dsh-irb

Assemble the ethics application before the committee does — and know which core statements are still missing.

An 18-item keyword checklist for ethics applications, an advisory risk level with the safeguards it implies, and a Markdown + SARIF export you can attach to the submission.

English · 简体中文

---

What it does

• irb_check — 18 items: title, PI/contact, purpose, design, participants/eligibility, recruitment, informed consent, assent for minors, risks, benefits, confidentiality/data protection, vulnerable groups, compensation, conflicts, retention, dissemination, animal ethics (3Rs), biosafety/human samples. Present items get character locators; missing core items get advice.
• irb_risk — structured flags (vulnerable groups, minors, prisoners, pregnancy, cognitive impairment, deception, sensitive topics, invasive procedures, biological samples, genetic data, human challenge, gene therapy, automated decisions) → advisory level plus the safeguards each factor requires.
• irb_export — Markdown (item table, risk table, safeguards) + SARIF 2.1.0, plus REP evidence when the application is checked.

What it does not do

• No approval, no legal advice. A checklist cannot substitute for the IRB/IACUC decision; the export says so.
• No adequacy judgement. It detects whether a statement exists, not whether it is sufficient.
• No risk classification standard. The level is advisory and must be confirmed by the committee.
• No DOCX/PDF. Convert to .md/.txt (PDF text: dsh-doc-anchor).
• No submission portal integration.

Risk factors and safeguards

Flag	Safeguard advised
vulnerable groups	extra consent/assent procedures and independent oversight
minors	age-appropriate assent plus parental consent
prisoners	independent advocate and protections against coercion
pregnant women/fetuses	specific fetal and maternal risk assessment
cognitive impairment	capacity assessment and authorised-representative consent
deception	debriefing and a justified waiver of full disclosure
sensitive topics	referral pathways and interviewer training
invasive procedure	clinical oversight, adverse-event reporting, stopping rules
biological samples	sample handling, biosafety, secondary-use consent
genetic data	specific consent and incidental-finding policy
human challenge	safety monitoring board and dose-escalation rules
gene therapy	long-term follow-up and regulatory approval
automated decisions	human review and an appeal route

Tools

irb_check

{ text? | path?, persist? } → { ok, analysis, evidenceId?, recordPath? }.

irb_risk

{ vulnerableGroups?, minors?, prisoners?, pregnantWomen?, cognitiveImpairment?, deception?, sensitiveTopics?, invasiveProcedure?, biologicalSamples?, geneticData?, humanChallenge?, geneTherapy?, aiDecisionMaking? } → { ok, risk }.

irb_export

{ analysis, risk?, title? } → { ok, markdown, sarif }.

Error codes

INVALID_INPUT · NOT_FOUND · PATH_ESCAPE · NOT_A_FILE · TOO_LARGE · UNSUPPORTED_FILE · STORE_ERROR

Compatibility & install

DeepSeek Harness	dsh-v0.2.1-rc.5 (verified 2026-11-08); peer >=0.2.1-0 <0.3.0 || >=0.2.2-0 <0.3.0
Node	^22.19.0 || >=24.9.0 (tested 24.22.1), ESM only
Peers	@perrylink/dsh-research-core ^0.1.0, @deepseek-ai/cordis ^4.0.5-alpha.1, @deepseek-ai/dsh-tools, @deepseek-ai/schemastery ^3.18.5-alpha.1
Runtime dependencies	zero

dsh plugin --profile <name> add @perrylink/dsh-research-core dsh-irb
dsh --profile <name> --dump-config | grep dsh-irb

Configuration

Key	Default	Meaning
enabled	true	false mounts nothing
root	.irb	store/record root
workspaceRoot	process.cwd()	application paths must resolve inside it
maxChars	1000000	character cap (fails, never truncates)
persist	true	store the checked application as REP evidence + record

Security boundaries

• Read-only on the application; writes only under root.
• realpath containment; DOCX/PDF rejected with a conversion hint.
• With persist: true (the default) the checked application text is stored as REP evidence under root; it can contain personal data, so set persist: false or protect the store accordingly.
• The application text is data; instruction-like content is never executed.

Evaluation

13 tests: checklist present/missing items and core-missing list, generic consent/ethics-committee mentions not mistaken for assent/animal content, empty-text rejection, minimal risk, minors+deception safeguards, all thirteen risk flags, runtime (REP record validity, optional-missing checks recorded as skipped, traversal/binary/oversize/empty rejection), malformed analysis/risk tool guards, tool schemas, check→risk→export end to end, failure branches, package contract.

No benchmark is claimed: keyword presence is pinned by fixtures; there is no labelled corpus of ethics applications.

Known limitations (v0.1)

• English/Chinese keyword cues only; a differently worded statement can be missed.
• Risk flags come from the caller, not from the application text.
• No jurisdiction-specific rules (HIPAA, GDPR, 涉及人的生命科学和医学研究伦理审查办法 details).
• No animal-protocol section structure beyond the 3Rs keyword check.

License

Apache-2.0. No third-party runtime code.