WTV1: Whole-Trajectory Verifier Canary
Status: prospectively frozen before verifier scores
Date: 2026-08-10
Host: pinned Qwen3.6-35B-A3B
Scope: already-open source-disjoint development diagnostic; holdout sealed
Thesis
DSET, GSET, and ISET disagree on only 125 of 1,908 presentations. Every disagreement contains at least one exact complete trajectory, while the 14 rows missed by all three are unanimous. Their coherent whole-trajectory oracle is therefore 1,894/1,908 (99.266%). Majority voting is worse than ISET (1,830 versus 1,838), so frequency is not a useful commitment rule.
WTV1 asks whether the same frozen Qwen host can semantically score each distinct complete trajectory given the exact source and original model-owned draft. It selects one complete lineage; it never averages fields, edits a candidate, reads the gold answer, calls a verifier/tool, or repairs output at inference. Verdicts use counterbalanced A/B next-token log odds to cancel fixed label preference. The trained DSET checkpoint is used only to retain the qualified same-family host and its pinned NF4 loading path; no weights are updated.
Frozen canary
Only the 125 disagreement groups are scored. Exact duplicate trajectories are deduplicated before scoring. The remaining 1,783 unanimous rows are carried through unchanged: 1,769 correct and 14 wrong. Candidate construction, host, checkpoint, prompt, context length, counterbalancing, and selection by maximum score are immutable before output.
The canary passes only if all conditions hold:
- at least 105/125 disagreement groups select an exact trajectory;
- combined exactness is at least 1,874/1,908;
- choice exactness is at least 220/256;
- no verifier prompt is truncated.
A pass authorizes a causality control with source/draft associations shuffled and then one trained compact verifier gate on source-disjoint candidate data. A miss closes this zero-shot verifier without prompt, score, checkpoint, or threshold variants. No sealed holdout is opened by this canary.
Claim boundary
A pass is evidence that complete candidate lineages contain useful complementary reasoning and that a model-owned semantic commit can recover it. It is not yet a standalone architecture claim because it uses three trained candidate producers and an extra model pass. Compute, latency, and matched single-path controls remain required before promotion.