Sentinel relevance from experience: actual reconstructed experiment results
I generated this report by executing the accompanying script.
I did not recover the original experiment or verify its earlier F1 figures.
Run UTC: 2026-10-28T14:30:19.663947+00:00
Python: 3.15.1
Script SHA256: bbcaad257dadf4e22ece581d0207c68420f0c1187ebdb3859b1f876a4ff86d6e
Dataset SHA256: 97184ea671a61fb8ff9ec79afbb641a83f1b477c98d2ab4faeb568dbf2bece88
Seed: 20261028
Observations: 15000; fitting: first 10000; held out: last 5000
I learned from 9999 delayed progress outcomes revealed before the split.
Held-out relevance positives: 834
Held-out observed progress positives: 1164
Fixed parameters: switch=0.025, correct action=0.75, trace decay=0.85,
history=13, threshold=0.5, progress probabilities=0.9/0.1, Laplace=1/1
I freeze consequence evidence at the split and continue updating action state.

I score positive-class relevance F1 and two distinct Brier targets:
method             precision recall   F1       relBrier progBrier   TP   FP   FN   TN
trace              0.951153 0.840528 0.892425 0.058591 0.116867  701   36  133 4130
history13          0.921271 0.799760 0.856226 0.063531 0.120434  667   57  167 4109
current_action     0.762797 0.732614 0.747401 0.075310 0.133993  611  190  223 3976
resource_only      0.000000 0.000000 0.000000 0.116565 0.159686    0    0  834 4166
never_positive     0.000000 0.000000 0.000000 0.166800 0.232800    0    0  834 4166
shuffled_trace     0.000000 0.000000 0.000000 0.116581 0.159563    0    0  834 4166

I learned these conditional progress probabilities (inferred activity x resource):
0 0.807554 0.105675 0.155945 0.102041 0.164207 0.073214
1 0.174569 0.088176 0.812081 0.088300 0.169355 0.089936
2 0.130769 0.114833 0.164425 0.120536 0.817404 0.128550

Trace minus shuffled F1: 0.892425
Fixed alignment rule (gap >= 0.10): PASS
Trace minus 13-action history F1: 0.036199
Fixed history improvement rule (gap >= 0.01): PASS
Invariant checks: PASS

I report one designed synthetic world with one fixed seed, without uncertainty intervals.
I do not infer actual preferences, causality, verified personal opportunities or a general thinking law.
I have not reproduced the historical 0.925, 0.924 and 0.336 F1 claims.