DIVERGE-NCP1 Natural Command Pointer

Status: confirmed across five fixed source-disjoint boards.

Capability hypothesis

EAL2 still receives each transfer program as an engineered list of operation symbols. NCP1 removes that list:

A content-addressed CTC pointer can read one raw variable-length natural command, bind unseen episode-local operation names against an eight-entry alias table, preserve operation order and repetition, and drive the frozen EAL2 law packet and recurrent executor without typed program symbols.

The 273,794-parameter pointer uses a shared byte embedding, separate bidirectional GRUs for the command and alias table, normalized dynamic pointer scores, one learned blank score, and greedy CTC collapse. The output classes are episode-local alias positions rather than a global operation vocabulary. Training-only alias spans provide frame supervision plus an auxiliary CTC loss; inference receives no spans, target length, typed symbols, or alignment.

This is a practical scaffold-removal gate, not a novelty claim for CTC, pointer networks, dynamic vocabularies, or recurrent encoders. The qualified EAL2 temporal reader, exact register binder, exact law-support solver, typed initial state/query, and exact Z/97 executor remain frozen.

Bounded feasibility calibration

Before freeze, one explicitly non-claim pilot established that the interface can learn. A 10,000-row treatment reached 2,048/2,048 fixed-sample program exactness after 1,000 updates; its matched shuffled-table arm stayed 0/2,048. On the pilot board, normal, reverse-order, and end-to-end state exactness were 99.6826%, while shuffled table, source scrub, and shuffled-table model were 0%. Renamed aliases exposed a generator-distribution mismatch. Correcting only that pre-freeze generator to use the same random consonant-vowel carrier as EAL2, then using 1,500 updates, produced 4,096/4,096 exact normal, renamed, and reversed programs and 4,096/4,096 terminal states at every depth 12--32; all three negative controls remained 0/4,096.

Those opened pilot boards cannot count as final evidence. Final development is rotated to fresh seed 2026080798; conditional confirmation seeds are 2026080799--2026080803 and must be built before final development scoring.

Frozen data and schedule

Training seed 2026080791 generates 100,000 command records at depths 4--20. Every record has eight fresh, randomly permuted opaque aliases and one raw command rendered from held-composition action/connector pairs. Development and confirmation use fresh EAL2 episodes and commands at depths 12--32. The typed symbols list is removed from every candidate-visible transfer. All final sources, names, and episode identities must be disjoint from EAL2, NLS1, the opened pilot boards, and each other, and every final byte must independently regenerate.

Two matched 273,794-parameter arms start from the same state and train for exactly 1,500 AdamW updates, batch 128, learning rate 0.001:

• treatment receives the correct episode alias table;
• shuffled-table control receives the same commands, targets, minibatches, and span supervision, but its alias table is cyclically shifted.

Both arms receive the same parameters, data, update count, optimizer schedule, and charged examples. EAL2 is not retrained.

Frozen development gate

All conditions are conjunctive:

1. inherited EAL2 temporal reading and law commitment are each at least 99%;
2. raw-command program exactness is at least 99%, with at least 95% at every depth 12--32;
3. independently renamed aliases and reversed command order are each at least 99% program exact;
4. source-deleted terminal-state and late-query exactness are each at least 99%, including at least 99% state exactness after alias renaming;
5. command-source scrub, shuffled alias table, and the independently trained shuffled-table model are each at most 5% program exact;
6. initialization, parameters, data, updates, batch, and optimizer schedule are matched; checkpoint/report custody is exact; and
7. the candidate runtime contains no exact alias search, typed program list, target length, or training alignment at inference, and command bytes are absent before execution.

A development miss closes NCP1 without width, layer, update, seed, learning-rate, renderer, decoder, threshold, or duration variants. A pass opens the five already-built confirmation boards exactly once with unchanged weights. NCP1 does not authorize continuation pretraining or an unrestricted reasoning claim.

Final result

The final corpus independently regenerates byte-for-byte and has zero source, opaque-name, or episode-identity overlap against EAL2, NLS1, both opened pilot boards, development, or any other confirmation seed. Training contains 100,000 rows at SHA-256 3078eb9f14abfa0bcc54dd97cc36e6d45a4fc8fdbe4c1ffc81a03c1cbe2c28ef.

The first treatment allocation, job 744741 on evc33, exposed no CUDA device and was canceled without a scientific result. The matched shuffled- table control 744742 and exact unchanged treatment replacement 744745 complete 1,500 updates from the same initial state. Treatment reaches 2048/2048 on its fixed training sample; control reaches 0/2048. Treatment/control checkpoint SHA-256 values are f830757fb2c735909732c75b8ce042a41993f4f47c59fe39f58e697381c0da46 and ec5f3bde4d9f4a8bbcd4c210463123a04b032ce685f9696544187a00b827dc19.

Development job 744797 passes every frozen condition. Normal, independently renamed, and reverse-order raw commands are each 4096/4096 program exact with a 100% depth floor through depth 32. Normal and renamed execution are each 4096/4096 terminal-state and 8192/8192 late-query exact. Source scrub, alias-table shuffle, and the independently trained shuffled-table model are each 0/4096 program exact. The inherited temporal reader is 6144/6144 and all 256/256 episode laws commit. Development report SHA-256 is be9347aed80529e5f6c4651973bb1d1d1199b180c1daf85ad5f8d21d8a60f292.

Five fixed confirmations under seeds 2026080799--2026080803 all pass with unchanged weights and evaluator. Aggregate normal, renamed, and reversed programs are each 20,480/20,480; normal and renamed states are each 20,480/20,480; normal and renamed queries are each 40,960/40,960. Each information-breaking program control remains exactly 0/20,480. Aggregate SHA-256 is ca02448696540a3b58d4baf0e944cb6abfbfdde5b1a74d761b82975c092ff8da.

NCP1 therefore qualifies the narrow capability claim: a learned natural- command owner can replace an engineered typed transfer-program list while preserving unseen alias binding, variable order and repetition, recurrent execution, and late-query correctness. It does not qualify unrestricted language reasoning. The explicit alias table, exact register binder, exact law-support solver, typed initial state/query, bounded algebra, and exact executor remain engineered.