VTE1: Verified Transaction Equivalence Learning

Status: closed negative after the one frozen source-disjoint evaluation. KCR1, NDR1, and every earlier edit lane remain closed. Broad development, the draft-hidden control, and holdout were not opened.

Execution receipt: CPU job 750779 admitted 11,218 sources, 33,654 presentations, and 65,206 verified candidate transactions with zero truncation (maximum 3,107/4,096 tokens). One-update mechanics replay 750781 passed with finite loss/gradient, exact candidate grouping, 2,704,896 trainables, and 35.74 GB peak memory. Fit 750791 completed all 256 updates; dispatcher 750814, shards 750841--750844, and merge 750845 completed the one immutable source-disjoint canary.

Final result

Aligned VTE1 scores 1285/1566 = 82.0562% semantically, below the immutable KCR1 parent at 1294/1566 = 82.6309%. Presentation scores are natural-owner 379/522, verified-CONTINUE 414/522, and verified-KEEP 492/522; only 324/522 sources are correct in all three states. All 1,566 outputs are valid transactions, but the model emits RESTART on every row. Exact execution falls to 69/1566, KEEP byte preservation is 0/692, counterfactual action consistency is zero, and 197 generations exhaust the 768-token budget.

The set-valued objective therefore removed canonical-label punishment but collapsed to universal regeneration. It misses the semantic, per-state, parent-margin, all-three, KEEP-preservation, and exhaustion gates. Exact VTE1 closes without any temperature, candidate-family, delimiter, rank, layer, duration, seed, decoding, parser, or threshold retry. Exact hashes and gate receipts are preserved in SHOHIN_VTE1_RESULT.json.

Capability hypothesis

KCR1 imposed a unique latent program label even when several programs produce the same correct final semantics. Its source-disjoint canary contains 244 semantically correct outputs with a noncanonical action; 210 occur on expected RESTART rows. VTE1 tests one structurally different learning mechanism: maximize probability mass over the complete equivalence class of coherent, independently verified transactions, rather than cross-entropy to one arbitrary KEEP/CONTINUE/RESTART serialization.

This is not a retrospective KCR1 pass. KCR1 remains failed. VTE1 changes the training object from a labeled branch to a set-valued executable program and gets one bounded pass.

Immutable substrate

• Qwen3.5-9B revision c202236235762e1c871ad0ccb60c8ee5ba337b9a;
• KCR1 update-512 parent SHA-256 07e08abe2480782afc77e35031d23bea71a737d019f307066af2bde786dd2ebd;
• final-four rank-8 LoRA, exactly 2,704,896 trainables;
• the same 11,218 admitted training sources and 522 source-disjoint canary sources;
• 4,096-token complete-sequence custody, no truncation;
• frozen deterministic transaction parser/executor and semantic assessors;
• no verifier, solver, gold label, hidden host repair, or task route at inference.

Verified equivalence sets

For each source-local presentation, the builder reconstructs the visible draft and independently verified final trajectory from the immutable KCR1 source artifacts. It enumerates only transactions that the frozen executor and assessor certify:

1. KEEP iff the visible draft is already the verified trajectory;
2. CONTINUE with the exact missing suffix iff the verified trajectory has the draft as a byte prefix;
3. RESTART with the complete verified trajectory for every presentation; and
4. for non-code wrong natural drafts only, CONTINUE with a fixed correction delimiter plus the complete verified trajectory, admitted only when the independent assessor confirms the executed result.

Duplicate byte-identical transactions collapse. Every retained source must keep all three presentations, every equivalence set must be nonempty, every candidate must parse and execute, and every candidate must independently score correct. The runtime row exposes only the source/draft prompt plus its verified transaction set during training. Assessor fields are never model inputs.

Set-valued objective

For transaction t, ell(t) is KCR1's normalized loss: action-marker and payload/EOS spans each receive one half of presentation mass when a payload exists; KEEP averages marker and EOS. With frozen temperature tau = 0.1,

L_eq(E) = -tau * log((1 / |E|) * sum_{t in E} exp(-ell(t) / tau)).

The model therefore needs to assign high probability to at least one complete verified transaction, without being penalized for selecting a different semantically equivalent lineage. There is no auxiliary action classifier. The generated transaction is parsed and executed exactly once at inference.

VTE1 resets optimizer state from the immutable KCR1 parent and runs exactly 256 updates, batch 1, accumulation 8, LR 2e-5, seed 2026081021, and data seed 2026081020. One update must first prove finite gradients, candidate grouping, charged-token accounting, and zero protected-weight mutation.

Matched arms

1. aligned verified equivalence sets;
2. identical fit with the entire draft span causally hidden at every layer;
3. immutable KCR1 update-512 parent with no VTE training.

The hidden arm uses identical prompts, candidate sets, parameters, updates, optimizer, and token/FLOP schedule. No action-label permutation is meaningful once labels are quotiented by verified execution.

Frozen source-disjoint gate

On the exact 1,566-row KCR1 canary, aligned must satisfy all conditions:

• at least 1,410/1,566 = 90.04% executed semantic correctness;
• at least 85% semantic correctness on each of verified KEEP, verified CONTINUE, and natural-owner presentations;
• at least 99% valid transactions and 99% KEEP byte preservation;
• at least 418/522 = 80.08% sources semantically correct in all three states;
• at least 78 answers over the immutable KCR1 parent and at least 13 over the matched draft-hidden fit;
• at most ten decode-limit exhaustions; and
• complete set-cardinality, parameter, token, FLOP, memory, latency, and hash receipts.

Only this conjunctive pass opens broad development. Broad development retains KCR1's 603/1,289 total and 223/349/17 domain floors, requires at least 13 answers over the hidden arm and immutable KCR parent, at least 13 net repairs on wrong exhausted drafts, and no more than 400 payload exhaustions. Only a conjunctive broad pass opens one sealed holdout.

Stop rule and claim boundary

Any CPU admission, mechanics, source-disjoint, or broad gate miss closes exact VTE1 without temperature, candidate-family, delimiter, rank, layer, duration, seed, decoding, parser, or threshold variants. A pass would establish that set-valued executable supervision resolves KCR1's branch-label nonidentifiability and improves model-owned temporal revision. It would not prove optimal planning, unrestricted reasoning, or a native 125M mechanism.