07-reaudit

07 · Re-audit & Attribution

Closes the loop. Without this command, the pipeline is open-loop — actions go out, no proof they did anything. With this command, every action gets graded: did it move queries, which queries, by how much, with what confidence.

---

Inputs

• clients/<slug>/visibility_baseline.json (current — about to be replaced)
• clients/<slug>/baselines/round-<N-1>.json (previous round)
• clients/<slug>/distribution/log.jsonl (every action since previous round)

Output

• clients/<slug>/reaudit/round-<N>.json — validates against schemas/attribution_diff.schema.json.
• clients/<slug>/reaudit/round-<N>.report.md — human-readable monthly review.
• (side effect) clients/<slug>/baselines/round-<N>.json — new baseline preserved.
• (side effect) Updates clients/<slug>/visibility_baseline.json to the new round.

---

Procedure

Step 1 — Snapshot previous baseline

Before re-running audit, copy current visibility_baseline.json to baselines/round-<N-1>.json if not already there. Idempotent.

Step 2 — Invoke 02-audit

Call 02-audit with meta.audit_round = N. Use the SAME target_queries list as the previous round (do not silently drop or add queries — that breaks the diff). If brand_context.target_queries has changed, log the delta but keep the audit on the union for one round; signal to user that next round will adopt the new list.

After 02-audit completes, the new baseline is at clients/<slug>/visibility_baseline.json.

Step 3 — Load both rounds + the action log

PREV=clients/<slug>/baselines/round-<N-1>.json
CURR=clients/<slug>/visibility_baseline.json
LOG=clients/<slug>/distribution/log.jsonl

Filter log entries to those with shipped_at between PREV.captured_at and CURR.captured_at. These are the actions in this round's window.

Step 4 — Per-query diff

For each query in PREV.summary.per_query:

• Look up matching query in CURR.summary.per_query (by query_id, fall back to query string). Keep the query text byte-identical across rounds — query_id is derived from the query string (02-audit Step 1), so any edit (even punctuation) mints a new id and silently drops the query from the diff. To revise a tracked query, retire the old one and add the new under a fresh id rather than editing in place.
• Compute delta: mention_rate_delta, position_delta.
• Determine verdict_change:
  • improved — mention_rate up by ≥10pp OR position improved by ≥1.
  • regressed — mention_rate down by ≥10pp OR position worse by ≥1.
  • newly-won — was 0 mention_rate, now > 0.3.
  • newly-lost — was > 0.3 mention_rate, now 0.
  • stable — within ±10pp and ±1 position.