DIVERGE-EAL2 Identifiable Temporal Interface

Status: qualified on development and five fixed independent confirmation boards.

Structural correction

EAL1 asked one reader to infer two independent coordinates for every numeric mention: temporal identity (BEFORE/AFTER) and episode-local register identity (X/Y). The source explicitly expresses temporal identity, but the random register names have no intrinsic X/Y meaning and EAL1 did not provide their episode-local table to the reader. Its 50% raw-role plateau and approximately 31% complete accuracy are the expected identifiability failure.

EAL2 does not change data, seed, width, duration, optimizer, coefficient catalog, law state, executor, controls, or thresholds. It assigns the two coordinates to their qualified owners:

• a 397,250-parameter byte-GRU owns only natural BEFORE versus AFTER semantics;
• exact whole-register scanning owns the disclosed episode-local X/Y table;
• their Cartesian composition yields the four complete transition roles;
• unchanged factorized support intersection induces each unseen operation; and
• unchanged sealed recurrent execution answers the depth-12--32 transfers.

This is a structurally different interface, not an EAL1 hyperparameter rescue. It asks the learned component to infer only information observable in its input while preserving model ownership of natural temporal semantics.

Frozen data, training, and controls

EAL2 generates a fresh 100,000-statement/256-episode corpus under seeds 2026080761 and 2026080762. Training and development use disjoint renderer pairs, but every individual before phrase and every individual after phrase appears in both splits; this holds out compositions without demanding lexical knowledge absent from a scratch byte reader. Sources, names, and matrix catalogs must have zero overlap and a second generation must be byte exact. Training remains 1,000 AdamW updates, batch 256, peak learning rate 0.003, cosine decay, and exactly half normal plus half temporal-counterfactual views. The full charged update budget matches EAL1.

Every EAL1 intervention remains unchanged: temporal counterfactual, temporal scrub, shuffled episode evidence, unrelated law transplant, reset, one-example underdetermination, and oracle temporal roles. The sealed packet still contains no demonstrations. The transitive EAL1/EAL2 candidate runtime must contain no exact operation table or runtime verifier.

Conjunctive gate

EAL2 passes only if all frozen EAL1 thresholds hold with temporal-reader semantics substituted for the impossible four-role reader:

• normal and temporal-counterfactual complete temporal assignment at least 99%, with at least 95% per renderer;
• temporal-scrub complete assignment at most 30%;
• learned law commit, coefficient-row, terminal-state, and query exactness at least 99%, with at least 95% at every held depth;
• oracle temporal roles give 100% law commit, state, and query exactness;
• one-example unique commits exactly zero;
• shuffled-evidence and unrelated-transplant state/query exactness each at most 5%;
• reset always abstains; and
• source deletion and runtime-source audits pass.

A miss closes EAL2 without local seed, width, update, learning-rate, renderer, or duration variants. A pass may open fixed independent confirmation boards and one composition with qualified Shohin owners. It does not authorize long continuation pretraining.

Development result

The first admitted neural job, Newton 744712, completed the unchanged 1,000 updates in 2m20s. The frozen 397,250-parameter reader is exact on all 6,144 normal and 6,144 temporal-counterfactual development statements, including a 100% renderer floor. Temporal scrub retains 1,738/6,144 = 28.2878% complete accuracy, below the frozen 30% ceiling.

All 256 learned episode-law packets commit. All 4,096 coefficient rows, 4,096 terminal states, and 8,192 late queries are exact, with 100% at every held depth from 12 through 32. Oracle temporal roles are also exact. One example never commits; reset always abstains; shuffled-evidence and unrelated- law controls reach 0/4,096 states and 76/8,192 queries. Source deletion and runtime-source audits pass.

Immutable hashes:

• checkpoint: c53cafe8478ab783ea4cd8d630a5f52245587d7957ed6565b960c48355636a36;
• training report: 4650a289111a61f72e73e3823bf6f8ec0400d8543f4f2350c65f2f79687c218a;
• development report: e4934f8447c9298663a6544c26b4aceff110a63f534b83e4bd81bd29f9585743.

Fixed confirmation

The same frozen checkpoint was evaluated without retraining on five fixed source-, name-, and identity-disjoint boards under seeds 2026080763 through 2026080767. The original training, development public, development assessor, and data-report hashes are bound before board generation. Each board contains 256 episodes and must independently satisfy the unchanged development gate. The aggregate passes only if all five boards pass, normal and counterfactual reading remain exact, temporal scrub remains at most 30%, and learned state and query execution remain exact.

Jobs 744713--744717 all pass in 22--46 seconds. Across the five boards, normal and counterfactual reading are each 30,720/30,720; learned terminal state is 20,480/20,480; and late-query execution is 40,960/40,960. Temporal scrub is 8,683/30,720 = 28.26497%. Shuffled-evidence and unrelated- law-transplant controls each retain only 4/20,480 states and 390/40,960 queries. Every per-seed original gate also passes.

The confirmation-data report SHA-256 is bafad3ddccaf66911eead6bdfe15e085a659481af071a18a2ac8b45a0c7b9816. The aggregate report SHA-256 is 3445cd0e797029fbbff88d1f597637ad8cab860fb9611f862cbb515a7a0dc953; its six-file manifest SHA-256 is 50c92146f86ecba38fb8d4c3faa1bf691c27efb759369273e5f95fff279a2f0d. All data and results are hash-matched and read-only locally and on Newton.

EAL2 therefore qualifies a controlled episode-local system-identification mechanism. This is not unrestricted language reasoning: exact register scanning, typed transfer programs, a bounded linear-law catalog, and the synthetic finite-field world remain engineered. The qualified reader should not be retrained; the next experiment must remove one of those larger scaffolds or integrate the mechanism into the strongest qualified Shohin path.