# V1 / V2 / V3 Overview

| Project | Status | Case set | Primary metric | Anti-hack coverage | Notes |
| --- | --- | --- | --- | --- | --- |
| V1 | historical | early PDF-derived hard/all seeds | `holistic_quality_v1` | limited | Useful for process and trajectory history, not directly comparable to V3. |
| V2 | historical/reference | VDiff v2.2 `case1..case13` | `vdiff_rubric_v1` + `offline_quality_v1` | limited | Cached v-diff rubric artifacts and visual scoring did not headline absolute-position or screenshot-background hacks. |
| V3 | active, 70-cell complete | VDiff v2.3 `case-h1..case-h10` | Gate Score v2 (`vharness_gate_score_v1` compatibility id) | explicit computed DOM | Evaluator-owned shared grounding, Playwright text geometry, source-only semantic visual bank, and computed anti-hack audit. |

V3 metrics are not directly identical to V1/V2 cached-rubric metrics. Use this
table as a project-line overview, not a numeric leaderboard.

The canonical V3 result is:

```text
runs/real-v3-extra-hard10-gate-score-70-20260709-03/final-scores.csv
viewer-bundles/real-v3-extra-hard10-gate-score-70-20260709-03/index.html
```

Corrected V3 means are L1 `2.254`, L2 `2.248`, and L3 `2.934`. These
numbers replace the archived scaffold score, where unavailable L1/L2 geometry
was incorrectly counted as 5/5 and the visual bank was only low-level image
statistics.