kinoTou

experiment log

Auto-Consolidate S vs H Probe — 2026-06-27

File
2026-06-27_auto_consolidate_sh.md
Size
2.6 KB
SHA-256
d3ce6859d26b8fef…
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.

Auto-Consolidate S vs H Probe — 2026-06-27

Question

Can an auto-consolidate path in the multi-fact stressor behaviorally discriminate architecture S (scalar) from architecture H (hybrid consolidation strength modulation)?

Method

Added --auto-consolidate to src/oczy/experiments/multi_fact_stressor.py. When active:

  • The agent's auto_consolidate config is set to True.
  • consolidation_pressure_threshold is lowered to 0.05 so a single 512-token high-drift turn can trigger consolidation.
  • After metabolize(), if agent.should_consolidate() is True, the probe runs agent.consolidate(strength=...). For H mode, strength scales by (1.0 + digest.drift_max) capped at 10.0.
  • The result includes auto_consolidated boolean.

Real-driver runs at length 512, no prefix, mode scalar and hybrid.

Results

mode auto_consolidated cold_drift consolidation_strength recall_a recall_b co_recall
scalar 1 0.867 10.0 0 0 0
hybrid 1 0.867 10.0 0 0 0

Interpretation

  • Both modes auto-consolidate because the lowered threshold is easily crossed.
  • Both hit the 10.0 strength cap, so hybrid's additional scaling has no effect.
  • Exact-token recall remains 0/0 because cvec-only consolidation cannot force target tokens.

The S vs H difference is masked by the cap and by the exact-token limitation.

Implication for architecture

The multi-fact stressor in its current form cannot discriminate S vs H. To do so, one of these changes is needed:

  1. Remove or raise the 10.0 consolidation-strength cap so hybrid can scale beyond scalar.
  2. Use multi-turn pressure accumulation so scalar and hybrid diverge in when consolidation fires.
  3. Measure domain-level recall (e.g. whether the answer mentions the right project) rather than exact target tokens, since cvec steering shifts domain.
  4. Switch the exact-recall path to hippocampus-derived prefixes and measure whether hybrid produces better/compressed prefixes.

Open questions

  1. Does removing the strength cap destabilize generation or improve recall?
  2. Does hybrid mode save memory (fewer traces for equivalent recall) when the cap is removed?
  3. Is domain-level recall in the multi-fact probe a more sensitive metric?

Artifacts

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

Commits

  • a2814e7 — Add --auto-consolidate mode.
  • fcbb9cd — Update SUMMARY.md with run #92 result.

Run

Run #92: benchmark code_qa_accuracy=1.0, fast suite 301 passed.