# Pentabrid 27B — Reproducibility Package

Artifacts supporting the Nature Medicine Matters Arising response to Oermann & Vishwanath (2026), "General-purpose LLMs outperform specialized clinical AI tools on medical benchmarks" (s41591-026-04431-5).

## Core finding
At fixed scale and offline (no retrieval), the specialization procedure determines medical reasoning performance. A short-CoT fine-tune (V13) collapsed the base model's reasoning (about 3,100 to about 760 tokens) and dropped MedXpertQA accuracy by 16.9 points; a corrected long-CoT approach (V14) restored reasoning length and recovered accuracy to near-parity. Framing: recovery to near-parity (41.7 vs 43.8), reversing the regression, not superiority over the base model. Fine-tune A corresponds to V13 and Fine-tune B to V14 in the deposited results table.

## Results (MedXpertQA, full 2,450 questions; scorer-verified)
See results_summary.csv. Base 43.76, V13 26.90, V14 41.67, V15 41.35, V16 (alpha=64) 42.20. V15 and V16 are two independent null ablations confirming the recovery is robust to configuration.

## Contents
- scripts/merge_medxpertqa.py — the scorer (exact match on 'Answer: X'; audit this for the no-inflation claim)
- scripts/eval_generic_medxpertqa.slurm, eval_mcq_hf.* — evaluation harnesses
- scripts/train_v14.py + .sbatch, merge_lora_v14.py — training and adapter merge
- scripts/probe_*.py — reasoning-structure probes
- results_summary.csv — the locked results table

## Base model and benchmarks (not redistributed here)
Base: Qwen3.6-27B (official Qwen repository). Benchmarks: MedXpertQA, MedQA, MedMCQA, used under their own licenses, not included here.

## Authors
Prof. Adnan Agha (ORCID 0000-0002-2704-8931) and Eram Anwar (ORCID 0009-0006-9335-9208), UAEU College of Medicine and Health Sciences / Tawam Hospital. IP: UAEU Application #2442.

## License
CC-BY-NC-ND-4.0