Project ShohinOPEN BUILD
Menu
← Complete research archive
Architecture researchPreregistered56 lines

ETV1: End-to-End Whole-Trajectory Verifier

TSVC1 reaches 99.61% validation on held-out deterministic clean/fault pairs but transfers to only 69/125 model-generated disagreements, exactly equal to the source-shuffled control. A frozen representation plus small head therefore learns corruption style rather than source-condi…

docs/research/SHOHIN_ETV1_END_TO_END_TRAJECTORY_VERIFIER.mdOpen original Markdown ↗

ETV1: End-to-End Whole-Trajectory Verifier

Status: prospectively frozen after TSVC1 failure and before ETV1 output
Date: 2026-08-10
Host: pinned Qwen3.6-35B-A3B
Scope: source-disjoint development; holdout sealed

Thesis

TSVC1 reaches 99.61% validation on held-out deterministic clean/fault pairs but transfers to only 69/125 model-generated disagreements, exactly equal to the source-shuffled control. A frozen representation plus small head therefore learns corruption style rather than source-conditioned semantic correctness.

ETV1 makes the verifier representation itself trainable. It retains the frozen Qwen base, router, and experts, but updates the existing 1,179,648-parameter shared post-MLP revision residual during a dedicated verifier pass together with the existing process-verifier head. Each training unit is the same source with one verified clean and one deterministic faulted complete trajectory; pairwise ranking and balanced binary losses force source-candidate comparison. At inference ETV1 scores each complete candidate and commits to one lineage. It cannot edit, regenerate, merge, execute, or consult a host verifier.

Frozen experiment

  • exact TSVC1-r3 train/aligned/shuffled candidate corpora;
  • pinned Qwen+DSET initialization and NF4/BF16 compute;
  • process-verifier leader scope, 300 updates, minimum 100, gradient accumulation 8, two candidates per identity, backbone LR 2e-6, head LR 2e-4, seed 20260809;
  • 4,096-token context with zero accepted truncation;
  • exact 125-group aligned and source-shuffled model-candidate diagnostics;
  • label-blind shape selector and frozen TSVC/WTV results as controls.

Gate

All conditions are conjunctive:

  • training, aligned evaluation, and shuffled evaluation have zero truncation;
  • internal final split on training identities selects at least 90%;
  • aligned selects at least 105/125 disagreement trajectories;
  • combined exactness is at least 1,874/1,908;
  • choice exactness is at least 220/256;
  • aligned exceeds source-shuffled by at least 13 disagreement rows;
  • aligned exceeds shape-only by at least 13 disagreement rows.

A pass authorizes one sealed holdout and candidate-producer consolidation. A miss closes this verifier family on the current data/host without scope, LR, duration, head, layer, seed, or prompt variants.

Claim boundary

A pass would establish a trained same-host semantic commit over coherent model-owned trajectories on a current MoE. It would still incur extra candidate and verifier passes; efficiency and transfer remain separate gates.