DIVERGE-OPB1: Evidence Operation Pointer

Status: frozen before data materialization, training, or neural score.

Capability Hypothesis

DIVERGE-SNL1 confirmed the composition of model-owned spanless value events, neural unseen-law synthesis, natural command decoding, recurrent execution, and late query readout. Its largest remaining compiler shortcut is an exact whole-word search that assigns each evidence statement to one of eight episode-local operation aliases.

OPB1 changes only that interface:

A shared byte encoder and dynamic alias-table encoder can select the one operation named by a raw evidence statement, preserving SNL1's complete end-to-end execution under fresh opaque renaming and coherent table permutation without exact string search.

All qualified OQB1, NCP1, SVE1, NLS1, and executor weights remain bit-identical. OPB1 does not retrain or fine-tune them.

Candidate Mechanism

For source-token states h_t and episode-local alias states a_j, OPB1 computes normalized projected compatibility plus a learned source-token gate:

e_tj = exp(s) * cos(W_h h_t, W_a a_j) + g(h_t)

and commits one operation position with:

argmax_j logsumexp_t e_tj.

The runtime receives only raw evidence bytes and the eight-entry alias table. It contains no regex search, substring search, alias index lookup, operation label, numeric parser, or support solver. The selected hard positions group the already-qualified SVE1 value events before the frozen NLS1 synthesizer.

Frozen Data And Training

• 100,000 deterministic training rows, seed 2026080861;
• development seed 2026080862, 256 episodes;
• conditional confirmations 2026080863--2026080867, 256 episodes each;
• 1,000 AdamW updates, batch size 128, learning rate 0.001;
• treatment and independently initialized-identical decoy-table control;
• fresh names, sources, and episode identities disjoint from the complete SNL1 lineage;
• operation targets and scrub aliases exist only in the assessor/training supervisor, never in candidate-visible evaluation records.

The decoy arm receives eight fresh names unrelated to the operation present in the sentence while retaining the treatment target labels, update order, initialization, parameter count, and compute.

Frozen Arms And Gate

Positive arms are normal source/table, unseen full operation-plus-register rename, and coherent operation-table reindex. Each requires:

• operation binding at least 99%;
• evidence, initialization, query, and law packets at least 99%;
• terminal states and answers at least 99%;
• at least 95% terminal-state exactness at every held depth;
• all frozen NCP1 program paths at least 99%.

Causal controls are fixed:

• cross-owner operation-table reindex preserves local binding at at least 99% but must reduce states to at most 5% and answers to at most 10%;
• replacing the named operation by an unseen out-of-table name must reduce operation accuracy to at most 20% and states to at most 5%;
• the matched decoy-table model must remain at most 20% operation accuracy and at most 5% state accuracy.

The development gate is conjunctive. A miss closes this exact mechanism without width, seed, duration, threshold, renderer, or loss variants. A pass opens the five fixed source-disjoint confirmations once; all five must pass.

Claim Boundary

A confirmed pass would remove exact evidence-to-operation string binding from the controlled SNL1 trajectory. Exact repeated-register occurrence quotienting, the bounded 25-row coefficient vocabulary, and modular Z/97 execution would remain engineered. This is an architecture-mechanism result, not yet an open-domain or public-benchmark reasoning claim.