Trust-aware retrieval & memory security

Available materials

The public repository contains development materials and a README. Verified public on 29 October 2026.

Documented technologies

Python.

Development status

A security engineering harness documenting retrieval provenance, temporal claims, memory gates and deterministic action authorization. Effectiveness has not been independently evaluated here.

Use and limitations

Review the repository licence before reuse. Repository visibility does not imply an open-source licence, peer review or clinical safety.

Research and engineering questions

Retrieval systems may combine sources with different provenance and age. A security harness can help frame questions about how retrieved text, remembered claims and proposed actions are evaluated before an action is authorised.

Suggested inspection sequence

Read the threat assumptions, inspect the documented evaluation commands and distinguish deterministic policy checks from model-generated judgments. Review how test cases represent stale information, conflicting claims and untrusted retrieved instructions.

Limits of a harness

Passing a selected test set would establish behaviour on those cases only. It would not prove protection against every attack or deployment configuration. Independent evaluation, broader cases and documented failure analysis remain useful next steps.