← All the evidencesentinel-personal-significance-results.txt
SHA-256 c4508faf69097a7753e223d7f9449e2c38da6745bc40cb71ff1323d18b8bbc27 · download the original
Sentinel personal significance candidate: synthetic, not real personal experience
Seed: 20261028 Observations: 24000 burn-in (no parameter fitting): 8000 validation: 6000 untouched test: 10000
Hidden preference changes: 376 constraint rate: 0.35525 action missing rate: 0.160625
The hidden preference is available ONLY to the synthetic evaluator.
Evidence model assumes visible constraint status, free-action reliability .77,
statement reliability .92, correction reliability .985 and hazard .018.
These are human-designed simulation assumptions, NOT learned values.
VALIDATION (bits of log loss lower is better)
model logloss accuracy coverage90 error|conf90 switch-delay
last_action 1.1269 0.5508 0.0000 nan 8.14
all_actions_window 0.9012 0.6857 0.0000 nan 42.40
last_statement 0.7739 0.8145 1.0000 0.1855 16.77
context_window 0.5472 0.8667 0.1020 0.0229 25.98
context_bayes 0.3078 0.9248 0.7067 0.0255 8.71
context_bayes_no_corrections 0.3375 0.9162 0.6602 0.0275 8.87
Tested hidden preference switches: 84
TEST (bits of log loss lower is better)
model logloss accuracy coverage90 error|conf90 switch-delay
last_action 1.1318 0.5478 0.0000 nan 6.26
all_actions_window 0.9379 0.6179 0.0000 nan 43.97
last_statement 0.6668 0.8449 1.0000 0.1551 15.35
context_window 0.5473 0.8669 0.0864 0.0093 24.75
context_bayes 0.3128 0.9230 0.7005 0.0250 8.77
context_bayes_no_corrections 0.3436 0.9130 0.6605 0.0263 9.78
Tested hidden preference switches: 153
TEST ROWS WITHOUT NEW EXPLICIT STATEMENT/CORRECTION: 9420
context_window logloss=0.5537 accuracy=0.8645 coverage90=0.0841
context_bayes logloss=0.3255 accuracy=0.9194 coverage90=0.6868
context_bayes_no_corrections logloss=0.3496 accuracy=0.9111 coverage90=0.6524
All scoring is on simulator labels, not independently assessed human priorities.
This does not establish real-world preference discovery, causal importance, or consent.
SENSITIVITY CHECKS on untouched test
shuffled constraint status logloss=0.8613 accuracy=0.7315 coverage90=0.4119
very sticky old preference logloss=0.4848 accuracy=0.8985 coverage90=0.8998
overreacts to changes logloss=0.4777 accuracy=0.8768 coverage90=0.2942
A correction rejecting an action must control permissions independently of these predictions.