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.

1. Load, ECHO, and lock ids

This is a two-part gate. Parsing is not loading — a json.load that exits 0 only proves the file is well-formed, not that you have the findings in context. You must echo the actual content before any adjudication.

1a. Parse and echo the finding table.

Print every finding from the parsed findings[] array — not from memory, not from the branch name, not from surrounding files:

python3 - <<'PY'
import json
r = json.load(open('.beagle/llm-artifacts-review.json'))
f = r['findings']
print(f"git_head={r.get('git_head')} scope={r.get('scope')} count={len(f)}")
print("| id | file | line | category | description |")
print("|----|------|------|----------|-------------|")
for x in f:
    desc = (x.get('description') or '').replace('|', '\\|')[:80]
    print(f"| {x['id']} | {x.get('file')} | {x.get('line')} | {x.get('category')} | {desc} |")
print("ids=" + ",".join(str(x['id']) for x in f))
PY

Pass: the command exits 0 and the table (one row per finding) appears in your output.

The only source of findings is the parsed findings[] array. Do not infer findings from the branch name, the working directory, or surrounding files. If your mental model of the findings differs from the echoed table, the table wins — discard the mental model and adjudicate only the rows above.

1b. ID lock (hard gate, before any adjudication).

Record the exact set of ids from the ids= line above and state it now, e.g. Locked ids: {1, 2, 3, 4, 5, 6, 7}. This is the locked id set. Every result you write in step 4 must map 1:1 to this set: no id added, none dropped. The output id-check in step 4 references this locked set, not a re-derived one.

If, while verifying, you find yourself about to adjudicate a finding whose id is not in the locked set — or about to write a result for a file that does not appear in any locked row — STOP. That is an agent error (you are reasoning from memory or context, not the report). Re-read findings[] via the echo command above and restart adjudication. Do not record such a finding as false_positive (see step 3, Status discipline).

Record git_head and scope from the report (already printed by 1a). If the working tree no longer matches (optional strict mode: compare to git rev-parse HEAD), warn that line numbers may drift.

2. Order findings

Default order:

1. category == "dead_code" or fix_action == "delete" or risk == "High"
2. Remaining findings by (risk descending, id ascending)

With --priority-only, stop after processing category dead_code and all fix_action: delete (still write full output for those processed).

3. Verify each finding

For each finding, follow references/verification-checklist.md. Its first check for every category is the existence precondition: confirm the cited file exists at source_git_head before running any symbol/usage check.

Minimum evidence per finding:

• Existence first: Confirm file exists at source_git_head (git cat-file -e <head>:<file> or test -f). A nonexistent cited file is not routine — it is either a deleted-file finding (note it) or a sign you are not looking at the real report. A wall of missing-file results means STOP and re-read findings[] (step 1a). Do not absorb missing files silently.
• Read the file at the cited location and enough context to judge (parent symbol, imports).
• For unused/dead claims: search the repo (symbols, exports, string hooks) unless the issue is purely stylistic with no removal.

Pass: checks_performed lists only checks you actually ran (e.g. file_exists, read_symbol, ripgrep_symbol); notes cite the decisive observation.

Assign one status:

status	Meaning
confirmed_issue	The finding in the report is valid; acting on it is appropriate.
false_positive	The finding in the report is invalid (factually wrong, or harmful if "fixed"); do not auto-fix.
inconclusive	Needs human or product context; treat like risky in fix-llm-artifacts.

Set confidence: high | medium | low based on how direct the evidence was.

Status discipline (hard rule): false_positive means "the finding present in the report is invalid." It never means "this finding is not in the report." If you encounter an apparent finding that cannot be matched to an entry in the locked id set (step 1b), that is agent error, not a false positive — STOP, re-read findings[] via the step-1a echo command, and restart adjudication. Writing a false_positive (or any status) for an id outside the locked set is forbidden.

4. Write output

Create .beagle if needed. Write .beagle/llm-artifacts-verification.json:

{
  "version": "1.0.0",
  "created_at": "2026-04-19T12:00:00Z",
  "source_report": ".beagle/llm-artifacts-review.json",
  "source_git_head": "<from review>",
  "review_scope": "all|changed",
  "results": [
    {
      "id": 1,
      "status": "confirmed_issue|false_positive|inconclusive",
      "confidence": "high|medium|low",
      "checks_performed": ["file_exists", "read_symbol", "ripgrep_symbol", "export_trace"],
      "notes": "1-3 sentences of evidence"
    }
  ],
  "summary": {
    "confirmed_issue": 0,
    "false_positive": 0,
    "inconclusive": 0
  }
}

Validate the file you wrote:

python3 -c "import json; json.load(open('.beagle/llm-artifacts-verification.json'))"

Pass: command exits 0; re-open the file and confirm (a) summary matches results (count each status), and (b) the set of results ids equals the locked id set from step 1b exactly — no id added, none dropped. If the id sets differ, the pass is broken: do not ship; re-read findings[] (step 1a) and reconcile.

5. Summarize for the user

Print a short markdown table: id, category, original one-line description, verdict, confidence.

End with:

• Counts of confirmed vs false positive vs inconclusive.
• Recommendation: run fix-llm-artifacts only on confirmed (see that skill when verification file is present).

Rules

• Do not invent new issues; only adjudicate existing findings[] entries.
• Prefer inconclusive over confirmed_issue when removal could break dynamic or cross-repo usage.
• Preserve finding id values exactly as in the source report.

Integration

• fix-llm-artifacts: When this file exists, use it to skip false_positive ids and to treat inconclusive like risky fixes.
• fix_action custody: The fix_action field (refactor/delete/simplify/extract) is emitted by review-llm-artifacts and consumed by fix-llm-artifacts as a risk gate; verification carries it through unchanged and does not re-validate it.