docs/visual-judging.md

Visual assessment

The visual judge compares trusted reference and submitted-object renders. It scores visible components, shape and placement, surface appearance and rendering artifacts. Objective USD measurements remain separate. A visual score does not establish mass, inertia, contact behavior, hidden geometry or physical repair acceptance.

simrepairbench.visual_judge.judge_visuals returns not_run with a null score by default. Execution requires an explicitly authorised controller, an injected JudgeClient, a configured provider/model and passing objective gates (submission_valid, preservation_pass and objective_success). The adapter contains no provider SDK or credential lookup. A client must declare its provider_id, honour the supplied timeout and return the actual served-model identifier in ProviderResponse.

The verifier renders the submitted USD itself. Agent screenshots are not grading inputs. Keep the render directory outside the actor filesystem and pass its canonical path as trusted_root. load_trusted_render_evidence checks an independently supplied render-receipt hash, case identity, submitted/reference USD digests, paired view identities, matching image sizes and PNG content/hashes. It retains immutable image bytes for the client request. Obtain the expected receipt hash directly from the trusted render job, never from a submission.

The receipt uses schema: simrepairbench-trusted-renders-v1 and contains:

• case_id, candidate_asset_sha256, reference_asset_sha256 and a boolean fixture label.
• renderer: name, version, code_sha256, profile_id and settings, applying to all paired views.
• images: records with role (reference or candidate), view_id, a confined relative path and sha256.

Supported inputs are non-interlaced 8- or 16-bit PNG renders. Each view needs one reference/candidate pair. Keep camera, lighting and exposure matched in the render profile. Unsupported or occluded evidence must be marked unassessable, with a null criterion score.

The frozen rubric is visual-judge-rubric.json. Results follow visual-judge.schema.json, recording provider/model settings, rubric and renderer identities, input/request/response hashes, concise observations and reported usage. Missing model provenance, invalid output or a provider failure produces no score. The adapter computes the weighted score from validated criterion values and leaves it null if any criterion is unassessable. It never changes the objective result.

Offline tests use explicitly labelled fixtures and mocked clients. Fixture evidence requires an additional allow_fixtures flag and remains labelled kind: fixture; these records are not model evaluations or repair episodes.