The library
Rendered from docs/reports/2026-09-26_draft-review-branches.md; the markdown file is the record.
Draft-review reaction branches (pre-computed Sat Sep 26, ~09:30)
Sent for review (12:03, superseding the 09:25 plan): submissions/20260926-0914_M4_sweep_A3_w0.2_E_CANDIDATE_our_direction.tsv (sha256 fb783076…), copied as ~/Downloads/CrossExam_predictions.tsv (team Cross Exam).
- Rule: A3 w0.2 + E. M1b v3 60% + handcrafted v5 20% + M5 20% in rank space, with M3 used only to suppress false alarms. Ratified by Nathan at 12:20 (GO A3) after NSA said the Docker image is not required; the pre-declared rule said replace. The runner's default is
models/fusion_v2/constants.jsonsince 12:35 and reproduces this file (exports: 7e-16; live: Spearman 1.0, max 9.3e-4). The 09:25 plan sent the 08:13e_on_afile (096f3f0c…); that file is the fallback, loadable via--fusion models/fusion_v1/constants.json. - Ask NSA for: minDCF, P_FA, P_miss and EER.
Every branch is already a logged TSV in 1,671-row template order, so reacting to NSA's number is a file copy.
| NSA's returned minDCF | Meaning | Final TSV to submit (copy, rename to TeamName_predictions.tsv) |
|---|---|---|
| 0.00–0.20 | Our score direction; the test set behaves like our holdout. | submissions/20260926-0914_M4_sweep_A3_w0.2_E_CANDIDATE_our_direction.tsv (the file sent at 12:03; unchanged). Until 12:20 this row pointed at the 08:13 e_on_a file. |
| 0.20–0.45 | Our direction; the test set is more like In-the-Wild or out of family. | The same A3 file. "Most FA-robust" means lowest ITW brief-cost minDCF, the false-alarm-weighted, threshold-swept metric the rules were chosen on; A3 w0.2 + E has the best ITW brief (0.228) and averse (0.239) of every ratified rule. |
| 0.45–0.90 | Ambiguous. Do not flip. | submissions/20260926-0928_BRANCH_C_M1b_alone_our_direction.tsv (M1b v3 alone, same 0.001 + 0.999·p map; unchanged). |
| 0.95–1.00 | NSA's code is reading our scores inverted. | submissions/20260926-0914_M4_sweep_A3_w0.2_E_CANDIDATE_FLIPPED_only_if_NSA_scores_inverted.tsv. |
If NSA returns EER instead: 2–5% means our direction, so use the first row. 95–98% means their code's polarity, so use the last row.
Also pre-computed, not tied to a band:
…BRANCH_C_M1b_alone_FLIPPED.tsv, for the case where NSA's number points to inversion and M1b alone is still preferred.submissions/20260926-0914_M4_sweep_A3_w0.2_E_CANDIDATE_*.tsv: the A3 candidate (positive result, not shipped).
All of these files are archived at /Volumes/Crucial P3 NVME Gen 3 2TB/hearsay/submissions/.
Docker / runner (12:35): the runner points at models/fusion_v2/constants.json (A3 w0.2 + E) by default; the image (not a graded deliverable, NSA ~12:00) stays on e_on_a as reproducibility evidence. It already supports --flip (since 9614c18: scripts/run_pipeline.py, pipeline.py), which maps 1 − p through the same determinate map and keeps the gate block at the bottom, and the Docker smoke test covers it. So an inversion verdict changes only which TSV is sent.