DIVERGE-NPL2: Equivariant Natural Verified Plasticity

Status: PASS on the single development run and all five frozen confirmation seeds. The exact result and claim boundary are recorded below.

Capability hypothesis

Oracle PL1 established that 64 session scalars updated by branch-local, verifier-certified eligibility can recover unseen noncommuting mini-languages and retain that capability after all demonstrations are deleted. EIC1 now establishes a source-sensitive, renderer-invariant QUERY transaction owner; NVE1 already establishes the natural verifier-message EVIDENCE owner.

NPL2 tests the missing conjunction:

Natural WORLD programs, verifier feedback, and late queries can be compiled by the qualified owners into protected typed transactions that drive the unchanged PL1 session update and improve held-out deeper composition after every source string and feedback message is deleted.

Only the 8 x 8 PL1 policy matrix is mutable. EIC1, NVE1, the structural WORLD parser, exact executor, verifier, and query reader remain hash-checked and immutable. NPL2 adds no continuation pretraining or durable consolidation.

Execution boundary

The structural lexer identifies already-declared aliases, integer spans, and exact occurrences but has no operation mapping. Every legal verifier message for a public branch/attempt plan is passed through NVE1; exact scanners expose candidate spans, while NVE1 alone assigns STEP/VALUE and TARGET/DISTRACTOR roles. EIC1 compiles each late query into one of the two physical register identities through its confirmed involution projection.

The supervisor may compare candidate traces to the hidden episode only to produce a bound PASS/first-error receipt. Candidate policy code receives the natural message, its authenticated transaction commitment, and no hidden mapping, trace, terminal state, or answer. After acquisition, all source, feedback, branch transcript, and residual bytes are deleted; transfer retains only typed programs, the 64-scalar policy, and late typed QUERY transactions.

Matched arms and controls

NPL2 reuses the exact PL1 proposal seeds, eight branches, twelve attempts, verifier-call budget, write projection, score clip, and controls:

• STATIC, CONTEXT_ONLY, DIVERGE_ONLY, FAST_WEIGHT, TRANSIENT_GRAD;
• oracle PL1 as a ceiling;
• NPL2 with natural semantic transactions;
• shuffled feedback, wrong-branch credit, reset, unrelated state transplant, no eligibility, poison plus rollback, and protected-owner mutation.

The semantic compiler is run for every legal message before outcome selection, so hidden verifier outcomes cannot choose which messages receive model inference. Inference counts, verifier calls, writes, retained bytes, and wall time are reported.

Split and gate

Development uses the existing 256 source-disjoint NPL1 episodes. Conditional confirmation uses five new 256-episode seeds 2026080911 through 2026080915, generated and audited before any NPL2 score. Exact alias, episode, and natural-program identity overlap against all PL1 splits, NPL1 development, and between confirmation seeds must be zero.

The confirmation gate is conjunctive:

1. WORLD, EVIDENCE, and QUERY semantic receipts are each at least 99.5% aggregate and 99% per seed/renderer;
2. NPL2 transfer is at least 80% aggregate and 75% per seed;
3. NPL2 remains within five points of oracle PL1 aggregate and on 4/5 seeds;
4. NPL2 exceeds every non-oracle arm by at least ten points aggregate and by five points on 4/5 seeds, with paired bootstrap lower bounds above zero;
5. attempt 12 improves at least 50 points over attempt 1;
6. reset loses at least 25 points and returns within three points of STATIC;
7. shuffled, wrong-branch, and unrelated-transplant controls remain within three points of STATIC;
8. eligibility removal loses at least five points;
9. poison changes behavior and rollback restores every output/hash;
10. every protected owner hash, verifier receipt, and source-deletion check is exact.

Before confirmation, the single 256-episode development run must satisfy the same aggregate semantic floors, at least 80% late-query exactness, remain within five points of oracle PL1, beat every non-oracle arm by ten points, improve at least 50 points from attempt 1 to attempt 12, and reproduce the reset, shuffled-credit, wrong-branch, no-eligibility, rollback, and protected- owner controls above. Every legal verifier message is compiled before any branch outcome is selected; no hidden outcome may trigger model inference.

A development miss closes NPL2 without owner, rank, width, branch, attempt, seed, renderer, prompt, write-budget, or duration variants. A pass opens the five frozen confirmation seeds exactly once. Even a confirmation pass is a controlled synthetic reasoning result, not yet open-domain reasoning or authorization for long pretraining.

Result

Development job 744637 passes every condition. All 122,880 legal natural verifier messages are compiled before any branch outcome is selected; 122,872 are exact. EIC1 decodes all 8,192/8,192 late queries. After source deletion, NPL2 answers 84.8145% exactly, identical to oracle PL1, while the strongest non-oracle arm reaches 3.4668%. The immutable development report SHA-256 is 6a856799991ba744b480f54656d5e873f2db08fcca160d4859bd14442b576920.

Five independent one-H100 confirmation jobs 744638--744642 evaluate the pre-generated seeds without tuning. NPL2 equals oracle PL1 on every seed:

Aggregate late-query exactness is 35,066/40,960 = 85.6104%; exact terminal state-pair transfer is 85.3223%, and exact operation-map recovery is 85.0781%. Oracle PL1 is exactly 85.6104%; the strongest non-oracle arm is 3.9185%. Natural EVIDENCE compilation is 614,321/614,400 = 99.9871%, and QUERY compilation is 40,960/40,960.

The causal controls remain low: reset 1.0571%, shuffled credit 3.7646%, wrong-branch credit 1.0669%, unrelated-state transplant 1.0278%, and no-eligibility 1.1182%. Poison changes all 1,280/1,280 sessions and exact rollback restores all 1,280/1,280. Every protected owner and source-deletion receipt is exact. CPU aggregate job 744643 passes all fourteen conjunctive conditions. Aggregate SHA-256 is ef585debf543e137838ab92ef303740ad318b8eda7dd911f8a692a927aba5e79.

NPL2 therefore qualifies a controlled natural verified-plasticity reasoning mechanism: learned natural transactions guide coherent branch-local session updates, and the retained model-owned policy solves unseen deeper programs after demonstrations and feedback text are gone. It still relies on a structural WORLD lexer plus exact executor/verifier and is not unrestricted natural-language, math, code, or open-domain reasoning. No long continuation pretraining is authorized by this mechanism result alone.