ARCHIVE STUDY · HYBRID LEARNING EVIDENCE
Five weights learned
Archive applicability choices.
A reward-trained selector improved under substantial scripted assistance. The frozen selector then transferred to full-start campaigns. This is local policy learning. Beam improvement remains unproved.
Saved events only. This page makes no model calls and runs no campaign engine.
Loading hash-verified study projection...
What changed after training
FROZEN BEFORE AND AFTERReward training across three initializations
Inspect checkpoint weights and training blocks
Retained memory saves charge
96 MATCHED CONTRASTSBoth trained arms escape in all 96 trials. Erasure does not prevent escape in this study. The scripted controller recovers facts through audits.
The 96 pairs share an identical prefix before intervention. They represent 16 public root groups, each with two twins and three initializations. They do not establish full-world memory necessity or a global memoryless bound.
Inspect all 96 paired charge contrasts
Replay an actual campaign witness
SIX PRESELECTED PUBLIC TRACESThe same public twin pair has an initial retained, trained retained, and trained erased witness. All six use initialization 28001. These witnesses illustrate the cohort and do not replace its 288 trial receipts.
Recorded event
Actual public state
Inspect the complete public record for this event
The witness records canonical compare_seal argument order. The submitted input preserves the recorded request order. Both compare the same unordered record pair.
Inspect every recorded action and notebook write
What the original ambition still requires
Receipt scope and source commitments
The independent learning audit reports a pass with scope limits. It recomputes recorded updates and probabilities. This viewer's exporter verifies sanitized manifest hashes, cohort structure, paired prefixes, and witness public-state digests. It does not rerun private replay verification.