verify-llm-artifacts

Verify LLM Artifacts Findings

Second-pass verification for .beagle/llm-artifacts-review.json. The detection pass optimizes for recall; this pass optimizes for precision so agents do not remove or “clean” code that is still required.

When to run

• After the review-llm-artifacts skill (especially full-project scans).
• Before the fix-llm-artifacts skill when findings include deletions, dead code, or High risk.
• Whenever past runs flagged artifacts that should not have been removed.

Inputs

• Required: .beagle/llm-artifacts-review.json from a completed review.
• Optional: $ARGUMENTS — --priority-only (verify dead_code and any fix_action of delete first; then others), --id N (single finding id).

If the review file is missing, exit with: Run the review-llm-artifacts skill first.

Prerequisite skills

1. Load the review-verification-protocol skill — general anti–false-positive discipline, including the Anti-confabulation gate (echo the artifact from a freshly read source before any verdict). The Load + ECHO gate in step 1 is this skill's concrete instance of that rule.
2. Load the llm-artifacts-detection skill — category criteria for what counts as a real issue.

Instructions

Hard gates

Objective pass conditions before you claim verification is done:

1. Input parse: The JSON load command in step 1 exits 0 (no traceback). Pass: valid JSON on disk at .beagle/llm-artifacts-review.json.
2. Echo before adjudicate: Step 1 has printed the full finding table (one row per findings[] entry: id, file, line, category, description) sourced from the parsed JSON in this turn. Pass: the table exists in your output and its row count equals len(findings) — you have not begun any verdict before it.
3. ID lock: Step 1 has recorded the exact id set from findings[] and stated it explicitly. Every results entry maps 1:1 to a locked id — none added, none dropped. Pass: the locked id list is printed; if at any point an apparent finding has no matching locked id, you STOP (see step 1, ID lock).
4. Evidence before verdict: For each finding you adjudicate, you have applied references/verification-checklist.md for its category (or documented why the category is N/A) and recorded matching strings in checks_performed. Pass: no status without at least one checklist-backed check or an explicit N/A note in notes.
5. Output contract: After writing .beagle/llm-artifacts-verification.json, the validate command in step 4 exits 0; summary counts equal the number of results entries by status; the results id set equals the locked id set from gate 3 exactly. Pass: schema-valid JSON and results ids == locked ids == source findings[] ids.