Project ShohinOPEN BUILD
Menu
← Complete research archive
Architecture researchTheory57 lines

DIVERGE-CVG1: Completion-Verified Whole-Lineage Gating

QPT1 is not a uniformly stronger or weaker model than B1. On the already-open 538-example development board, a static oracle that may select one complete B1 or QPT1 answer reaches 73.798% five-domain macro and 366/538 solved, versus QPT1 62.588% / 317 and B1 55.630% / 248. The un…

docs/research/DIVERGE_CVG1_COMPLETION_VERIFIER_GATE.mdOpen original Markdown ↗

DIVERGE-CVG1: Completion-Verified Whole-Lineage Gating

Status: conditional successor, frozen before implementation or score. Launch only if the exact SAG1 prompt-only gate closes.

Measured Motivation

QPT1 is not a uniformly stronger or weaker model than B1. On the already-open 538-example development board, a static oracle that may select one complete B1 or QPT1 answer reaches 73.798% five-domain macro and 366/538 solved, versus QPT1 62.588% / 317 and B1 55.630% / 248. The union retains 34/40 code answers and reaches GSM8K 96/100, MATH-500 64/100, GPQA 97/198, and BBH logic 75/100. This oracle is diagnostic, not a model result.

The observed bottleneck is therefore arbitration after a complete trajectory, not absence of complementary expert capability. SAG1 tests prompt-only arbitration. CVG1 is structurally different: it evaluates completed candidate trajectories before committing.

Architecture Hypothesis

  1. Keep the qualified B1 and pointer-transaction expert as frozen, coherent whole lineages.
  2. Generate one complete candidate from each lineage under identical decoding.
  3. A small model-owned critic reads prompt + candidate completion and emits calibrated probability of task success plus an abstention risk.
  4. Commit to exactly one whole candidate. Never average logits, tokens, or internal fields across candidates.
  5. If critic evidence is insufficient, choose B1 exactly.

The critic receives no evaluator task name, benchmark identity, gold answer, external solver output, or execution result at inference. It is trained only on source-disjoint rollout pairs whose outcomes can be established during training by exact math/code/logic/science assessors. Evaluation labels remain sealed.

Fast Gate

  • Freeze both generators before rollout collection.
  • Build a balanced outcome corpus containing base-only wins, expert-only wins, both-correct, and both-wrong pairs; split by source and normalized prompt identity before training.
  • Compare against always-B1, always-expert, random disagreement choice, teacher-forced-loss choice, and SAG1 prompt-only routing.
  • Require on the unchanged development board: at least 30/40 code, at least +3 macro points and +15 solved over the strongest frozen single lineage, gains in at least three domains, and no domain regression over two points.
  • Require calibration and selection accuracy on a source-disjoint rollout holdout before opening development evaluation.
  • A development pass opens one larger source-disjoint confirmation. A miss closes exact CVG1 without critic-width, threshold, seed, duration, or benchmark-label variants.

This lane tests whether model-owned post-trajectory verification can recover the measured 49-answer headroom beyond QPT1. It does not claim the static oracle score as achieved capability.