---
name: "Fable Forensic Skill — verify the artifact, not the performance"
description: "Use when a claim must be verified against the artifact, not the model's performance."
---

# Fable Forensic Skill — verify the artifact, not the performance

## When
Use when a demo, eval, or article might be theater and the file is the evidence.

## Stop
Never accept a narrative without the artifact. Never 'improve' the evidence.

## Steps
1. High assurance: production impact, destructive/expensive action, external publication, or hard-to-reverse work. Read references/verification-checklist.md first.
2. Standard: material consequence or multi-stage verification.
3. Light: low consequence and direct verification.
4. Establish the task contract (outcome, constraints, authorization, observable completion test). Never treat tool access as permission to go beyond the request.
5. Inventory evidence. Distinguish observed facts from inference. Expose decisive sources, assumptions, and gaps in the final response.
6. Decompose without losing ownership. Delegate only bounded packets. Treat every subagent response as an unverified claim. A subagent's `completed` status is workflow metadata, not proof of completion.
7. Execute within the boundary. Maintain a failure trail. Do not silently replace a failed method.
8. Isolate and verify the deliverable itself (render, execute, inspect diffs, reopen files). Do not accept the agent's own statement of completion as evidence.

## Load
Fable Forensic SKILL.md. The artifact files. Read-only.