DIVERGE-CAB1: Content-Addressed Register Bus

Status: closed conjunctive development failure; confirmation unopened

Parents: confirmed NCP1 command owner, confirmed EAL2 temporal owner, closed JRB1 attribution

Objective: replace canonical register coordinates with one episode-local content-addressed basis

Why this successor exists

JRB1 learned evidence and natural initial-state binding exactly and preserved exact recurrent execution, but its pooled query reader reached only 96.58%. JRB1's purported shuffled-table negative also exposed a representational symmetry: every owner learned the same global coordinate swap, so canonical state tuples changed while semantic answers remained correct. A coherent basis change is not an information-destroying intervention.

CAB1 removes the artificial canonical-coordinate requirement. Every learned register owner predicts a position in a randomly ordered, episode-local table. Law packets, initial state, recurrent execution, and late query reading stay in that table-relative basis. Canonical coordinates exist only inside the independent assessor. The query path replaces whole-sentence mean pooling with token-level pointer evidence aggregated by logsumexp.

Frozen mechanism

The trainable owner is a shared byte encoder with dynamic register candidates:

1. numeric mention spans point to table positions for evidence and initial state;
2. each query token scores both table entries;
3. a learned scalar gate selects informative query tokens;
4. token evidence is aggregated into one query-to-position decision;
5. the already-qualified EAL2 temporal owner and NCP1 command owner remain bit-identical;
6. all downstream state is represented and executed in the selected table basis.

The candidate sees raw natural evidence, raw commands, natural initial-state text, natural late queries, operation aliases, and the two-entry register table. It does not see canonical register identities, numeric role labels, typed initial states, typed operation sequences, terminal states, or answers.

Exact numeric-span boundaries, the explicit two-entry table, exact law support, the bounded Z/97 executor, and output comparison remain engineered scaffolds. CAB1 tests basis-coherent semantic binding; it is not an open-domain reasoning claim.

Frozen corpus

• Training seed: 2026080821
• Development seed: 2026080822
• Conditional confirmation seeds: 2026080823 through 2026080827
• Training: 100,000 generated records
• Development and each confirmation board: 256 episodes
• Each episode: 24 evidence transitions, 16 noncommuting programs at depths 12--32, and 32 late natural queries
• Table order is deterministic but pseudorandom and balanced independently of canonical identity
• Opaque names and source texts are disjoint from JRB1 training and every JRB1 development/confirmation board; JRB1's signed transitive overlap receipt binds the older EAL2/NCP1 lineage
• Every artifact is deterministically regenerated and hash-bound before model access

Confirmation files are materialized and sealed with the same generator and evaluator before development scoring. They may open only after an unchanged development PASS.

Frozen optimization

Two matched 290k-parameter-scale arms start from identical initialization and receive the same sampled rows, update count, batch size, and optimizer:

• Treatment: candidates are the two source-owned register names in random table order.
• Decoy-table control: candidates are two equally shaped episode-local names absent from evidence, initial state, and query while targets and all other inputs remain unchanged.

Each arm receives exactly 1,000 AdamW updates, batch 128, learning rate 0.001, weight decay 1e-4, and the unweighted sum of evidence, initial, and query cross-entropies.

Frozen development arms

1. Treatment: ordinary source and table.
2. Renamed: unseen register names consistently replace source and table.
3. Whole-table permutation: evidence, initial, state, and query owners all receive the reversed table. This is a positive equivariance test.
4. Cross-owner permutation: only evidence compilation receives the reversed table; initial state and query remain in the ordinary basis. This is the causal coherence-breaking control.
5. Source scrub: source register names are replaced by unrelated names while the ordinary candidate table remains.
6. Decoy-table model: the independently trained matched control is evaluated on the ordinary board.

No arm averages fields across bases. The executor sees only the committed law packet, table-relative initial state, NCP1 operation sequence, and table-relative late query.

Conjunctive gate

CAB1 passes development only when every condition is true:

• qualified EAL2 and NCP1 hashes and reports match;
• command exactness is at least 99%, with at least 95% at every depth;
• treatment, renamed, and whole-table-permutation evidence binding, temporal assignment, complete roles, initial binding, query binding, and law commit are each at least 99%;
• those three positive arms reach at least 99% terminal-state and answer exactness, with at least 95% terminal exactness at every depth;
• cross-owner permutation reaches at most 5% canonical terminal-state exactness and at most 55% answer exactness;
• source scrub reaches at most 30% evidence grouping, 35% initial binding, 55% query binding, 20% state exactness, and 35% answer exactness;
• the decoy-table model stays below those same calibrated ceilings;
• treatment/control initialization, data, charged examples, schedule, and configuration match exactly;
• checkpoint/report/data/runtime hashes match, parent weights remain bit-identical, typed initial/query carriers are absent, and runtime source has no exact register-name scanner.

Every one of the five confirmation boards must independently pass the unchanged evaluator. A development failure closes this exact CAB1 rule without width, duration, seed, threshold, renderer, or loss variants. One read-only attribution may locate the failed owner but cannot rescue the result.

Claim boundary and next decision

A confirmed PASS qualifies one content-addressed coordinate bus that preserves whole-state coherence across natural evidence, initial state, recurrent execution, and late query. It does not authorize continuation pretraining or a frontier-capability claim. The next successor would remove another explicit scaffold, selected from numeric spans, law support, or the exact executor.

A FAIL means this token-evidence bus does not clear the controlled composition gate under the frozen budget. Its exact failure is preserved and not tuned.

Immutable result

CAB1 is a conjunctive development FAIL. Matched treatment and decoy-table jobs 744838/744839 completed all 1,000 frozen updates from the same initial state. Both fit their fixed 2,048-row training sample exactly, but that fit did not establish content addressing. The unchanged development evaluator ran once in job 744841; its report SHA-256 is ffd29db045d22478367338253d3fc004695f63fdb5f5107eb6ee759652931bdc.

Program compilation remains exact at 4096/4096. The treatment misses the positive component, execution, and depth gates. Source scrub exceeds its answer ceiling, and the independently trained decoy model is far above every calibrated chance ceiling. The cross-owner intervention does produce the predicted causal collapse, but that one positive causal result cannot rescue the conjunctive failure.

The one permitted read-only attribution finds no recurrent-depth degradation. Treatment state exactness is flat at roughly 211/256 for every depth from 12 through 32 and follows the 211/256 episode-law commit count. Renamed and decoy arms show the same pattern at 222/256 and roughly 198/256. The failed owner is therefore episode-global register compilation/binding, not repeated execution. The training-name correction also gave real and decoy namespaces a shared deterministic serial/index suffix. The decoy's 77.258% answer rate is consistent with exploitation of that generator structure rather than source presence. This invalidates the intended matched control and is preserved as a data-construction defect; it does not explain away the treatment's own large miss.

No confirmation file was accessed. No width, duration, seed, threshold, renderer, loss, or corrected-control rerun is authorized. CAB1 closes exactly. The ordered successor is structurally different: an episode-local occurrence quotient supplies discrete identity addresses for exact repeated opaque names, while learned owners remain responsible for semantic roles. It is not a retraining or repair of the CAB1 similarity metric.