Screening Manager (title/abstract screening)

Goal: produce an auditable screening log that can be traced back to the protocol.

Inputs

Required:

• output/PROTOCOL.md

Optional candidate pools (choose one if available):

• papers/papers_raw.jsonl
• papers/papers_dedup.jsonl
• papers/core_set.csv

Outputs

• papers/screening_log.csv

Output schema (recommended)

Columns (minimum viable):

• paper_id
• title
• year
• url
• decision (include|exclude)
• reason (short, protocol-grounded)
• reason_codes (protocol clause IDs, e.g., E3 or I2+I4)
• reviewer (HUMAN|CODEX)
• decided_at (ISO date/time)
• notes (optional)

Workflow

1. Read output/PROTOCOL.md and extract:

   • inclusion criteria
   • exclusion criteria
   • time window
   • any mandatory outcomes/metrics

2. Choose a candidate list

   • If a papers/*.jsonl or papers/core_set.csv exists, use it as the row source.
   • Otherwise, ask the user for the candidate list format and load it into the workspace first.

3. Screen each candidate deterministically

   • Decide include only when the title/abstract clearly satisfies inclusion and does not trigger exclusion.
   • If information is insufficient, decide exclude and state the missing requirement explicitly in reason.

4. Write papers/screening_log.csv

   • One row per candidate.
   • Reasons must cite protocol clause IDs from output/PROTOCOL.md (e.g., reason_codes=E3).
   • Avoid generic reasons like “not relevant”; state the missing requirement explicitly.

5. Quick QA

   • Check every row has decision + reason.
   • Spot-check that include rows satisfy the protocol.

Definition of Done

• papers/screening_log.csv exists and covers all candidates.
• Every row has a decision + protocol-grounded reason.

Troubleshooting

Issue: you do not have a candidate pool file

Fix:

• Import/export the candidate list into one of: papers/papers_raw.jsonl, papers/papers_dedup.jsonl, or papers/core_set.csv, then rerun screening.

Issue: too many borderline cases

Fix:

• Tighten output/PROTOCOL.md with more operational inclusion/exclusion rules (then re-screen for consistency).