Skip to main content

​
Purpose

DEP-02 is a deterministic local reference for testing whether retrieval-induced adaptive state can make hidden retrieval activity distinguishable to an authorized observer. It is an intentionally controlled mechanism, not an external attack or vulnerability finding.

​
Four-arm design

All arms reset from the same canonical initial state: Each arm then runs the same one-query observer probe and the same deterministic region decision. The vulnerable policy uses /derived_state/shared_access_counter as its only adaptive mediator. The pure-read ablation does not change target state.

​
Observer boundary

The observer view contains only:
  • its own authorized query;
  • ordered response memory identifiers and contents;
  • the response result count.
The canary, hidden assignment, evaluator oracle, trace, full state, and wall-clock timing are not part of the view or classifier. Direct canary disclosure is classified separately and cannot count as RISI-C.

​
Frozen result contract

The comparison reports controlled_mechanism_recovered only when:
  • the vulnerable pair is classified risi_c with advantage 0.5;
  • the pure-read pair is classified no_failure with advantage 0.0;
  • both pairs respect the exact one-probe budget and exclude the canary;
  • the shared counter is the sole vulnerable-pair state difference and the pure-read pair has none;
  • all four synthetic region decisions are safe.
Model-free replay verifies the evidence inventory and hash-chained traces, reconstructs every final state, cross-checks retained observer and decision artifacts, and recomputes both pair assessments.

​
Run it

Do not generalize this synthetic reference result to another model, memory system, deployment, or population. External testing requires separate authorization, isolation, adapter implementation, and evidence review.