Imported from SyedHasnain04/federated-clinical-intelligence (AGENTS.md). Install upstream with npx skills add SyedHasnain04/federated-clinical-intelligence. Copyright stays with the author.

AGENTS.md

Instructions for any AI coding agent (Claude Code, Cursor, Copilot, etc.) working in this repository. Read this before touching any file.

What this project is

A federated, multimodal, self-evolving clinical intelligence platform. 3-5 simulated hospital nodes train a shared model on MIMIC-IV COVID severity/mortality data without centralizing raw records, combining tabular EHR + clinical notes, with automated drift-triggered retraining.

Before doing anything

1. Read 00_SHARED_REPO_PROTOCOL.md — folder ownership, git workflow, error-rectification rule.
2. Identify which person's plan applies to the current task: PersonA_Implementation_Plan.md (federated infra/privacy), PersonB_Implementation_Plan.md (multimodal modeling), PersonC_Implementation_Plan.md (MLOps/dashboard).
3. If the task is Phase 7+ (research rigor, ablations, statistics), also read RESEARCH_PROTOCOL.md and EVAL_MODULE_SPEC.md.
4. Never edit files outside the current person's owned folders (see protocol) except eval/, which is co-owned and requires cross-review.

Standing rules for every task

• Follow the exact function signatures, config keys, and file paths specified in the relevant implementation plan — do not invent alternate names or restructure.
• Every code change must be followed by running the specified test command. Do not report a task complete without showing actual test output.
• On test failure: fix the implementation, not the test, unless the test is provably wrong. Re-run until green. Do not proceed on red.
• Never fabricate results, metrics, or claims of success without actual command output to back them.
• Never modify config/base.yaml without flagging it — this file is shared across all three people's modules.
• Never commit real patient data, API keys, or credentials.
• Replace print() with proper logging calls in any code intended to reach Phase 6 or later.
• Follow the exact git cycle in 00_SHARED_REPO_PROTOCOL.md: pull main → branch → build → test/fix loop → commit (only own files) → push → PR → resync if main moved → re-test → human review → merge.

Phase 7+ additional rules (once active)

Read the Master Contract and Hard Rules (Section 39) in RESEARCH_PROTOCOL.md in full before any Phase 7+ task. Key points: no cherry-picked seeds, no tuning until results "look right," no unsupported clinical-safety claims, every experiment logs config/seed/git-commit/dataset-version, all metrics come from eval/ — never inline one-off calculations.

When something doesn't fit

If an instruction here or in a plan document conflicts with what you're being asked to build, stop and flag it rather than silently reinterpreting or working around it. Contract mismatches between modules (e.g. a model's output shape doesn't match another module's expected input) must be raised explicitly, not patched with silent casts.

Reference index