Validation and before/after examples

Local release-candidate validation

The dashboard is generated from the reproducible commands and evidence recorded in VALIDATION.md. Hosted Windows, macOS, and Linux links remain pending until the public repository workflows run.

The release includes synthetic, rights-cleared fixtures so validation behavior is inspectable without downloading a model, engine, editor, or proprietary asset.

Sprite baseline and loop correction

The candidate changes frame placement only: pixel content, palette, dimensions, and alpha remain fixed. Baseline variance improves from 3 px to 0 px and the loop seam from 4 px to 0 px. The source is preserved and fixture approval is explicit.

Machine-readable evidence: representative-quality-evidence.json

Other representative gates

| Fixture | Evidence checked | |---|---| | Real-time mesh | Triangle budget, manifold topology, normals, UV overlap, LODs, collision | | Rig and locomotion | Unweighted vertices, foot slide, loop error, root-motion continuity | | Ambience loop | True peak, loudness, loop seam, clipped samples | | Greybox scene | Spawn-to-goal reachability, navigation, p95 frame time, memory, draw calls | | Visual regression | Different-pixel percentage, SSIM, camera drift, human review | | Gameplay smoke | Launch, core action, win/loss feedback, restart, crashes |

Run the evidence tests:

python -m unittest tests.test_quality_fixtures -v

These fixtures prove the validators and evidence contract, not the output quality of any external model. Production assets must generate their own artifacts under the same gates.