# VeriHarness v0.2.3

Status: initial five-case verification-prompt ablation completed on 2026-07-18.

This small version is isolated from the completed VeriHarness v0.2 experiments. It owns its prompt
modules, MCP implementation, model-profile configuration, experiment manifests, runs, and analysis.

## Current module status

- Tool V1 · Initial MCP-Aligned: review candidate.
- Task V1 · Initial: review candidate.
- Verification V0 · Mentor Minimal: retained baseline comparison.
- Verification V1 · Initial Full: review candidate.
- Verification V2 · Overall Compact: retained comparison.
- Verification V3 · Two-Checkpoint Closure: active follow-up comparison.
- Verification V4 · External Verdict Authority: active iteration-2 candidate.
- Verification V5 · External Failure Adjudication: active iteration-3 candidate.
- Verification V6 · Atomic Confirmed Repairs: active iteration-4 candidate.

See [Prompt module index](prompts/README.md).

Iteration 1 (C3/V2 versus C4/V3, case5 + case25, three replicates) completed 12/12
generation cells and frozen-rubric offline evaluation. V3 increased median duration 2.86× and
verification-fix cycles by 9 without macro-score loss, but exact two-checkpoint closure held in only
3/6 V3 cells. See [Iteration 1 report](docs/results/20260719-verification-loop-iteration1-v0.1.md).

Iteration 2 (C4/V3 versus C5/V4) reached exact two-checkpoint closure in 6/6 V4 cells and retained
the long loop, but macro score fell by 0.307. See [Iteration 2 report](docs/results/20260719-verification-loop-iteration2-v0.1.md).

Iteration 3 (C5/V4 versus C6/V5) retained 93% median duration and 6/6 exact closure while improving
macro score by 0.250, but verification-fix cycles fell by 2.5. See [Iteration 3 report](docs/results/20260719-verification-loop-iteration3-v0.1.md).

Iteration 4 made the loop 2.44× longer and added 8 verification-fix cycles, but macro score fell by
0.423 and exact closure fell to 5/6. V6 is rejected. See [Iteration 4 report](docs/results/20260719-verification-loop-iteration4-v0.1.md).

Current follow-up: five-case final confirmation compares C3/V2 with the retained C6/V5 candidate.
- MCP: version-local implementation with inline text, UTF-8 text-file, and image inputs.
- Active frozen candidate dataset: [VDiff Gold + Discovery five-case set](data/datasets/vdiff-gold-discovery-5-v0.1/README.md).
- Preserved predecessor: [three-case set](data/datasets/vdiff-gold-discovery-3-v0.1/README.md).

## Completed initial ablation

- Manifest: [C0–C3 × five cases × three replicates](experiments/v023-verification-ablation-5case-v0.1.yaml).
- Conditions: C0 Tool+Task; C1 + Mentor Minimal; C2 + Initial Full; C3 + Overall Compact.
- Generation: 60/60 completed, first attempt, observed concurrency 10.
- Offline evaluation: 370 unique VDiff Gold+discovery rubrics and 4,440 candidate judgments.
- Result: [Chinese experiment report](docs/results/20260718-v023-verification-ablation-5case-v0.1.md).
- Machine analysis: [analysis.json](runs/v023-verification-ablation-5case-20260718-01/offline-evaluations/vdiff-gold-discovery-v1-20260718-01/analysis.json).

The prompt source files are intentionally not protected by expected SHA values. Every attempt
records the actual source files, resolved modules, assembled prompt, and their hashes in its
trajectory snapshot, so the completed run remains auditable without declaring the drafts frozen.