Imported from abdelb75103/GBI-DE-MVP (agent-instructions/extraction/AGENTS.md). Install upstream with npx skills add abdelb75103/GBI-DE-MVP --skill extraction. Copyright stays with the author.

FIFA GBI Live Extraction Instructions

Scope And Source Of Truth

Use this branch for manual Tabs 1–10 extraction, live apply, extraction QA, Backlog 2 batches, translated-paper extraction, population layouts, and extraction-status changes.

Follow skills/gbi-live-extraction/SKILL.md end to end. Load only the paper-relevant references it routes to. Track general review in fifa-gbi-data-extraction/docs/review-backlog.md and second-search Backlog 2 work in fifa-gbi-data-extraction/docs/second-search-extraction-review-backlog-2026-07-03.md.

Model And Write Rules

• Do not use Gemini unless Abdel explicitly requests it.
• Default to local GPT-5.5 medium when available; record any model substitution.
• Prefer additive-only writes. Do not overwrite nonblank extraction values unless Abdel requests a correction.
• Preserve studyId as papers.assigned_study_id.
• Assign selected records to AbdelRahman Babiker (00000000-0000-0000-0000-000000000001) unless instructed otherwise.
• Keep human-review candidates processing; preserve true exclusion statuses.
• Never change screening votes, resolver decisions, or protected manual-review fields during extraction.

An explicit extraction, review, apply, correction, refresh, or batch-processing request authorizes the skill's live apply and focused integrity gate unless Abdel says local-only, dry-run, or asks to review a proposal first. Verify live-backed corrections in the live record; do not stop at local JSON or Markdown.

Backlog 2 Parallel Batches

For multiple batches, use one coordinator plus one fresh-context agent per fixed five-paper batch:

1. Reserve every selected unassigned live paper in one conditional update.
2. Create numbered five-record batches and pre-create backlog rows.
3. Give each agent only its own five IDs using skills/gbi-live-extraction/references/parallel-batch-agent-prompt.md.
4. Each agent owns source review, pre-apply completeness, live apply, the focused integrity gate, and its five backlog rows.
5. Keep exclusions/systematic reviews in their original batch; never replace them.
6. Do not add cross-agent extraction review. Human review is the next substantive review.
7. The coordinator performs one final reconciliation of membership, status, assignment, population count/layout, and backlog rows. Trust a successful agent gate for source and field-level checks.

Do not repeatedly poll agents, inspect intermediate live writes, or rerun gates unless an agent reports failure, conflicting evidence, or a blocker.

Tracking And Review State

Use filenames and opening summaries that identify search wave, extraction stage, date, and artifact type. Persist only the batch input and final live audit by default; create a correction audit only when a post-apply correction was required.

When Abdel says a batch is complete, mark its rows ✅ reviewed_complete, set remaining in-scope live papers to extracted, and add the completion date. When he explicitly flags an extraction paper for exclusion, set live/backlog status to flagged, record the reason, and mark it reviewed complete only when he says review is complete.