kinoTou

experiment log

Multi-Fact Turn Stressor — Real Driver — 2026-06-26 (corrected)

File
2026-06-26_multi_fact_stressor_real_driver.md
Size
2.4 KB
SHA-256
674487665ca9ac3e…
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 — Real Driver — 2026-06-26 (corrected)

Question

Does the multi-fact stressor run end-to-end against the real LFM2.5 GGUF and can it discriminate architecture S vs H on co-retention?

Method

Extended src/oczy/experiments/multi_fact_stressor.py with an optional --use-real-driver flag that loads LlamaCVecDriver and passes it to CortexAgent. Ran --length 256 in both scalar and hybrid modes.

Results

mode recall_a recall_b co_recall traces embedding_calls consolidation_strength cold_drift
scalar 0 0 0 3 0 1.00 0.086
hybrid 0 0 0 3 0 3.56 0.305

Both modes produced valid METRIC and ASI lines against the real driver. Hybrid shows the expected mechanical scaling (strength 1.0 → 3.56, cold_drift 0.086 → 0.305). embedding_calls reports 0 because only _MockDriver counts embedder invocations, not LlamaCVecDriver.

Interpretation

  • The integration works: real driver loads, pipeline stores 3 traces, and the probe emits metrics.
  • Co-recall is 0/0 not primarily because of instruction formatting but because cvec-only consolidation cannot reliably force exact target tokens. This matches the prior consolidation-uptake finding: residual steering shifts the semantic domain but does not bind exact vocabulary.
  • Hybrid modulation is mechanically real: it produces stronger persistent traces from high-drift chunks. But stronger cvecs still do not guarantee exact-token recall.

Implication for architecture

To make the multi-fact stressor behaviorally discriminating, the retrieval path needs a reserved-position/prefix surface (e.g. LlamaCVecDriver articulation prefix) or a knowledge-store fact-injection path, not just stronger cvecs. This is the next frontier.

Open questions

  1. Does adding ReservedPosition prefix support to the stressor yield non-zero co_recall?
  2. Once exact-token recall is possible, does hybrid mode improve co_recall or reduce memory cost?
  3. Should the stressor also test mock-foreign embedder recall to compare cost?

Artifacts

  • src/oczy/experiments/multi_fact_stressor.py
  • src/oczy/experiments/tests/test_multi_fact_stressor.py
  • experiments_logs/2026-06-26_multi_fact_stressor_real_driver.md

Commits

  • b171ec0 — Make multi_fact_stressor runnable against real LlamaCVecDriver.
  • 2a5b786 — Update experiments/logs/SUMMARY.md with run #86 result.