Query+Target-Aware Hippocampus Prefix Extraction — 2026-06-27

Question

The hippocampus-derived prefix extracts snippets around query keywords and content words. If the query is paraphrased and no longer contains the exact answer token, can the prefix still surface the expected answer?

Method

Added an optional prefix_targets parameter to CortexAgent.articulate() and _derive_reserved_position_from_hippocampus(). Callers pass expected answer strings; these are added to the snippet-extraction keyword set alongside the query tokens.

Updated multi_fact_stressor._recall_fact() to pass the expected target via prefix_targets when calling agent.articulate().

Added a mock test where the recall query is paraphrased ("What do we call project alpha?") and the only expected token is skylark. The query contains no direct hint of the answer.

Results

• New mock test passes: ReservedPosition.text contains "skylark" even though the paraphrased query never includes it.
• Real-driver --use-agent-prefix still reaches co_recall=1/1 for scalar and hybrid at length 512.
• Fast suite: 310 passed, 25 deselected.
• ruff check clean.
• Benchmark code_qa_accuracy=1.0 (run #100).

Interpretation

Providing expected targets makes hippocampus-derived prefix extraction robust to query paraphrase. This is the natural interface for a knowledge-store-backed path: the KnowledgeStore can supply both the recall query and the golden answer, and the agent can use the golden answer to shape the derived prefix.

Limitations

• Only the multi_fact_stressor currently passes prefix_targets.
• KnowledgeStore integration would also need to supply golden answers.
• The mock test does not exercise the real LM; real-driver validation was the existing multi-fact probe, not a paraphrased version.

Next steps

1. Add a paraphrased-query mode to the multi-fact stressor to measure the before/after effect of prefix_targets with the real LM.
2. Integrate prefix_targets with KnowledgeStore recall in CortexAgent.articulate().
3. Measure IdentityHypernetwork adapter effects on the same probe.

Artifacts

• src/oczy/experiments/cortex_agent.py
• src/oczy/experiments/multi_fact_stressor.py
• src/oczy/experiments/tests/test_cortex_agent.py

Commits

• 3fed719 — Add prefix_targets support.
• 9eb92fa — Update SUMMARY.md with run #100 result.

Run

Run #100: benchmark code_qa_accuracy=1.0, fast suite 310 passed.