OCET1: On-Policy Counterexample Edit Training
Status: prospectively frozen before proposal output, 2026-08-10
Scope: source-disjoint training/development; holdout sealed
Diagnosis and hypothesis
RIFT1's tied second pass is 99.84% syntactically valid and strongly draft-causal, but marks KEEP on many wrong model-generated proposals. Existing training used deterministic answer substitutions; evaluation asks the model to judge its own natural errors. The objective is therefore off-policy.
OCET1 changes the learning distribution, not rank, seed, duration, prompt, or decoder budget. The immutable aligned DSET-Q35 owner generates one mandatory rewrite for every source-disjoint training presentation. The generic executor materializes it. Training labels are then derived mechanically:
- exact executed proposal: KEEP;
- incorrect local executed proposal: REPLACE_LAST from its model-owned final surface to the independently verified final surface;
- malformed proposal: fail-closed to the original draft and its existing verified transaction.
The original textual draft and proposal feedback are removed before the next attempt. Only source, executed model-owned proposal, and training-only exact transaction remain. At inference, the system is the RIFT tied recurrent transducer with no verifier, label, solver, or semantic host.
Frozen stages
- Generate exactly one immutable aligned proposal per 15,278 training row (7,639 sources) with the pinned DSET checkpoint, greedy decoding, and the FRET executor. Hash every proposal and execution.
- Audit complete source/proposal/target retention, local edit derivability, source-disjointness, corruption-family balance, and zero holdout use.
- Warm-start the exact 1,179,648-parameter DSET residual, reset optimizer, and run one 256-update on-policy fit. No router/expert/base changes.
- Evaluate the tied two-transition architecture on the exact opened 1,908-row development board against DSET, ISET, RIFT, and a matched label-permuted on-policy control. Hidden draft remains the causal control.
Before capability output, exact gates will retain RIFT's minimums: >=1,874
overall, >=220/256 choice, >=945/954 clean, >=859/954 fault, +13 over
ISET and every newly trained matched control, >=95% valid scripts, and zero
exhaustion. One pass may open one sealed holdout. A miss closes OCET1 without
data-count, update, rank, layer, LR, seed, threshold, or recurrence variants.
Claim boundary
OCET1 tests whether closed-loop learning on immutable self-generated counterexamples makes a model-owned recurrent edit architecture operational. DAgger, self-training, edit models, and verifier-derived training labels are prior art; no primitive is claimed as novel. The possible contribution is the causal same-family draft-to-edit fixed-point system and its matched controls.