# Task 3 Hard V2 — preclinical PICO exact-span extraction

This is a 20-document closed-book confirmation run.

For each abstract, extract every source-supported exact span belonging to:
- Species
- Strain
- Induction
- Intervention
- Comparator
- Outcome

## Source annotation conventions
1. Focus on the actual in-vivo animal experiments.
2. Do not annotate interventions performed only in isolated tissue/cells in vitro or ex vivo.
3. Background/objective/conclusion text is generally excluded unless needed to interpret
   an experimental description or abbreviation.
4. Preserve the shortest meaningful source phrase; exclude unnecessary punctuation.
5. Spans must not overlap.
6. Repeated eligible mentions at different offsets are separate entities.
7. Experimental genetic manipulations can be role-bearing interventions/inductions when the
   abstract uses them as the experimental manipulation; do not automatically classify every
   genotype-like phrase as Strain.

## Files
- `pico_span_extraction.xlsx`
- `pico_cases.jsonl` — exact source strings for character indexing

## Required span format
Every one of the six span cells must contain a JSON list.

Example:
`[{"start":392,"end":396,"text":"mice"}]`

No entity:
`[]`

Offsets are 0-based and end-exclusive:
`abstract_text[start:end] == text`

Sort spans by start, then end.

Also fill:
- confidence = HIGH | MEDIUM | LOW
- extraction_note = one concise sentence describing the main role/boundary issue

## Closed-book rule
Use only the supplied files. Local scripts for exact offsets are allowed.
Do not use web search, PMC/source lookup, the original corpus, or external annotations.

## Return
1. `pico_span_extraction_output.xlsx`
2. `run_report.txt`

The report should include:
- approach;
- local scripts/automation;
- LOW-confidence count;
- 8–15 hardest case IDs;
- major ambiguity/error modes;
- failures/limitations;
- whether external/network lookup was used.