PICO exact-span extraction run report — Hard V2

External/network lookup used: NO

Approach:
Read the supplied task instructions, span guide, workbook, and all 20 exact abstract strings from pico_cases.jsonl. Annotated only source-supported in-vivo animal experiment spans for Species, Strain, Induction, Intervention, Comparator, and Outcome; generally excluded title/background/objective/conclusion-only material. Used shortest meaningful non-overlapping phrases, retained repeated eligible mentions at distinct offsets, and sorted all spans by start then end.

Local scripts/automation:
Used local Python scripts only to segment supplied abstracts for review, locate exact substrings, calculate 0-based end-exclusive offsets, serialize each span cell as a JSON list, and validate abstract_text[start:end] == text for every span. A second validator checked sorting and global non-overlap across labels before the workbook was written. The provided .xlsx template was edited with artifact_tool and exported locally.

LOW confidence documents:
4 (PICO-V2-001, PICO-V2-006, PICO-V2-009, PICO-V2-018)

Hardest cases:
1. PICO-V2-001
2. PICO-V2-003
3. PICO-V2-006
4. PICO-V2-007
5. PICO-V2-008
6. PICO-V2-009
7. PICO-V2-013
8. PICO-V2-016
9. PICO-V2-018
10. PICO-V2-019

Major ambiguity/error modes:
Genotype-like phrases sometimes functioned as experimental interventions rather than strains; multi-arm designs could make the same conceptual exposure an induction in one mention and a comparator in another; several abstracts used group numbers or indirect control wording; ex-vivo measurements had to be separated from ex-vivo-only interventions; abbreviated disease challenges and unusual punctuation complicated exact boundaries; and some result clauses mixed outcome names with directionality or mechanistic interpretation.

Failures/limitations:
No external metadata, source articles, corpus annotations, or network lookup were used. Some abstracts did not explicitly state control treatments or induction procedures, so those roles were left empty rather than inferred. Ambiguous role assignments were resolved from the supplied abstract text only, and LOW confidence was used where multiple defensible PICO role boundaries remained.