DSET1: Draft-Span Edit Transducer
Status: frozen prospective mechanism gate, 2026-08-09. DSEO1-v0 and the read-only DSEC0 binary-gating ceiling are closed. No DSET1 model output exists at freeze time. Development and holdout benchmarks remain unopened.
Why DSEO1 is insufficient
DSEO1 makes the edit action identifiable, but its full-response target remains
the same for clean and faulted drafts. The final generator can therefore learn
the action and still ignore it. Measured evidence is aligned action 94.73%
versus swapped 5.91%, while answer accuracy is statistically and practically
control-equivalent (89.21% aligned, 89.01% swapped, 90.82% final-only).
DSEC0 then grants the binary action a deterministic execution role: KEEP
copies the draft and FIX uses DSEO1's generated rewrite. It reaches only
1,870/2,048, ten answers above final-only, with 83.40% fault repair. Even
oracle actions reach only 84.57% fault repair. Binary selection is not the
remaining bottleneck; full-response rewrite is.
Changed mechanism
DSET1 replaces full-response regeneration with a model-owned edit script and a deterministic copy/edit transducer. The only legal scripts are:
<KEEP>
and
<REPLACE_LAST>
old_surface
new_surface
KEEP emits the exact visible draft. REPLACE_LAST finds the last exact
occurrence of old_surface in the draft, replaces it once with new_surface,
and emits the resulting complete trajectory. Empty or multiline surfaces,
missing old surfaces, malformed scripts, and unknown actions fail closed. No
verifier, solver, answer label, host repair, or hidden source access exists at
inference. The transducer is part of the decoder contract, analogous to a copy
head: the model owns every discrete edit field, while deterministic execution
only materializes those fields.
For a fixed source x, clean and fault drafts require different script tokens
and produce different executed trajectories. The target is therefore not
solvable by the DSEO1 shortcut:
[ (x,d_c)\to KEEP\to d_c, \qquad (x,d_f)\to REPLACE_LAST(o,n)\to d_f[o\mapsto n]. ]
The old surface o is draft-specific. A source-only model may infer n, but
cannot emit the exact script without observing the faulted draft.
Data repair and custody
DSET1 derives a fresh deterministic split from DSEO1's immutable, already
decontaminated 28,997-candidate source pool; it does not trust DSEO1's selected
pairs or clean_verifier_passed flag. A new CPU builder independently requires:
- clean and fault trajectories differ only at the registered span;
- the span is the final occurrence of the clean/fault surface;
- the registered span is the structurally final explicit-answer or boxed answer surface, with token-boundary checks that exclude LaTeX command letters;
- exact execution restores the complete clean trajectory, so scoring does not depend on an ambiguous broad answer extractor;
- both script presentations fit the pinned OLMoE 4,096-token context without truncation; and
- train and diagnostic source identities remain disjoint.
This explicitly removes the discovered DSEO1 corruption bug where an
incorrect one-letter metadata label could mutate the e in \\text{F} while
leaving answer F unchanged. Every dropped pair and reason is reported.
Host, initialization, and optimization
- pinned
allenai/OLMoE-1B-7B-0125-Instructrevisionb89a7c4bc24fb9e55ce2543c9458ce0ca5c4650e; - frozen attention, routers, experts, embeddings, and LM head;
- unchanged all-16-layer rank-18 shared post-MoE residual, exactly 1,179,648 trainables;
- immutable MPR1-hidden update-256 owner initialization;
- three arms trained independently for 512 updates, LR
2e-5, seed2026080916, one pair/microstep and eight pairs/update; - script-token CE normalized per presentation, then averaged across the pair;
- greedy script decoding with at most 32 new tokens.
The richer script receives 512 updates prospectively, not as a rescue chosen from output. All arms use identical pair order, source multiset, parameter count, optimizer, context, and update count.
Matched arms
- Aligned: correct script with the aligned visible draft.
- Within-source swapped: clean receives its paired fault script and fault
receives
KEEP; source, pair, token budget, and initialization are fixed. - Draft hidden: correct script target with the complete draft span causally hidden through every model layer.
The immutable DSEO1 final-only score (1,860/2,048) remains historical context,
not a DSET1 gate input. DSET1 uses a fresh diagnostic split and exact complete-
trajectory scoring, whereas DSEO1 used different identities and broad answer
extraction. Comparing those rows as if they were matched would be invalid. The
draft-hidden DSET1 arm is the equal-geometry source-only causal control.
Frozen gate
All conditions are conjunctive on the source-disjoint paired diagnostic:
- aligned exact script accuracy
>=90%overall and per corruption family; - aligned counterfactual pair consistency
>=90%; - aligned executed answer accuracy
>=95%overall; - aligned clean-copy accuracy
>=99%and fault-repair accuracy>=90%; - aligned exceeds swapped and hidden execution by at least 13 answers each;
- swapped and hidden exact script accuracy are each
<=60%; - zero accepted malformed edits, missing old surfaces, or non-final replacements; and
- complete data, checkpoint, parameter, token, memory, latency, and hash receipts.
A one-update mechanics pass is required before the three fits. Any gate miss closes exact DSET1-v0 without script syntax, seed, width, duration, layer, or threshold variants. No capability benchmark or holdout opens on a miss.
Claim boundary
A pass would establish that a draft-conditioned model can own a compact edit program whose deterministic execution materially improves complete answers on a MoE. It would not establish general planning, arbitrary-span editing, natural-draft fault localization, or large-MoE transfer. Those require later, separately frozen gates.