Skip to content

VALIDATION.md

Latest commit

 

History

History
83 lines (60 loc) · 3.1 KB

VALIDATION.md

File metadata and controls

83 lines (60 loc) · 3.1 KB
Raw

Validation

Release Scope

This repository distributes one Codex plugin containing one Agent Skill: antibes-holiday.

The release intentionally excludes:

  • third-party screenshots and source-account information;
  • copied compositions, characters, signatures, or logos;
  • rejected generation attempts and internal comparison sheets;
  • private absolute paths, credentials, task logs, and local configuration;
  • a bundled image model or undeclared external service.

Structural Checks

The following checks must pass before release:

  1. Codex quick_validate.py accepts the Skill.
  2. Codex validate_plugin.py accepts the Plugin.
  3. .agents/plugins/marketplace.json parses and discovers the Plugin.
  4. A clean ZIP extraction contains SKILL.md, both reference files, the calibration asset, seven selected example outputs, UI metadata, and the plugin manifest.
  5. Installed Plugin files match the release source.
  6. Repository-wide scans find no private paths, credentials, or source-platform URLs.

Forward Tests

The drawing system was tested across unrelated subject families:

  • a person without recurring accessories;
  • a non-human living form;
  • a functional object;
  • an abstract physical relationship;
  • an expressive display mark;
  • a multi-actor narrative scene;
  • a dominant-object narrative scene.

At least one isolated-subject test and one narrative-scene test used only the bundled non-semantic stroke calibration image. No third-party image was supplied at runtime for those tests.

Primary Acceptance Gates

Every generated candidate must pass:

  • physical pen behavior;
  • contrast between fast sweeps and slower recognition marks;
  • structural drift plus visible loss and recovery of control;
  • at least one imperfectly registered junction or displaced restart;
  • shorthand recognition through incomplete forms;
  • unforced proportion, spacing, and terminals;
  • transfer without imported motifs;
  • original subject and composition.

Narrative scenes must also pass:

  • causal readability;
  • hierarchy through scale, crop, density, and active blank space.

There is no aggregate score. Weak physicality cannot be offset by strong semantics.

Compatibility

  • Codex: full Plugin and Skill loading, with built-in image generation when available.
  • Other Agent Skills loaders: the portable SKILL.md, references, and asset remain usable when the loader preserves relative paths.
  • Other image-capable agents: use their native renderer.
  • Agents without rendering: output a renderer-ready prompt and validation checklist.

Platform-specific UI metadata under agents/openai.yaml may be ignored safely.

The public example gallery is documentation only. It is not loaded as style evidence by the Skill and is not required at generation time.

Residual Limitations

  • Image quality still depends on the renderer and may require focused iterations.
  • Small production logos require authored vector reconstruction and optical testing.
  • Abstract relationships require a concrete physical metaphor to remain inferable.
  • Claude Desktop does not discover Codex plugins directly; it needs a separate instruction or connector setup.