kinoTou

experiment log

Paraphrased-Query Multi-Fact Stressor Mode — 2026-06-27

File
2026-06-27_paraphrase_mode.md
Size
2.5 KB
SHA-256
3cde9663cd047832…
Download raw file ↓Back to research →
Primary source. This is the verbatim Oczy document. The analytical field notes on the research page interpret and summarize these sources.

Paraphrased-Query Multi-Fact Stressor Mode — 2026-06-27

Question

Does the hippocampus-derived prefix path still achieve exact-token recall when recall queries are paraphrased and no longer contain the original keywords that made snippet extraction easy?

Method

Added --paraphrase to multi_fact_stressor.py. When enabled, recall queries become:

  • Alpha: "What name is used for project alpha?" (original: "What is the codeword for project alpha?")
  • Beta: "What do we call project beta?" (original: "What is the codeword for project beta?")

The prompt shown to the LM still uses the original question (QUERY_A/QUERY_B) because the expected target string is part of that literal text. The recall_query passed to agent.articulate() is the paraphrase, while prefix_targets=TARGET_A/TARGET_B is unchanged.

Ran real-driver scalar and hybrid with --use-agent-prefix --paraphrase --length 512, and compared to the non-paraphrased baseline with the same flags.

Results

mode paraphrase co_recall prefix_source consolidation_strength
scalar no 1/1 hippocampus 10.0
scalar yes 1/1 hippocampus 10.0
hybrid yes 1/1 hippocampus 35.99

The paraphrased queries do not degrade exact-token recall.

Interpretation

  • prefix_targets successfully guides snippet extraction even when the query text no longer contains the high-signal keyword "codeword".
  • Hippocampal replay is driven by the paraphrased query; the synthetic embedding similarity still retrieves the relevant traces because "project alpha" / "project beta" remain in the paraphrase.
  • This is a meaningful robustness test: the prefix is derived from expected targets, not just query keywords.

Limitations

  • Only two paraphrase variants tested.
  • The long-turn facts are still planted with the exact target tokens, so the memory surface is rich enough.
  • No measurement was made without prefix_targets under paraphrase; that would isolate the value of the target-aware feature.

Next steps

  1. Add a paraphrase-without-prefix_targets condition to isolate its value.
  2. Integrate prefix_targets with KnowledgeStore recall.
  3. Measure IdentityHypernetwork adapter effects.

Artifacts

  • src/oczy/experiments/multi_fact_stressor.py
  • src/oczy/experiments/tests/test_multi_fact_stressor.py

Commits

  • dc749d9 — Add --paraphrase mode.
  • 76faddf — Update SUMMARY.md with run #101 result.

Run

Run #101: benchmark code_qa_accuracy=1.0, fast suite 311 passed.