Skip to content

SAFETY_AND_COMPLIANCE.md

Latest commit

 

History

History
91 lines (70 loc) · 3.85 KB

SAFETY_AND_COMPLIANCE.md

File metadata and controls

91 lines (70 loc) · 3.85 KB

Safety and compliance

Intended use

Research-first workflow-support software for human review of radiology report text and related pilot operations.

Explicitly not intended for

  • autonomous diagnosis
  • treatment recommendation
  • patient self-service interpretation
  • unsupervised clinical deployment
  • claims of validated clinical decision support

Safety controls

  • research-only framing in the API and web product
  • evidence-based explainability for each flag
  • human reviewer action required for escalation or closure decisions
  • durable review-action and import-run audit surfaces
  • configurable thresholds and deterministic rationale codes
  • no opaque model-only decision path replacing the explainable baseline

Privacy posture

  • use synthetic or de-identified data in development and public examples
  • do not keep PHI in the repository
  • store secrets in environment variables
  • default to local-safe development settings, with documented pilot auth modes for controlled environments

Pilot and auth posture

The current implementation supports:

  • mock auth for local development
  • trusted-proxy auth for controlled pilot packaging
  • header-based actor forwarding for controlled pilot packaging
  • site-scope-aware access and import behavior

These modes support research and demonstration workflows. They are not a substitute for institution-specific governance, enterprise IAM review, or production deployment controls.

Compliance notes

This repository does not claim regulatory clearance, clinical validation, or production readiness for general hospital deployment.

Contributors should avoid:

  • language implying approved clinical use
  • examples that contain PHI
  • documentation that presents scores or trial matches as decisions instead of review support

Responsible open source guidance

  • document dataset provenance clearly
  • keep examples and tests de-identified
  • disclose limitations and workflow assumptions
  • publish benchmark methods, not inflated performance claims

Autoresearch posture

Pancreatic Signal v2 ships an opt-in autoresearch lab subsystem (docs/AUTORESEARCH.md) inspired by karpathy/autoresearch. The following safety constraints are enforced by the schema, the experiment driver, and the API:

  • Single edit surface. The autoresearch agent is permitted to modify exactly one file: data/ontologies/pancreatic_signal_rules.json. Any proposal that touches another path is recorded as decision = "off_surface" and discarded.
  • No live writes. The autoresearch loop never writes to the case store, audit log, imports tables, or any production data path. It writes only to autoresearch/runs/<id>/ (append-only) and a temporary candidate ontology.
  • No Python source mutation in v2.0. Editing Python source is explicitly out of scope for the agent and deferred to v2.2.
  • Human-gated promotion. A "kept" candidate is not adopted by the live triage engine until a human runs make autoresearch-promote RUN=<id> (or invokes the admin-only POST /api/v1/autoresearch/promote/{id} endpoint). Promotion writes a provenance record at autoresearch/runs/<id>/promotion.json.
  • Rollback is one command. make autoresearch-rollback restores the frozen baseline at autoresearch/baseline/pancreatic_signal_rules.json into the live ontology path.
  • Schema-enforced guardrails. The candidate ontology is validated by OntologyConfig (extra="forbid") before any evaluation runs. Recall floors, threshold ordering, no-silent-deletion, and a deterministic eval check are enforced by scripts/run_autoresearch_experiment.py.
  • Pilot/clinical positioning unchanged. Autoresearch does not change the product's research-only positioning: triage remains explainable, every flag retains its rationale, and no autonomous clinical decision is introduced.