Pre-register the YOLO-baseline evaluation & checkpoint-selection protocol (#71)

[jonfroehlich]

Closes #71.

Why now, specifically

The five YOLO recovery runs (jobs 37745358–363) are still training on ckpt-g2, and no YOLO checkpoint has been scored on any benchmark bundle — the only numbers anyone has seen are the internal val-split curves already committed in the training record and mirrored in #51''s status comments. Pre-registration only counts if it lands before test numbers exist, so this was a race against those jobs finishing.

What this adds

A "Pre-registered evaluation & checkpoint-selection protocol" section in scripts/model_comparison/yolo_baseline/README.md (the training record), fixing in writing:

1. Checkpoint selection: best-val, never test. best.pt exactly as Ultralytics saves it (best val fitness = 0.1·mAP50 + 0.9·mAP50-95, patience=20). This refines the issue''s draft wording ("best val mAP50") to the artifact the runs actually keep — and the two rules pick the same epoch (1) on all five committed curves, verified from runs/*/results.csv. RampNet parity stated: its published checkpoint was likewise selected on val (best validation loss, stage_two/train.py), never a benchmark number.
2. Config/headline selection on val only, with the honest caveat that tiles vs pano val sets differ — so the choice controls emphasis only; every trained config''s benchmark row is reported regardless.
3. Benchmark eval through the identical path as the whole roster: compare.py yolo provider → box centers → rampnet/detection_eval.py matcher at radius 0.022, 0.05 cache floor; tiles configs through the same perspective rig as the VLMs (imgsz 1024), pano configs whole-pano (imgsz 1280).
4. Operating point fixed in advance, not tuned on test: headline F1 at conf 0.25 — Ultralytics'' own default predict confidence, i.e. the "configure the baseline the way its authors recommend" choice. The sweep is still printed with its best-F1 row flagged as tune-on-test, as the harness already does.
5. AP with the truncation caveat vs RampNet''s committed 0.5-peak extraction (the low-floor RampNet curve is the Lower the deployment operating point (peak threshold 0.55 → ~0.25–0.35): a free, recall-first lever #54/Experiment: flip-TTA at deployment operating points — the evaluated config used it, deployed inference doesn't #78 line of work).
6. Unstable-run rule: an instability-affected best.pt is reportable only as a lower bound, superseded by the Rerun the collapsed YOLOv11 pano baselines with a stabilized (larger physical batch) config #70 stabilized rerun evaluated under this same protocol, unchanged.
7. Seeds aspirational (≥3 seeds of headline configs if ckpt capacity allows; all current runs are seed 0).

Cross-references added per the issue''s deliverable list: compare.py''s module docstring points at the protocol, and docs/model_comparison.md''s baseline-in-progress note now links to the committed section instead of just citing #71. The paper''s appendix "Baseline protocol" paragraph is the section, condensed.

pytest -q: 188 passed.

🤖 Generated with Claude Code (claude-fable-5)