extraction-form（提取形式）

Extraction Form (evidence review)

Goal: create a consistent, analysis-ready extraction table that is directly grounded in the protocol.

Inputs

Required:

• papers/screening_log.csv
• output/PROTOCOL.md

Optional:

• papers/paper_notes.jsonl (if you already have structured notes)

Outputs

• papers/extraction_table.csv

Workflow

1. Determine the included set

- From papers/screening_log.csv, collect all rows with decision=include.

1. Build/confirm the schema

- Use the extraction schema defined in output/PROTOCOL.md. - If the protocol does not define fields yet, stop and update output/PROTOCOL.md first.

1. Populate papers/extraction_table.csv

- One row per included paper. - If papers/paper_notes.jsonl exists, use it as a structured source for values/provenance (but keep the table schema governed by output/PROTOCOL.md). - Always include provenance columns: - paper_id, title, year, url - For each protocol-defined field: - fill concrete values (units explicit) - use an explicit sentinel for unknowns (recommended: empty cell + notes)

1. Keep it auditable

- If a value is inferred (not directly stated), mark it in a notes column. - Do not write synthesis; only extraction.

1. Quick QA

- Ensure 1:1 coverage: included papers == extraction rows. - Spot-check a few rows against the paper text/notes.

Definition of Done

• papers/extraction_table.csv exists.
• Every included paper from papers/screening_log.csv has exactly one extraction row.
• Column meanings match output/PROTOCOL.md (no ad-hoc columns without updating the protocol).

Troubleshooting

Issue: the protocol does not specify extraction fields

Fix:

• Update output/PROTOCOL.md (extraction schema section) and re-run extraction.

Issue: extraction table mixes narrative text with fields

Fix:

• Move narrative into a notes column and keep the rest as atomic values (numbers/enums/short strings).