GSET1: Paired Causal Fault Gate

Status: prospectively frozen before GSET1 model output
Date: 2026-08-09
Host: pinned Qwen3.6-35B-A3B
Scope: source-disjoint development; holdout sealed

Why this experiment exists

The trained Qwen DSET transfer executes 1,822/1,908 trajectories (95.49%) and beats the causally hidden-draft arm by 648 answers. Its exact gate fails on the choice family. Read-only attribution is unusually sharp: all 128 clean choice rows KEEP correctly; 79/128 faulted choice rows incorrectly KEEP; every one of the other 49 emits a correct replacement. Forcing the replacement action on the 128 faulted choice rows yields 128/128 exact scripts and 128/128 exact executed trajectories. The value generator is therefore qualified. The remaining defect is the decision owner.

GSET1 changes the architecture rather than retrying DSET rank, duration, seed, or prompt settings. It separates error detection from edit serialization:

1. The frozen DSET model encodes the complete source/draft presentation.
2. A small model-owned paired fault gate reads the final contextual state and emits KEEP or REPLACE.
3. KEEP deterministically copies the draft.
4. REPLACE is inserted as the causal first action token; the frozen DSET model generates the old/new surfaces, which a generic deterministic executor applies.

The final trajectory depends on the gate: forcing KEEP versus REPLACE changes the actual generation path. No verifier, answer label, task router, or semantic host repair exists at inference.

Frozen architecture and objective

The DSET checkpoint is fixed at SHA-256 166d8cbafd5fa5aa842a42c8856294436ed2704db3d51396a3554576399ea8bb. Its 1,179,648 residual parameters remain frozen. The gate is:

LayerNorm(2048) -> Linear(2048, 256) -> SiLU -> Linear(256, 2)

It has 529,154 trainable parameters. Training units remain paired by source: clean requires KEEP and the corresponding fault requires REPLACE. The loss is an equal-weight sum of action cross-entropy and a margin-2 paired contrastive loss on the actual KEEP/REPLACE logits. Training uses 1,024 updates, 256 pairs per update, AdamW at 3e-4 with cosine decay. Cached states are permitted because the Qwen base and DSET residual are immutable and hash-bound; no hidden state is used across train/development identities.

Matched arms

• Aligned: complete draft-visible state and correct paired labels.
• Swapped: identical aligned states, parameters, updates, and loss, but KEEP/REPLACE labels are swapped within every source pair.
• Hidden: identical parameters, updates, labels, and geometry, with draft keys causally hidden during both state extraction and replacement generation.

The already-completed forced-action diagnostic is attribution only and cannot make GSET1 pass.

Frozen development gate

All conditions are conjunctive:

• aligned action accuracy >=95% overall;
• aligned action accuracy >=95% in every corruption family;
• paired counterfactual consistency >=95%;
• aligned executed-trajectory exactness >=98%;
• clean copy >=99%;
• fault repair >=97%;
• aligned execution exceeds swapped and hidden by >=20 absolute points each;
• swapped and hidden action accuracy are each <=60%;
• no decode exhaustion;
• at most three literal old-surface execution errors, the known bounded DSET serialization boundary observed before this contract.

A conjunctive pass opens exactly one sealed holdout. A miss closes exact GSET1-v0; there is no width, seed, duration, margin, layer, or prompt rescue.

Claim boundary

A pass would establish a transferable, model-owned detect-then-edit temporal revision mechanism on a current MoE host. It would not establish general reasoning, public-benchmark improvement, or novelty of binary classifiers, low-rank adapters, or edit scripts in isolation.