TSVC1: Trained Source-Trajectory Semantic Commit

Status: prospectively frozen after WTV1 failure and before TSVC1 features
Date: 2026-08-10
Host: pinned Qwen3.6-35B-A3B
Scope: source-disjoint development; holdout sealed

Thesis

The edit arms contain an exact complete trajectory on all 125 development disagreements, but an untrained natural-language verifier selects only 56. TSVC1 replaces prompt-based judgment with a trained, small semantic commit head over frozen same-host hidden states. It receives the source and one complete candidate trajectory, scores candidates independently, and commits to one whole trajectory. It cannot edit, merge, regenerate, execute, or consult a verifier at inference.

The training objective is identifiable. Each training identity presents the same source with exactly two complete trajectories: the verified clean draft and its deterministic faulted counterpart. Exactly one is correct, candidate order is hash-randomized, and a pairwise ranking loss directly supervises the commit decision. The Qwen+DSET host stays frozen. Only the existing compact late-state correctness head is trained.

Frozen experiment

• Training: all 7,639 source-disjoint ISET training pairs, one clean and one faulted complete trajectory per source.
• Development: the exact 125 DSET/GSET/ISET disagreement groups from WTV1, with duplicate complete trajectories removed.
• Causal control: the identical development candidates with sources rotated within corruption-family and clean/fault buckets. A singleton bucket rotates within the same corruption family across pair member so every source is genuinely replaced.
• Baseline: the existing label-blind completion-shape reranker at matched candidate geometry.
• Frozen host features: final, tail-mean, and completion-mean pools from layers -1, -2, -4, and -8; complete source/candidate custody at 4,096 tokens.
• Head: existing 128-wide correctness head and frozen pairwise+BCE training implementation; no backbone, adapter, router, or expert update.

Gate

All conditions are conjunctive:

• zero source/candidate truncation in train, aligned, and shuffled features;
• aligned selects at least 105/125 exact disagreement trajectories;
• combined aligned exactness is at least 1,874/1,908;
• aligned choice exactness is at least 220/256;
• aligned beats the source-shuffled control by at least 13 disagreement rows;
• aligned beats the shape-only selector by at least 13 disagreement rows;
• the head's hash validation split on training identities selects at least 90%.

A pass authorizes one sealed holdout and an efficiency consolidation study. A miss closes this exact semantic-commit interface without head-width, layer, seed, duration, or loss variants.

Claim boundary

A pass would establish model-owned semantic commitment over complete source-conditioned reasoning trajectories on a current MoE. It would not yet show that three candidate producers are compute-efficient, nor establish a new reasoning primitive independent of candidate diversity.