Light review — action optional. These items were auto-approved and are already live. They came from official government sources (Federal Register, FDA, USDA, Congress, state bills via LegiScan, etc.) where factual accuracy is reliable. Items stay in this list until you click "Looks Good" or until your next login resets the timestamp — on each login, you only see items published since your previous session. Looks Good dismisses from this list (item stays live), Edit opens the full edit panel to update any field inline (saves immediately), Escalate Tier bumps tier up one if underrated. Focus on Tier 3+ or anything misclassified.

This is your full backlog — every item ever marked pending that you haven't acted on yet. Items accumulate across pipeline runs; the pipeline number (e.g. "7") only counts what was added in the last run, not the total waiting here. Sorted by tier — T1/T2 first (highest urgency). Use Reject T5 to clear low-value horizon items instantly. For each item: Approve, Reject, Duplicate, or Edit & Approve.

Items you rejected in the last 24 hours. Review here if you rejected something by mistake. Use Restore to move an item back to pending for re-review, or Approve to publish it directly.

Action required. These are approved items already live on the platform that have hit their scheduled review date. T1-T2 items are flagged after 30 days, T3 after 40 days, T4-T5 after 60 days. For each item decide: Extend 30 Days (still active, nothing resolved), Mark Resolved (bill enacted/died, rule finalised, litigation settled), or Mark Dormant (no updates, no outcome — moves to archive, stays searchable, can be reactivated). You can also Edit any field before resolving. These are different from the Review Queue — they are already published.

All approved items currently on the platform, grouped by age. Click any tile to view and manage those items. Use Edit, Resolve, Dormant, or Remove to keep the feed current. Items in the 60-day and 90-day+ buckets need the most regular attention — anything without a recent update should be resolved or made dormant.

Duplicates are items the pipeline found that already exist in the database (matched by title prefix + source URL). They are discarded every single run — the pipeline runs every 4 hours, so the same article that appeared in the last run will be skipped in every subsequent run. A high duplicate count (150-200) is normal and healthy — it means the pipeline is consistently checking sources and correctly ignoring things it has already processed. Duplicates are never sent to Claude and never stored — they vanish silently each run. No action needed.

Historical view of all pipeline executions. Each row shows fetch/triage/approve/pending/duplicate/error counts. Click any row to expand source breakdown and error details.

Auto-reject runs at the start of each pipeline execution. Items older than the threshold (measured in pipeline runs, not hours) are automatically rejected and removed from the Approvals queue. T1/T2 items are rejected silently. T3–T5 items are rejected and logged.

The primary tool for improving signal quality. The Claude triage prompt determines how every item is classified. Edit it here and save — the next pipeline run will use the new version automatically. All previous versions are preserved and can be restored. The pipeline fetches the latest version from Supabase at runtime, falling back to the static triage-prompt.txt file if the table is empty.

Phase 3: Ingredient Surveillance. Tracks mention frequency and velocity per ingredient across all intelligence items. ⚡ ACCEL = 25%+ increase in 30-day mentions vs prior window. 🔬 = PubMed science velocity acceleration. Risk Score is weighted by tier distribution of mentions.

30-day publication counts from PubMed for tracked ingredient keywords. Updated at pipeline start. ⚡ ACCEL = 25%+ delta vs prior 30-day window.