kinoTou

experiment log

Multi-Fact Turn Stressor — 2026-06-26

File
2026-06-26_multi_fact_stressor.md
Size
2.8 KB
SHA-256
ca640ec862db1d73…
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.

Multi-Fact Turn Stressor — 2026-06-26

Question

Can a single long turn store both a novel fact and a correction-styled fact, and can architecture H (hybrid consolidation-strength modulation) improve "co-retention" over architecture S (scalar stats gate)?

Method

Created src/oczy/experiments/multi_fact_stressor.py:

  • Builds a 512-token turn with filler and two facts:
    • Fact A (novel): "The codeword for project alpha is skylark."
    • Fact B (correction-style): "Correction: the codeword for project beta is not raven, it is rook."
  • Processes the turn through a CortexAgent with the ingestion pipeline active:
    • chunker: fixed-window, 64 tokens
    • salience: lexical-novelty
    • embedder: same-lm (mock driver)
    • aggregator: stats
  • Forces consolidation via agent.consolidate().
  • Queries each fact and checks target substrings:
    • Query A: expect "skylark"
    • Query B: expect "rook"
  • Supports --mode scalar and --mode hybrid to toggle DigestiveGateConfig.use_hybrid_consolidation.

Results (mock driver)

mode recall_a recall_b co_recall traces embedding_calls consolidation_strength cold_drift
scalar 0 0 0 3 4 1.00 0.10
hybrid 0 0 0 3 4 6.39 0.64

Both modes retain 0/0 facts with the mock driver, but the mechanical effect of hybrid mode is visible: consolidation strength is scaled by (1 + drift_max) capped at 10.0, and cold drift is correspondingly larger.

Interpretation

  • The mock driver's embeddings are deterministic hash-based vectors; they distinguish chunks but carry no semantic content, so retrieval cannot work even when the correct chunk is stored. This is the same limitation as the mock needle sweep.
  • The stressor does exercise the architecture S vs H difference on the consolidation-strength axis, which is meaningful for real-driver evaluation.
  • Lexical-novelty keeps only 3 traces for the entire 512-token turn — the embedding cost is already very low.

Implication for architecture

The multi-fact probe is now ready to run against the real LFM2.5 driver. If the real driver shows co_recall > 0, architecture H and S can be compared on memory cost vs retention. If the mock behavior is representative, H should produce stronger persistent traces for high-drift (correction) chunks.

Open questions

  1. Does the real-driver version of this probe show non-zero co_recall?
  2. Does hybrid mode improve co_recall at the same memory delta?
  3. Would boosting the novel fact's chunk persistence (not just strength) help?

Artifacts

  • src/oczy/experiments/multi_fact_stressor.py
  • src/oczy/experiments/tests/test_multi_fact_stressor.py
  • src/oczy/experiments/logs/SUMMARY.md

Commits

  • 0f566c4 — Add multi_fact_stressor.py probe and tests.
  • b3e1488 — Update experiments/logs/SUMMARY.md with run #85 result.