| name | vex-hyperframes-visual-repair |
| description | Diagnose and repair Vex HyperFrames renders with Scene Program V2, adaptive frame capture, blind/grounded/design critics, typed visual counterexamples, evidence-preserving patch operations, monotonic repair acceptance, and independent final judging. Use when a HyperFrames visual is irrelevant, generic, unreadable, semantically wrong, visually weak, or failing QA. |
Vex HyperFrames Visual Repair
Repair the executable visual program, not the rendered pixels and not free-form HTML.
Workflow
- Read
scene_program_v2.json, render_trace.json, and frame_contact_sheet.png.
- Compare
blind_critic.json, grounded_critic.json, and design_critic.json.
- Convert failures into typed counterexamples with exact element IDs, relation IDs, frame IDs, evidence IDs, and allowed repairs.
- Generate a signed
VisualPatchSet against the current program signature.
- Apply only the closed patch DSL and validate the complete scene program.
- Rerender and compare before/after snapshots.
- Keep the repair only when it is monotonic: fewer hard failures or a sufficient score gain with no quality, coverage, or issue regression.
- Run the independent final judge without repair history.
Hard Rules
- Never invent copy, facts, metrics, entities, interface states, or assets.
- Never rewrite arbitrary HTML, CSS, JavaScript, or shell commands.
- Never change claim-graph endpoints, relation types, evidence IDs, graph signature, or semantic signature.
- Never remove required objects or relations.
- Never publish a failed candidate because it is the best available candidate.
- Reroute to source footage, Manim, Blender, or no visual when the signed contract cannot be satisfied by HyperFrames.
References