Seven experiments, one narrowing wall

Experiment 01 introduced events designed to separate exact recall, domain recall, interference, and behavior per byte. It found five de-saturation events but no behavioral transfer: the mock behavior delta and discrimination score were both 0.0; exact recall was 0 while domain recall was 1.0.

The result is a tested null, not a failed benchmark. It showed the instrument could expose posture without confusing it for exact competence.

Experiment 02 asked whether a reserved KV route could beat the cvec ceiling on exact facts. The registered campaign result was a refutation: KV rank-1 count 0, while logit biasing reached all three targets. Separate HF work found KV splice could be rank-for-rank equivalent to a text prefix, but parity with a prefix was not robust fact recall.

The layer-L story is deliberately reported with both outcomes. The pre-registered S1.4 comparison refuted the claim across Qwen and LFM2.5: gaps of −0.083 and +0.058, both below +0.10. A later single LFM2.5 real-driver closure reached +0.109254 and accepted that one reproducibility run. It does not erase the two-architecture refutation.

Context-scoped attractors reached a scope-selectivity index of 1.0 in a single campaign run. That is evidence that state can be partitioned by context under the tested mechanism, with the important caveat that one run provides no seed distribution.

The metabolism-loop experiment executed four consolidations and measured a cold-state slope of 0.1755, yet both metabolism drift delta and drift uptake were 0.0. Internal state changed; captured behavior did not. The later bounded-growth campaign was a five-seed footprint positive: a 0.002079 M1 ratio, bit-identical footprints, and a spread within 20 bytes. It does not rehabilitate the earlier A0b shortcut, which regenerated a random matrix from a seed and therefore discarded the learned updates it was supposed to compress.

Scope

SSI 1.0Positive, but single-run.

Loop

Behavior delta 0.0Four consolidations did not produce captured behavior change.

Growth

Ratio 0.002079Five-seed storage result; not proof that learned content survived compression.

The conversation world-model work replaced a lexical correction detector with a predictive path. The July campaign’s registered condition measured marker-free uptake +1.0, while critic AUC improvement remained 0.0.

This must not be merged with the older June headline. That earlier +1.0 gap used a lexical baseline constructed to score zero and collapsed to 0.0 against a competitive token-overlap baseline. The correct current verdict is accepted-partial for the later uptake surface, null for the critic, and superseded for the earlier headline.

Across the seven experiments, mechanisms that visibly carried content were retrieval-like: prefixes, logit bias, KV content equivalent to a prefix, and the scope-slot reranker. Handwritten neural updates could alter internal state, but they did not reliably make a frozen model execute a newly learned behavior.

That finding motivated the move from hand-authored updates to meta-trained update and articulation rules. The wall was no longer “add a better memory organ.” It was “learn the protocol between experience, state, query, and frozen specialist.”