peak_umm_iter

20260509_153511_peak_umm_iter · complete · published 2026-07-05 · seed 42

Intent

Locate each Z's true peak unseen_mismatch training iteration more precisely than the six coarse waypoints of paper_long_iter allow: take the coarse argmax per Z, then run a dense 9-point geometric bracket around it.

Hypothesis

The coarse waypoints under-sample the peak: the true peak iteration is interior to the coarse bracket for mid/high Z, decreases monotonically with Z, and the refined peak values exceed the coarse ones where the coarse grid straddled a narrow peak.

Conclusions

Refined peaks per Z (from refined_peaks.csv in the analysis dir):

1. Three regimes in peak timing.
2. Z ≤ 0.7: no interior peak — the metric is still (noisily) rising at the right edge (peak pinned at ~10,000, values 30–46%, late_monotone for most). Uniform pairings never produce a clean agreement peak.
3. Z ≈ 0.8–1.1: interior peaks appear at ~1400–2000 iterations with large seed spread (std 13–31) — the transition region, where seeds split between solutions (connects to the B_recent/B_subject bimodality work).
4. Z ≥ 1.2: sharp early peaks, moving from ~2000 down to ~700–850 iterations as Z grows, values 86–96% with tightening spread.
5. Peak iteration decreases with Z across the interior-peak range, confirming the paper_long_iter coarse trend with ~2× finer resolution (peak_iter_vs_alpha.png).
6. The coarse grid did straddle narrow peaks: e.g. Z=0.8 refines to iteration 1427 (63.0%) where the coarse argmax was 2400; several Z in 1.1–1.6 refine to off-waypoint iterations (2018, 849, 714) with values a few points above the coarse ones (peak_iter_grid.png).

Caveat: the child's own summary CSV covers only the tail alpha slice (bracket runs are stored per-α under dense_bracket/experiments/); the authoritative merged table is merged_metric.csv + refined_peaks.csv in the analysis dir.

Comparison figures

Children

experiment.yaml