **Prevalence anchoring.** The six IMPROVE random-forest configurations predict a mean
+  probability of 0.39-0.40 against an observed rate of 0.0267, i.e. 14.6-15.1x too high. The
+  NNAlign comparator is anchored (1.1x).
+- **Discrimination.** AUROC 0.63-0.65 for the IMPROVE models, 0.61 for NNAlign
+  (patient-clustered 95% CIs roughly +/-0.04).
+- **Calibration.** Intercepts about -3.25 (CI excludes 0) with slopes about 1.1 (CI includes 1):
+  the IMPROVE scores are a ranking with the right spread and the wrong level. NNAlign:
+  intercept -0.09 [-0.33, 0.15], slope 0.97.
+- **Recalibration ladder.** Cross-fitted rung1 (intercept only, fitted on held-out patients)
+  cuts Brier from ~0.17 to 0.0258 for every IMPROVE model, a demonstrated repair. That repaired
+  Brier equals the prevalence-only constant predictor (0.0259): once the level is fixed, the
+  probabilities carry almost no Brier-scale skill beyond the base rate. NNAlign gains nothing
+  from recalibration.
+- **Unit of analysis.** Patient-clustered AUROC intervals are 1.3-1.7x wider than naive
+  per-peptide ones.
+- **Fragility.** Dropping patient RH-31 moves AUROC the most for six of seven models
+  (largest 0.013).