DIVERGE-SVE1: Spanless Value-Event Transduction

Status: confirmed across development and five fixed source-disjoint boards.

Capability Hypothesis

OQB1 established that exact repeated surface identity can be quotiented into an anonymous episode-local address while a learned owner attaches semantic value and query roles. SVE1 asks whether the remaining host-provided numeric spans and parsed integers can be removed. A byte-level model must emit complete, value-bearing transition and initialization events from raw source text. Its output, not a host digit parser, is the only value input to law compilation and state initialization.

SVE1 preserves the already-qualified OQB1 late-query owner and NCP1 command owner bit-for-bit. It changes only the evidence and initial-value interface. Long continuation pretraining and public-benchmark claims remain unauthorized.

Candidate Interface

The generic OQB1 occurrence quotient replaces the two declared raw register names with anonymous table-position markers and deletes the names. A shared 472,136-parameter two-layer bidirectional byte GRU emits CTC sequences through two heads:

• evidence: four tokens from role * 97 + value, where role is one of before-slot-0, before-slot-1, after-slot-0, after-slot-1;
• initial state: two tokens from slot * 97 + value.

The candidate runtime receives no numeric boundaries, integer values, typed initial state, terminal state, target register, or answer. It greedily decodes CTC outputs and fails closed unless the evidence roles are exactly {0,1,2,3} or initial slots exactly {0,1}. Exact bounded Z/97 law support, alias lookup, the modular executor, the two-name declaration, and exact repeated-surface identity remain engineered scaffolds.

Teacher-only frame alignment is allowed during training to accelerate the falsifier. Those alignments never enter the checkpoint interface, candidate runtime, development evaluator, or confirmation evaluator.

Frozen Data And Training

• training: 100,000 fresh records, seed 2026080841;
• development: 256 episodes, seed 2026080842;
• conditional confirmation: five 256-episode boards, seeds 2026080843 through 2026080847;
• treatment and shuffled-target control: identical initialization, architecture, data sources, sampler, 1,500 AdamW updates, batch 128, learning rate 0.001;
• shuffled control pairs every source with targets offset by 50,021 rows;
• all final source, name, and identity commitments must be transitively disjoint from the qualified OQB1 lineage and independently regenerate byte-for-byte.

The development and confirmation renderers, names, matrices, commands, depths, and identities are held out by the inherited generator split. Confirmation is not readable unless development passes every condition below.

Matched Arms

1. treatment;
2. unseen register renaming;
3. coherent table reindex across evidence, initial state, and query;
4. evidence-only table reindex;
5. value scrub replacing every digit byte without identifying spans;
6. occurrence-level quotient break;
7. independently trained shuffled-target model.

NCP1 program predictions and OQB1 query predictions are shared where the arm does not causally intervene on their interface.

Frozen Gate

Development passes only if all conditions are conjunctively true:

• NCP1 complete-program exactness is at least 99%;
• treatment, unseen rename, and coherent reindex each achieve at least 99% exact evidence events, initial events, query binding, law commits, terminal states, and answers;
• every positive transfer-depth cell is at least 95% exact;
• evidence-only reindex is at most 5% states and 10% answers;
• value scrub is at most 1% evidence events, initial events, states, and answers;
• occurrence break is at most 30% evidence events, 35% initial events, 20% states, and 35% answers;
• shuffled-target model is at most 1% evidence/initial events, 5% states, and 10% answers;
• schedules, charged examples, parameterization, initialization, data, reports, parent hashes, and checkpoint hashes match exactly;
• candidate runtime contains no numeric-span scanner or raw-integer parser.

A miss closes exact SVE1 without width, duration, seed, threshold, renderer, loss, alignment, vocabulary, or parser variants. A pass opens each fixed confirmation board exactly once; all five must pass unchanged.

Claim Boundary

A confirmed pass would qualify only spanless byte-to-value-event transduction inside this controlled bounded algebra and the already-qualified OQB/NCP system. It would not establish open-domain numerical parsing, alias/coreference equivalence, learned law search, unrestricted natural-program reasoning, or public benchmark capability.

Confirmed Result

Matched one-H100 jobs 744855/744856 completed all 1,500 frozen updates from identical initialization. Treatment reached 2048/2048 on its fixed training sample; the 50,021-row-offset shuffled-target control reached 0/2048. Treatment/control checkpoint SHA-256 values are 4d037d61b313d4237bf1d2309b00cf0905fe32478c2b399db56d2204d923d901 and 643ef2a9193ecb6a40d6957e3590bd0286f73329e24d86845624660e0bfebe15.

Development job 744857 passed every frozen condition. NCP1 programs were 4096/4096. Treatment, unseen rename, and coherent table reindex were each exact on 6144/6144 evidence event sequences, 4096/4096 initial event sequences, 4096/4096 terminal states, and 8192/8192 answers, with a 100% floor at every depth 12--32. Value scrub and the shuffled-target model produced zero complete events, states, or answers. Occurrence break produced zero states/answers. Cross-owner reindex preserved the local events but collapsed to 2/4096 states and 95/8192 answers. Development report SHA-256 is 93096b61b34528157ac6c1eb71dd43807aa2d87eff4bd1b077be0282051472d6.

The five fixed confirmation jobs 744860--744864 ran once without retraining and all passed. Aggregate treatment, unseen rename, and coherent reindex were each exact on 30,720/30,720 evidence sequences, 20,480/20,480 initial sequences and terminal states, and 40,960/40,960 answers. Value scrub, occurrence break, and the shuffled-target model each produced zero states and answers. Cross-owner reindex retained only 6/20,480 states and 400/40,960 answers. Aggregate SHA-256 is 41b31368e26d00dcb161c9f55c29f132520e78bff95d3a13e647254fd9603c4b.

SVE1 therefore qualifies the narrow mechanism claimed above. The next scaffold-removal target is the host's exact episode-law support solver; exact operation alias binding, bounded coefficient vocabulary, modular execution, and the two-name occurrence quotient remain explicit limitations.