INPUT_FORMAT.md

Input format

Pass a JSON object to modeling.py --input. Image paths resolve relative to --image-root. Return values are pass, unknown, or a supported failure category (fail for keyboard checks).

Full-screen regression

Use example-regression.json. Supply the approved reference screenshot followed by current, and the actual viewport dimensions. Checks cover clipping, missing content, contrast, image distortion and minimum control dimensions. The task assumes one introduced failure category.

Local requirement

Set task to requirement_check. Provide:

• requirement_type: contrast, text_clipping, required_content, occlusion, image_aspect, or target_size.
• requirement, reference_target (description and bbox_xywh), and viewport.
• Images in order: reference, current, reference_detail, current_detail. Detail records also contain full_image_crop_xyxy.

Extend the reference target by max(48 pixels, 20% of its dimension) on each side, clamped to the screen. Apply the same crop to both screenshots. Use the approved reference to choose the target and crop.

Keyboard reachability

Set task to keyboard_access. Include a public target description, screenshot and observed_keyboard_trace when available. The trace must cover a complete Tab cycle to establish unreachability. A screenshot alone is insufficient.

Output

The JSON response includes choice, timing and token counts. semantic_support ranks the allowed answers; it is not calibrated confidence. Supply only observable evidence and requirements, without gold answers or mutation metadata. Inputs over 8,192 tokens are rejected rather than truncated.