Skip to content
neuroflow0.2.31 · docs

← neuroflow

/neuroflow:ethics

Ethics and IRB workflow — the hard gate before human data collection.

Tracks protocol submissions and amendments, versions consent forms, and keeps the approval status — including the expiry date — visible across the whole project.

Modes

Mode What it does
(none) / --status Approval dashboard — status, dates, consent version, amendments, AI processing
--protocol Draft or update the ethics protocol; log amendments (append-only)
--consent Create or revise a consent form — always as a new version, never edited in place
--approved Record an approval: committee, reference, approval date, expiry date, and whether participant data may be read by the AI model
--erase A participant asked for erasure: record it, run the sweep, work through the checklist

Why it matters

  • Consent forms are versioned because participants signed a specific version — it must stay retrievable exactly as signed.
  • Expiry dates go to timeline.md, so /phase shows an approaching expiry without you running /ethics.
  • The gate is enforced in the workflow. status.md starts with a small machine-readable block (status, approval_id, expires, ai_processing, set_by, set_at). /data and /experiment read it and stop every step that involves participants while the approval is missing, expired, or recorded without your confirmation (set_by: model). Projects without participants note ethics: not-applicable.
  • The gate is checked: if data exists but no approval is recorded (or data predates it), /ethics --status and the sentinel audit both warn.
  • Reading participant data with an AI model is a data transfer. Everything the model reads goes to the model provider and into Claude Code's local transcripts. ai_processing (none, pseudonymised, identifiable) records what your approval, consent form and data-protection officer allow under applicable law (e.g. GDPR or HIPAA); with none, Claude writes the scripts, you run them, and Claude works from the aggregate output.
  • Erasure requests are a checklist, not a single delete. --erase lists every place a participant's data can hide — derivatives, project memory, git history, Claude Code's own transcripts and caches, personal and team caches, external services — and erasure_sweep.py finds every file that mentions the ID. It reports paths and counts only and never deletes; you run it in your own terminal and decide.
  • Participants only ever see validated software — stimuli, questionnaires and consent never run inside Claude Code.
  • /output --archive reads the consent version in force before letting a dataset go public.

Everything lives in .neuroflow/ethics/ with status.md as the single source of truth.

  • /preregistration — study design registration (scientific commitment; ethics is the legal/administrative one)
  • /experiment — participant sessions, pilots and allocation happen only behind this gate
  • /data — data intake happens only behind this gate