DIVERGE-SOT1: Stage-Owned Epistemic Transactions

Status: frozen before SOT1 implementation, data materialization, training, or neural result on 2026-08-06.

1. Decision

IEM1 disproved the universal-owner integration hypothesis. One shared encoder fit the evidence and query training interfaces but lost mandatory source semantics: SET transferred at 0/2 and invalidated all 256 programs. A read-only splice restored the immutable TOL3 source owner and immediately recovered 3,072/3,072 evidence receipts, 256/256 sealed episodes, and 256/256 sensitive answers. The IEM1 query owner remained incomplete at 280/768 exact natural query transactions.

DIVERGE-SOT1 tests a materially different architecture:

Give each semantic stage an isolated parameter owner. Owners exchange only typed, provenance-bound transactions through a small epistemic microkernel. A parameter update can alter one owner but cannot silently rewrite another owner's transaction vocabulary.

SOT1 is not an IEM1 transport, width, loss, duration, seed, or renderer repair. IEM1 remains closed. SOT1 does not authorize continuation pretraining.

2. Architecture

Stage owners

One StageOwnedEpistemicMachine checkpoint contains three disjoint neural owners and an owner manifest:

1. WORLD source owner: the exact frozen TOL3 local semantic anchor at checkpoint SHA-256 b8b9dfe54b7ab4a31a74739625b8650fa4ee93a41221ab5d82610ebc1c030328. It owns declaration, update, predicate, swap, and structural query tokens.
2. EVIDENCE owner: the exact frozen NVE1 whole-mention compiler at checkpoint SHA-256 1610815471c695b0d2d198922dd99369e1f45a5dabc1b1c5d8e986b30fd200ff. It owns STEP/VALUE and TARGET/DISTRACTOR evidence assignments.
3. QUERY owner: one fresh position-free two-layer bidirectional byte GRU, width 192 and zero dropout, with one whole-group permutation head. It owns only the hard TARGET/DISTRACTOR query assignment. It shares no embedding, encoder, normalization, or head parameter with the other owners.

The transaction microkernel is deliberately semantically small. It verifies schema, owner-state hash, source/packet commitment, complete mention identity, and monotone stage order. It can route a valid transaction to the existing factorized rational state engine or reject it; it cannot infer, repair, or average semantic fields. Every transaction records the owner that produced it and the exact owner-state commitment.

Two-phase plasticity

SOT1 uses owner-local optimization. A proposed update is written to a candidate owner state first. The composite checkpoint commits only after:

• all non-target owner hashes remain bit-identical;
• the target owner's training receipt is complete;
• transaction schemas and manifest hashes validate; and
• the candidate contains no cross-owner parameter alias.

This first gate updates only the fresh QUERY owner. The qualified WORLD and EVIDENCE owners are immutable. Later architecture-aware training may make an owner plastic only through a separately frozen retention gate; SOT1 does not silently unfreeze them.

Execution boundary

The runtime path is:

natural WORLD -> WORLD owner -> sealed program/fault-line packet
natural EVIDENCE -> EVIDENCE owner -> sealed monotone receipts
packet + receipts -> factorized recurrent state engine
natural QUERY -> QUERY owner -> sealed late read transaction
terminal support -> exact answer if invariant, otherwise ABSTAIN

After the WORLD owner seals its packet, raw WORLD bytes and hidden states are absent from evidence, execution, and query stages. A query cannot alter support or execution. Incompatible worlds remain complete and are never averaged.

3. Data and training

The QUERY owner trains on the immutable 50,000-row IEM1 query corpus at SHA-256 8dc6085f7632b56563416fa75c13ff611344ad53c7c2ca4350c54bbcc17301fa. The one fit uses seed 2026080616, 1,000 AdamW updates, batch 256, learning rate 3e-3 with cosine decay, betas (0.9, 0.95), weight decay 0.01, and gradient clipping at 1.0. Loss is class-balanced cross entropy over one hard two-group permutation. No answer, state, evidence, operation, or comparator label reaches the QUERY owner.

The opened IEM1 board is development-only for integration debugging. The scientific SOT1 result requires one fresh deterministic 256-program board at seed 2026080617, materialized after this specification and before training. It retains 12 binary fault lines and three query modes per program but uses new program identities and at least three query surface templates not present in either query training or IEM1 confirmation. Exact WORLD, evidence, query, and identity overlap with training and prior confirmation must be reported and zero where applicable. No model score may select rows.

4. Controls

• immutable separate TOL3 + NVE1 + typed-query ceiling;
• closed IEM1 shared-owner checkpoint;
• premature highest-support top-1;
• equal-memory complete particles;
• no-evidence factorized support;
• query-owner role swap;
• evidence-owner numeric and symbol role swaps;
• shuffled complete evidence sets;
• declaration-state reset;
• owner-manifest swap and packet/query provenance swap; and
• post-seal WORLD/evidence/query poisoning.

The source and evidence owner state hashes are checked before and after query training and after composite serialization. Any difference is a fatal two-phase-commit violation.

5. Frozen pass/kill gate

All conditions are conjunctive:

1. source programs are at least 250/256 exact on the fresh board and at least 1,000/1,024 exact on protected TOL3;
2. natural evidence is at least 3,041/3,072 exact and protected NVE1 recovery remains at least 250/256;
3. natural queries are at least 752/768 exact, every accepted query is valid, and all three query modes are at least 245/256 exact individually;
4. at least 245/256 sensitive answers are exact, including at least 95% exact conditional on initially wrong top-1, with extensional parity on every answered episode;
5. no-evidence sensitive queries abstain at least 245/256, invariant queries answer at least 245/256, and partial-evidence underdetermined queries abstain at least 245/256;
6. top-1 and equal-memory particles trail SOT1 by at least 50 points;
7. relevant owner-role swaps, shuffled evidence, and state reset each reduce exactness by at least 50 points;
8. every manifest or provenance swap rejects, post-seal poisoning is bit-invariant, and zero invalid transaction, false commitment, malformed packet, gold deletion, overflow, or valid-execution rejection is accepted;
9. non-target owner hashes are bit-identical before/after training and after composite checkpoint reload; and
10. parameters per owner, examples, bytes, updates, checkpoint/data hashes, wall time, peak memory, canonical storage, transaction counts, and logical/unique state applications are reported.

A QUERY owner miss closes this exact isolated-query design without a width, duration, seed, renderer, optimizer, or loss variant. A component-retention or composition miss closes SOT1. A pass qualifies the transaction architecture for one broader free-form board and architecture-aware training; it does not itself prove general reasoning.

6. Claim boundary

Owner-specialized encoders, typed messages, version spaces, exact ALUs, two-phase commit, and global-workspace routing all have prior analogues. The candidate contribution is their source-sealed conjunction as a learning architecture: semantic stages have cryptographically accountable parameter ownership, plasticity commits atomically against interface contracts, and a factorized recurrent state machine consumes only coherent transactions. The first SOT1 result remains controlled-language evidence even if it passes.