shadalyze-evidence-review

name	shadalyze-evidence-review
description	Review a Shadalyze single-event AI Evidence JSON exported from Unity Frame Debugger. Use to interpret embedded GLES source, Malioc vertex and fragment estimates, shader variants, and static-analysis limitations without reading files outside the evidence artifact.

Shadalyze Evidence Review

Read agent_contract.read_context before analyzing the artifact. Treat it as complete when context_is_complete is true; do not require this skill file to be available to analyze an exported JSON.

Procedure

1. Verify artifact_type and schema_version. State which required sections are absent.
2. Identify shader, pass, keywords, API, and Frame Debugger event from subject.
3. Analyze vertex and fragment separately. For each stage, use parsed for stable values and raw_report only when additional Malioc detail is necessary.
4. Compare static shader cost using total_cycles first. When longest_path_cycles is available, use it with bound_pipelines to attribute a modeled bottleneck; do not require it for cost comparison because variable loop counts, such as lights or reflection probes, can omit it.
5. Cite source evidence from sources.<stage>.content. If source is unavailable, say so; do not invent GLSL locations.
6. Produce the sections required by agent_contract.read_context.required_report_sections.

Interpretation

• BoundPipelines names the resource class on the modeled critical path. The bound-cycle value is a maximum of its listed pipeline cycle values, not their sum.
• A texture-bound finding requires Texture in the relevant path's BoundPipelines; high texture cycles alone are insufficient.
• A static Malioc estimate does not establish device frame cost, frame rate, cache behavior, thermal behavior, or whether a proposed optimization improves the final frame.
• Preserve vertex and fragment conclusions independently. A fragment optimization is not automatically the highest-impact frame optimization without draw count, coverage, and device timing.

Output Rules

For every conclusion, cite the exact JSON field and, where source is involved, source stage and line range. Rank recommendations by likely benefit and implementation risk. Each recommendation must name a validation capture or measurement that could falsify it.