Clinical Skill Architecture: a framework for translating rating-scale ontologies into LLM agents

Abstract

Changelog All notable changes to clinical-skill-architecture will be documented in this file. The format follows Keep a Changelog and this project adheres to Semantic Versioning. [0.2.0] — 2026-09-21 Revision release accompanying the revised npj Digital Medicine manuscript. No change to the skill files, prompts or pipeline logic (prompt version stays 1.0.0); the release corrects the documentation and adds the suite-independence tooling. Fixed docs/reproduce_paper.md, README.md, MODEL_CARD.md: the synthetic figures F1 0.840 / sensitivity 0.971 were attributed to Claude Sonnet 4.6; they were produced by Claude Opus 4.6 (Sonnet 4.6: F1 0.824). Removed figures that correspond to no result file (decision accuracy 0.94, weighted κ 0.92, Cohen's κ 0.68, Spearman ρ ≈ 0.41, an "Opus 4.7" run). Table numbers now match the revised manuscript's table titles rather than an earlier draft. docs/reproduce_paper.md: GLM-5 through openai_compat documented as the paper's primary configuration and primary reproduction route; the Claude runs are documented as CLI runs. extras/paper_reproduction/README.md: the scripts are adapted copies, not verbatim copies. README.md Step 5: the human-agreement guidance names the statistic and level of analysis (screening κ, subscale-presence κ, weighted-sum ICC) instead of an unqualified "κ ≥ 0.6". DISCLAIMER.md: aligned with the manuscript's classification (research software; not clinical decision support; not a medical device) and its seven preconditions for clinical use. config/output_format.md example: HOP coding weight 3 → 1 (HOP is a flat-weight subscale; no output was affected). Suite v1.0 figures for GLM-5 restated from the current evaluator (F1 0.629, precision 0.650, one false-positive coding fewer than the artefact archived in March 2026); Claude Opus 4.6 is unchanged. Added Concept DOI (10.5281/zenodo.20435438) recorded in CITATION.cff and the README software citation; the software citation's version was still 0.1.0 and is now 0.2.0. README: the "Adapting the framework to your own scale" recipe now warns that lifting examples into skill files and extrapolating from seed examples are where test-set independence is lost, and points at the gate; the citation block carries the revised article title. synthetic/independence_gate.py: skill-file independence gate (sequence ratio ≥ 0.7 or shared five-token phrase) and synthetic/suite_v11.json, the 90-instance de-overlapped suite v1.1; generate_all.py --strict refuses instances that fail the gate. csa.eval.evaluate_run(..., instance_ids=...) to evaluate on a subset such as v1.1. Expert-arm and construct-validity figures, with their reference standards and the calibration limit (weighted-sum ICC 0.21), in MODEL_CARD.md and README.md. [0.1.0] — 2026-04-23 Initial public release accompanying the npj Digital Medicine submission. Added score() and score_async() module-level functions and GGAgent class Anthropic backend with prompt caching enabled by default OpenAI-compatible backend (Requesty.ai, OpenRouter, vLLM, Ollama) 7 subscale skills (HOP, SAC, PMR, SOM, DAM, SEP, HOS) loaded via importlib.resources Two-stage screening + per-subscale detail pipeline csa-score console script csa.eval evaluation harness with hierarchical accuracy, Clopper-Pearson CIs, Cohen's and weighted kappa 3 demo synthetic instances bundled inside the wheel Full 150-instance synthetic suite generator at synthetic/ Copy of the paper pipeline scripts at extras/paper_reproduction/ 34-test suite covering fragmenter, scorer, validator, eval metrics, agent smoke