REAL MODEL OUTPUTS · SYNTHETIC RECORDS & MONEY
Memory retrieval and use
○ Agent □ Deterministic step ◇ Input Solid: used Dotted: data reference Dashed gray: disabled. Select a node for saved evidence.
ALL CASES · FIXED EQUAL FAMILY WEIGHTS
What changed the risk?
Proposed unauthorized loss is scored before the independent gate. Memory, model and critic differences are measured on matched fresh cases. Choose an earlier batch to inspect the calculation ablation.
What each change contributes
Positive values mean fewer harmful proposals. Lines show 95% model-based ranges; crossing zero leaves the direction uncertain. Each comparison holds the other settings fixed.
How uncertainty changes with more cases
Ranges use balanced prefixes of the same batch. They are model-based summaries, not independent experiments or simultaneous bounds.
Scenario evidence and assumptions
Families have equal prespecified weights and separate Jeffreys priors. Estimates describe this challenge mix and depend more on the prior when samples are small.