Audit 2026-05 — Evidence Dumps

Real artifacts captured locally on 2026-05-27 to back the rebuttal at audit-rebuttal-2026.md. Re-generate with:

/usr/bin/python3 scripts/expand_audit_rebuttal.py \
  > docs-site/docs/audit/evidence/105-alegacoes-ponderadas.md
/usr/bin/python3 scripts/dump_coverage_model_info.py
gh run list --workflow satellite-change.yml --limit 30 --json \
  databaseId,displayTitle,conclusion,createdAt,event,headBranch \
  > /tmp/runs.json
/usr/bin/python3 -m pytest tests/test_atoll_planet_export.py \
  test_offline_mode.py test_snap_anatel.py test_coverage_model_info.py -q \
  --deselect test_offline_mode.py::test_bedrock_invoke_model_canned_offline \
  --deselect test_offline_mode.py::test_bedrock_list_available_models_static_offline

Index

File	Backs claim #	What it shows
coverage-model-info.json	#2, #6	Real GET /coverage/model/info payload with the §B2 audit_gate block. Synthetic n=200 model: effective_cv_rmse_db=11.97, passed=false — gate correctly refuses unfit model. Production model trained on drive-test data is expected to pass.
satellite-change-last30.md	#5	Last 30 runs of satellite-change.yml: 22 success + 5 failure (full breakdown in the dump itself). Mix of schedule (weekly cron) and workflow_dispatch. Cadence — not success rate — is the point: refutes the "monthly tower update" claim.
pytest-anchors.txt	#3, #5, #6, #9	Anchor tests green: tests/test_atoll_planet_export.py, test_offline_mode.py (2 boto3 deselected for missing env dep), test_snap_anatel.py, test_coverage_model_info.py. 72 passed, 2 deselected.

How to read the audit_gate dump

"audit_gate": {
  "max_cv_rmse_db": 8.0,
  "effective_cv_rmse_db": 11.9684,
  "effective_source": "local_model",
  "passed": false
}

• effective_cv_rmse_db is the minimum of CV-RMSE across the heads actually loaded (residual LightGBM + local Ridge). Whichever the stack will use at predict time.
• effective_source tells which head won the comparison.
• passed is the single bit the auditor (or a CI gate) reads to verify the §B2 commitment of cv_rmse_db < 8 dB.

The synthetic 200-sample model intentionally fails the gate; the endpoint is honest about it. After B1 (drive-test pilot) lands a production residual_lightgbm artifact, this dump will flip to passed: true with effective_source: "residual_lgbm".