0008: Keep Visual Acceptance Human-Authoritative

Status: Historical. The human-authority principle remains current; the detached capture-plan, adapter, baseline, and receipt machinery is superseded by ADR 0025, and visual work now routes through ordinary delivery under ADR 0048.

This record explains why human visual acceptance matters. It is not a current recipe for building a visual-rehearsal pipeline.

Context

Game visuals can regress without breaking compilation or deterministic logic tests. Fixed screenshots help, but ordinary golden-image systems often conflate pixel equality with quality, update baselines automatically, inspect only one view, or hide procedural and temporal failures behind a favorable sample.

STAGE needs a visual workflow that agents can plan and analyze while preserving human control over aesthetics, readability, feel, and product intent.

Decision

• Represent visual intent as cataloged subjects, named states, rubrics, and stable evidence coordinates.
• Combine fixed regression views with bounded, seeded exploratory views.
• Keep engine adapters responsible only for staging, rendering, and reporting raw output.
• Independently verify every artifact against the capture plan and adapter index before review.
• Use objective image checks and temporal, spatial, and procedural analysis to prioritize review, not to grant acceptance.
• Keep AI review advisory and explicitly record its model, session, observations, and limitations.
• Require an identified human reviewer for acceptance or rejection.
• Promote baselines only from acceptance-ready evidence and refuse overwrite of an existing baseline.
• Match later captures by semantic shot identity rather than filename or array order.
• Treat visual comparison as diagnostic evidence. A pixel change is not automatically a defect, and pixel equality is not automatically acceptance.
• Bind final evidence and the human decision in an immutable receipt.
• Keep authored assets and inspected repositories unchanged unless a named project operation explicitly owns capture fixtures.

Consequences

• Agents can automate broad, repeatable visual inspection without becoming the final art authority.
• Humans review a prioritized evidence set rather than manually hunting every angle and seed from scratch.
• Projects must implement or adopt an engine-specific capture adapter.
• Visual evidence consumes storage and requires explicit retention policy.
• Temporal quality, interaction, performance, accessibility, and game feel may still require video, profiling, hardware, or live playtest evidence.
• A baseline update requires a new review instead of a convenient automated rewrite.