Configurable Hybrid Consolidation Cap — 2026-06-27

Question

Does the 10.0 consolidation-strength cap mask architecture H's effect in the multi-fact stressor? If the cap is removed, does H produce stronger persistent traces and better recall?

Method

Added --hybrid-cap to src/oczy/experiments/multi_fact_stressor.py. --hybrid-cap 0 disables the cap. Ran real-driver S vs H auto-consolidate probes at length 512, with and without reserved-position prefix.

Results

Without prefix (cvec-only)

With prefix

Interpretation

• Removing the cap exposes a real architecture difference: hybrid produces ~3.6x stronger consolidation strength than scalar on the same turn.
• However, the extra strength does not increase cold_drift or exact-token recall. The cvec-only path remains blocked at 0/0; the prefix path remains at 1/1 regardless of mode.
• This suggests the extra consolidation strength is either:
  1. Saturated inside agent.consolidate() before affecting trace strength,
  2. Detrimental or neutral rather than helpful for exact recall, or
  3. Manifesting on a metric we are not yet measuring (e.g. memory-per-byte).

Implication for architecture

Hybrid consolidation modulation is mechanically effective, but the cap of 10.0 was hiding that effect. Whether the effect is useful requires a non-exact recall metric or a memory-cost metric. The current probe design does not reward higher strength; it only checks exact target tokens, which cvecs cannot force.

Open questions

1. Does higher consolidation strength reduce the number of traces needed for equivalent recall, improving memory-per-byte?
2. Does it improve domain-level or paraphrase recall?
3. Is there an optimal cap, or is unbounded H destabilizing at high drift?

Artifacts

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

Commits

• 48e85f8 — Add configurable --hybrid-cap.
• ae14feb — Update SUMMARY.md with run #93 result.

Run

Run #93: benchmark code_qa_accuracy=1.0, fast suite 302 passed.