# Live CortexAgent Hippocampus Prefix Validation — 2026-06-27

## Question

Does the live `CortexAgent.use_hippocampus_prefix` path actually generate a
useful `ReservedPosition` during `articulate()` and produce exact-token recall
in a real multi-fact turn?

## Method

Added `--use-agent-prefix` to `src/oczy/experiments/multi_fact_stressor.py`.
When this flag is set:
1. The stressor enables `agent.config.use_hippocampus_prefix = True`.
2. `_recall_fact()` now passes `recall_query=query` to `agent.articulate()`, so
   the agent has a query to drive hippocampal replay.
3. The stressor wraps `agent.set_reserved_position()` to capture the source of
   any `ReservedPosition` the live agent sets (because `articulate()` clears it
   in `finally`).

Ran real-driver scalar and hybrid with `--auto-consolidate --hybrid-cap 0
--use-agent-prefix --length 512`.

## Results

| mode | prefix_source | recall_a | recall_b | co_recall | consolidation_strength |
|---|---|---|---|---|---|
| scalar | hippocampus | 1 | 1 | 1/1 | 10.0 |
| hybrid | hippocampus | 1 | 1 | 1/1 | 35.99 |

Both configurations reach exact-token co_recall via a prefix generated by the
live `CortexAgent`, not by the stressor wrapper.

## Interpretation

- The closed loop works end-to-end:
  - perceive long turn → hippocampus stores high-surprise traces
  - consolidate → traces remain available
  - articulate with recall_query → agent replays relevant traces
  - derive ReservedPosition from replay → apply as reserved-position prefix
  - exact-token recall achieved.
- The stressor wrapper's own `_derive_prefix_from_hippocampus` is no longer
  required when `--use-agent-prefix` is active; the agent handles it.
- Scalar and hybrid are indistinguishable at this probe resolution.

## Limitations

- Hippocampus keyword extraction still relies on query tokens and content words;
  it has not been tested on out-of-vocabulary or paraphrased queries.
- `--use-agent-prefix` is a stressor flag; production use requires enabling
  `CortexAgentConfig.use_hippocampus_prefix` directly.
- No comparison was made to the hand-coded prefix path in the same run, but
  earlier runs show it also reaches `co_recall=1/1`.

## Next steps

1. Remove or deprecate the stressor-only `--auto-prefix` now that `--use-agent-prefix`
   validates the agent path.
2. Generalize keyword extraction with explicit expected targets from a
   knowledge store.
3. Measure IdentityHypernetwork adapter effects on the same probe.

## Artifacts

- `src/oczy/experiments/multi_fact_stressor.py`
- `src/oczy/experiments/tests/test_multi_fact_stressor.py`

## Commits

- `e580526` — Add `--use-agent-prefix`.
- `ab1134a` — Update SUMMARY.md with run #98 result.

## Run

Run #98: benchmark `code_qa_accuracy=1.0`, fast suite `310 passed`.