ISET1: Idempotent Span-Edit Transducer

Status: prospectively frozen before ISET1 model output
Date: 2026-08-09
Host: pinned Qwen3.6-35B-A3B
Scope: source-disjoint development; holdout sealed

Thesis

DSET's replacement value generator is qualified: externally selecting REPLACE repairs all 128/128 faulted choice examples. DSET and GSET fail because they first classify whether a draft is wrong. ISET1 removes that latent binary decision. Every invocation emits one complete replacement transaction:

• clean draft: REPLACE_LAST(old, old), an identity transaction;
• faulted draft: REPLACE_LAST(old, corrected), a corrective transaction.

The same strict deterministic executor applies both. The architecture therefore learns one sequence-transduction problem instead of a detector followed by a conditional generator. A valid final trajectory is draft-dependent because the emitted old surface must be present in the draft; there is no source-only fallback, verifier, semantic parser, answer label, or task router at inference.

Frozen experiment

The host, final-16 rank-18 post-MLP residual, NF4/BF16 loading, 1,179,648 trainables, 256 updates, LR 5e-5 cosine, four paired sources per update, 4,096-token custody, Qwen tokenizer, and source-disjoint identities are exactly the completed DSET-Q35 transfer settings. The only changed factor is the edit language and target objective.

Arms:

1. Aligned: clean identity transaction and fault corrective transaction.
2. Swapped: the two complete transactions are swapped within the same source pair, with identical geometry and compute.
3. Hidden: aligned transactions with draft keys causally unavailable.

The clean identity transaction is derived deterministically from the verified fault transaction's corrected surface and is independently executed against the clean draft before admission. Every train and diagnostic row must retain its complete prompt and 32-token transaction.

Frozen gate

All conditions are conjunctive:

• =98% exact executed trajectories;

• =95% exact scripts in every corruption family;

• =99% clean identity edits;

• =97% fault repair;

• =95% paired transaction consistency;

• aligned execution exceeds swapped and hidden by >=20 absolute points each;
• zero execution errors and zero decode exhaustion.

A pass opens exactly one sealed holdout. A miss closes ISET1-v0 without rank/seed/duration/layer/prompt variants.

Claim boundary

A pass would establish an explicit, model-owned, host-agnostic draft-to-edit transducer that transfers to a current MoE. It would not by itself establish general reasoning or public-benchmark improvement.