DIVERGE-JRB1: Joint Register Binder

Status: development FAIL; confirmation closed Parent: confirmed DIVERGE-NCP1 command owner and confirmed DIVERGE-EAL2 temporal owner
Objective: remove exact register-name scanning plus typed initial-state and query-register inputs in one integrated gate

Capability hypothesis

NCP1 already maps raw commands to ordered operation pointers. EAL2 already maps natural evidence to BEFORE/AFTER semantics. Their remaining end-to-end path still uses an exact string scanner to assign each numeric mention to one of two episode-local registers, receives a typed initial vector, and scores both output coordinates rather than interpreting a natural query.

JRB1 gives one shared learned owner all three register-binding duties:

1. assign four evidence numbers to the two episode-local registers;
2. assign two natural initial-state numbers to those registers; and
3. select the register requested by a natural late query.

The owner is a permutation-equivariant dynamic pointer. A shared byte encoder represents the source, another shared encoder represents the two episode-local register names, and normalized dot products score source mentions or a pooled query against the table entries. Evidence and initial-state paths share the same mention projection; the query has one task-specific projection. The model has no fixed output head for register names.

The qualified EAL2 reader remains bit-identical but no longer receives register IDs from its exact scanner. Its BEFORE/AFTER logits are constrained using JRB1's learned register groups. The qualified NCP1 pointer also remains bit-identical. Exact numeric span extraction, operation-name lookup during law induction, the bounded law solver, modular executor, and the two-entry register table remain engineered and are outside this claim.

Frozen data

• Training seed: 2026080811
• Development seed: 2026080812
• Conditional confirmation seeds: 2026080813 through 2026080817
• Training rows: 100,000
• Development and each confirmation board: 256 episodes
• Every episode contains three demonstrations for each of eight unseen linear operations, sixteen noncommuting programs at held depths 12 through 32, and two late natural queries per program.
• Register and operation names are opaque and episode-local.
• Training and evaluation use disjoint compositions of independently recurring initial-state and query renderer primitives.
• Training, development, confirmations, and prior EAL2/NCP1 boards must have zero source-text hash, opaque-name, and episode-identity overlap.
• Candidate-facing boards contain raw evidence, raw command, raw natural initial-state text, raw natural query text, and two dynamic tables. They do not contain typed initial states, typed operation sequences, register indices, numeric role labels, terminal states, or answers.

Frozen optimization

Two matched arms start from identical weights and receive the identical row order and schedule:

• Treatment: the dynamic register table is aligned with the targets.
• Shuffled-table model: the table is cyclically permuted during training while targets remain in canonical coordinates, deleting the only usable episode-local name-to-coordinate relationship.

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

Frozen evaluation arms

1. Treatment: normal register names and normal dynamic table.
2. Renamed registers: unseen names consistently replace registers in evidence, initial state, query, and table.
3. Table permutation: the same semantic table is presented in reverse order and predictions are mapped back to canonical identity. This is a positive equivariance test, not a negative control.
4. Register-source scrub: evidence, initial state, and query use unrelated decoy names while the candidate table remains original.
5. Shuffled-table model: the independently trained matched control is evaluated on the normal board.

The same frozen NCP1 program predictions are used for every JRB1 arm. Programs, initial states, law packets, and late query choices become typed only after the learned owners have run; raw source text is not consulted by the executor.

Conjunctive development gate

JRB1 passes only if every condition is true:

• qualified EAL2 and NCP1 parent hashes and PASS reports match;
• NCP1 program exactness is at least 99%, with at least 95% at every depth;
• treatment evidence-register, initial-register, query-register, complete-role, initial-state, terminal-state, and answer exactness are each at least 99%;
• treatment law commitment is at least 99% and terminal-state exactness is at least 95% at every depth;
• renamed-register evidence, initial, query, terminal-state, and answer exactness are each at least 99%;
• table-permutation semantic equivariance for those same outputs is at least 99%;
• register-source-scrub terminal-state and answer exactness are each at most 5%;
• shuffled-table-model terminal-state and answer exactness are each at most 5%;
• treatment/control initialization, data, update count, batch, and learning rate match exactly;
• checkpoints match their reports, parent weights are bit-identical, the JRB1 runtime contains no exact register search, and typed initial/query carriers are absent.

The five confirmation boards are opened only after a development PASS. Every seed must independently pass the unchanged evaluator. A development failure closes this exact JRB1 rule without width, duration, seed, threshold, renderer, or loss variants. Attribution may distinguish binding, compilation, execution, and query failure, but may not rescue the gate.

Claim boundary

A confirmed PASS would qualify controlled model-owned register binding across natural evidence, natural initial state, and natural late query, composed with previously qualified command and temporal owners. It would not establish open-domain language understanding, unrestricted arithmetic, general reasoning, or superiority on public benchmarks. A FAIL would show that this shared dynamic pointer does not remove the three register scaffolds under the frozen budget.

Development result

Final corpus report SHA-256 is d240756fa82052cc204a93596064524dec74ec0a3da8f84199f5106863ecc2dd; training SHA-256 is 2181593d3813636a398cd75090577d3e3aaf60f15269cfac509cef9de7000c38. Treatment/control jobs 744822/744823 complete the unchanged schedule from identical initialization. Treatment reaches 2048/2048 joint fixed-sample exactness. Its 290,177-parameter checkpoint SHA-256 is bc3fc348acac017e0782e09465036640b32819da5245811f418c2b8f3c56552e.

Frozen development replacement 744828 passes the substantive evidence, initial-state, and recurrent-execution path:

• parent NCP1 programs: 4096/4096;
• evidence register binding: 6144/6144 normal and renamed;
• complete transition roles and law commits: 6144/6144 and 256/256;
• natural initial-state binding: 4096/4096 normal and renamed;
• terminal recurrent state: 4096/4096 normal and renamed, with 100% at every depth 12--32;
• natural query binding: 7912/8192 = 96.5820% normal and 7844/8192 = 95.7520% renamed;
• exact natural-query answers: 7914/8192 = 96.6064% normal and 7846/8192 = 95.7764% renamed.

The conjunctive result is FAIL. Report SHA-256 is df8827dbea25428f843e06751bd722c0f52d362f7e8619550d23ffd9c3473171; confirmation remains unopened.

One read-only attribution closes the lane. Query errors are spread across all eight held renderer compositions rather than one missing primitive; the worst normal pair is 952/1024 = 92.97%. Mean-pooling the complete query is the remaining positive-path bottleneck. Register-source scrub gives 1395/6144 = 22.71% evidence, 1029/4096 = 25.12% initial, and 4044/8192 = 49.37% query exactness, which is near the relevant grouping and binary chance levels. Its downstream state/answer rates (15.70%/28.25%) show that the frozen 5% ceilings were not chance-calibrated.

The shuffled-table model commits all 256 laws but produces only 30/4096 canonical terminal states and 7811/8192 correct answers. This is the expected signature of one coherent global coordinate swap: the state tuple is permuted, while a correspondingly permuted query still retrieves the correct semantic answer. A fixed table rotation is therefore an equivalent representation, not an information-deleting control. The preregistered verdict remains FAIL; the result is not relabeled. The successor must represent state in table-relative content-addressed slots, focus query evidence token-wise, and break owner coherence only in causal controls.