# V3 Gate Score V2 Audit

## Contract

The promoted contract revision is
`vharness_gate_score_v2_shared_grounding_semantic_bank_computed_dom`.

- Each source case has one evaluator-owned grounding bank generated without a candidate image.
- All seven cells for that case use the same target text blocks and bounding boxes.
- Candidate text boxes come from Playwright `getBoundingClientRect()` relative to `.page`.
- Each case has up to eight target-only semantic visual questions and cached target answers.
- Candidate answers are compared by symmetric answer equivalence.
- Missing grounding, DOM measurement, VQA answer, or equivalence evidence is `unavailable`, never an automatic pass.
- The earlier scaffold score and first r1-derived-grounding reevaluation remain archived and are not headline results.

## Results

| Level | Cells | Gate mean | Gate >= 4 | Semantic visual | Geometry | Structure |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| L1 direct | 10 | 2.254 | 0/10 | 3.876 | 0.292 | 3.124 |
| L2 native Claude SDK | 30 | 2.248 | 0/30 | 3.536 | 0.236 | 3.165 |
| L3 v-harness v2.3 max6 | 30 | 2.934 | 1/30 | 4.439 | 1.029 | 3.498 |

L1 and L2 are effectively tied. L3 leads L2 by `+0.686`, including `+0.903`
on semantic visual checks and `+0.793` on computed text geometry. The advantage
is meaningful but not sufficient: only one L3 cell reaches the strict 4/5
threshold, and all 30 L3 cells exhaust six epochs without native gate acceptance.

## Anti-Hack Findings

| Level | Visible absolute/fixed text | Hidden text | Large media/background |
| --- | ---: | ---: | ---: |
| L1 | 8/10 cells | 2/10 | 4/10 |
| L2 | 27/30 cells | 4/30 | 8/30 |
| L3 | 12/30 cells | 0/30 | 12/30 |

These are audit flags, not all confirmed malicious hacks. In particular, a
large image can be a legitimate page asset. The finding remains important when
the media covers most of the page, refers to a target-derived background, or
coexists with weak live-text geometry. Absolute positioning is likewise not
forbidden for decoration, but visible text pinned by coordinates violates the
V3 flow-layout requirement and is scored as a structure failure.

## Interpretation

The corrected evaluator does not support the earlier claim that L1/L2 beat L3.
L3's repeated target feedback improves semantic fidelity and reduces visible
absolute-text usage, while the independent geometry bank shows that much of the
remaining layout is still spatially inaccurate. L1 and L2 often look plausible
to a holistic visual judge but rely heavily on fixed positioning and score very
poorly on independently grounded text geometry.

The independent grounding also prevents an L3 replicate from being evaluated
against the exact grounding cache it optimized during generation. This matters:
an intermediate r1-derived evaluation gave L3 r1 a geometry mean of `4.048`
versus about `1.42` for r2/r3. With evaluator-owned grounding, the replicate
means are `1.302`, `0.878`, and `0.908`, a much more credible spread.