← Complete research archive
Architecture researchTheory93 lines

DIVERGE-MZE1 Model-Owned Z/97 Executor

The confirmed CWC1 - EWC1 - NPL2 path should not need a hard-coded mapping from operation identity to state transition. A tied, outcome-supervised finite-field law owner can infer each opaque operation's complete 2x2 linear action and recurrently execute it without importing the …

docs/research/DIVERGE_MZE1_MODEL_OWNED_Z97_EXECUTOR.mdOpen original Markdown ↗

DIVERGE-MZE1 Model-Owned Z/97 Executor

Status: confirmed across development and five fixed confirmation seeds.

Capability hypothesis

The confirmed CWC1 -> EWC1 -> NPL2 path should not need a hard-coded mapping from operation identity to state transition. A tied, outcome-supervised finite-field law owner can infer each opaque operation's complete 2x2 linear action and recurrently execute it without importing the exact PL1 operation function.

MZE1 changes only the executor. Confirmed CWC1 selection, EWC1 structural transcription, NVE1 evidence, NPL2 branch-local plasticity, EIC1 late queries, data, seeds, prompts, verifier, and all thresholds remain fixed.

Architecture

For each of eight opaque operation identities and two output registers, MZE1 learns a categorical distribution over coefficient rows (a,b) with a,b in {-2,-1,0,1,2}. Training exposes only random input states and observed successor states. It marginalizes every row compatible with each observed outcome:

L = -log sum_{(a,b): ax+by = target mod 97} p(a,b | operation, output)

Autonomous execution commits the highest-weight row once and applies the same learned law at every recurrent step. The candidate runtime contains no import of diverge_pl1_data or the exact apply_operation function. Modular arithmetic and the bounded coefficient catalog are explicit architectural priors; the operation semantics are learned.

Frozen component gate

One treatment and one equal-parameter/equal-update shuffled-outcome control start from identical state and receive 256 updates of 4,096 transitions. The component passes only if:

  • treatment is exact on all 8 * 97 * 97 = 75,272 one-step transitions;
  • treatment is exact on 2,000 unseen programs at each depth 4, 8, 16, and 32;
  • the shuffled-outcome control is at most 5% exact against true operations;
  • the runtime source imports no exact operation implementation; and
  • parameter and update budgets match.

A miss closes MZE1 without coefficient-range, seed, update, learning-rate, or catalog variants.

Frozen composition gate

Only a component pass opens one development run. The learned transition owner replaces the exact transition by hash-bound injection before any NPL2 arm is evaluated. The exact verifier remains unchanged and is explicitly disclosed. The development report must pass every existing CWC1/EWC1/NPL2 condition and must additionally preserve the learned executor checkpoint/state receipt.

Only a conjunctive development pass opens the five already-fixed confirmation seeds. A confirmation pass qualifies removal of the hard-coded operation semantics from this controlled reasoning path. It does not remove engineered candidate generation, the exact verifier, the synthetic mini-language, or the finite-field architectural prior, and it is not a public reasoning claim.

Frozen result

The 400-parameter treatment and equal-size shifted-outcome control each train for 256 updates of 4,096 outcome-only transitions. Treatment is exact on all 75,272/75,272 one-step transitions and all 2,000 held programs at each depth 4, 8, 16, and 32. The shifted control scores 0.2657% against the true operations. The candidate runtime imports neither the PL1 data module nor the exact operation function. Checkpoint and component-report SHA-256 values are 0526e0e4...a9124c and 5846528b...349bf.

Development job 744698 passes the complete composition at 84.8145%, exactly equal to oracle, but takes 16m45s because it re-materializes the committed coefficient rows on every microstep. A semantics-preserving cache commits those same rows once. Replay 744700 completes in 4m34s and reproduces all score, control, gate, state, and custody fields other than elapsed time and packaging paths. Its result SHA-256 is 5fc8d0d3...34cec4.

Five fixed confirmation jobs 744701--744705 run concurrently, followed by aggregate 744706. Aggregate late-query exactness is 35,066/40,960 = 85.6104%, exactly equal to oracle on every seed. The strongest non-oracle remains 3.9185%; all original WORLD, EVIDENCE, QUERY, plasticity, reset, shuffled-credit, wrong-branch, transplant, eligibility, rollback, source-deletion, and owner-custody gates pass. Aggregate SHA-256 is 14f05b8a8a7f30f01418c0d00034a91e528da18395652304dd20d92e0325dbdd.

MZE1 therefore removes the hard-coded operation semantics from the strongest controlled Shohin path. It does not remove exact verification, explicit complete-candidate construction, the narrow mini-language, or the bounded linear-law catalog. Those are the next capability boundaries.