Compaction Laundering: provenance destruction

Evidence record

Preprint; experimental compaction and execution results.

Read the original source ↗

Claim-level citations (3)

• supportsFlattened conversation roles can turn tool text into apparent user approval: With 15 trials per harness, five role-flattening harnesses laundered 13-15 forged approvals; four role-typed harnesses laundered none. All clean controls were 0/15.

  Experiment E10, page 13 and Figure 4; three models times five trials

• supportsFlattened conversation roles can turn tool text into apparent user approval: The provenance check caught 77% at 2% legitimate-claim demotion in its chosen setting; it was not tested against an adaptive attacker.

  E9 threshold 0.15; Limitations, defense-adaptivity bullet

• supportsFlattened conversation roles can turn tool text into apparent user approval: E11 isolated execution after forced laundering, with five trials per harness; this is not the 87-100% summary-laundering success metric.

  Experiment E11, pages 13-15