The statements a reader checks

Five minutes before a submission, somebody writes "Ethics approval was obtained" and "The authors declare no conflicts of interest". Neither sentence can be checked by anyone, and both are the kind a reader is right to be suspicious of. So they are a record here: plan/disclosure.json, validated when it is written and again by r0, printed as disclosure.md inside the submission folder.

The rule this skill exists for, in one line: this package can check that a statement is COMPLETE and it cannot check that it is true — and it will never decide whether a study needed approval. That is an institution's determination. An agent that answers it is answering for somebody else, and the person who pays for it is the researcher.

The record

rs_plan slug=completion-trial part=disclosure disclosure={
  "authors": [
    { "name": "A. Okonkwo", "credit": ["conceptualization", "methodology", "writing-original-draft"] },
    { "name": "R. Silva", "credit": ["formal-analysis", "visualization", "writing-review-editing"] }
  ],
  "ethics": { "approval": "approved", "body": "the University Research Ethics Committee",
              "reference": "UREC-2026-0142", "consent": "written" },
  "funding": [ { "funder": "the National Institute of Education", "grant": "NIE-2026-77",
                 "role": "The funder had no role in the design, analysis or the decision to submit." } ],
  "conflicts": "none",
  "data": { "availability": "on-request", "location": "the corresponding author, within the terms of the ethics approval" },
  "aiUse": { "used": true, "tools": ["a generative writing assistant"], "scope": "writing",
             "disclosure": "A generative writing assistant was used to edit the prose. All claims, analyses and references are the authors' own." }
}

plan/disclosure.json written, and plan/disclosure.md regenerated from it
- path: C:\Users\ana\Desktop\vncode-research\completion-trial\plan\disclosure.json
- 2 author(s), ethics approved, 1 funder(s), no conflicts, data on-request, AI use declared

`format=package` writes these statements as `disclosure.md` beside the title page, and NEVER into the manuscript: a disclosure names people and funders, and a blinded review must not receive it. This package checks that a statement is COMPLETE and cannot check that it is true - whether a study needed approval is your institution's determination, not this tool's.

Six fields, and only authors is required. Each one answers a question a reader actually asks.

field	what it must say, or it is refused
authors	every author needs at least one CRediT role — an author with no declared contribution is a name on a list
ethics	approval is one of approved, exempt-declared, not-required, not-applicable, pending; an approved or exempt-declared needs the body, the reference and how consent was taken
funding	a list of { funder, grant }. An empty list is a statement — "none to declare" — and is different from leaving the field out, which says nobody recorded it
conflicts	the string none, or a list of { author, what }
data	availability is open, on-request, restricted or none; the first two need a location, the last two need a reason
aiUse	used true or false. If true, tools and scope are required; if false, naming a tool or a scope is refused as a contradiction

What is refused, and why the wording matters

Refused (DISCLOSURE_INCOMPLETE): an ethics `approval` of approved needs `body`: "approved" with no approving body is a sentence nobody can check

• DISCLOSURE_INCOMPLETE — the statement is shaped correctly and says nothing a reader can check. An approval with no body, no reference or no consent; consent recorded as waived with no reason; data declared open with no address ("available on request" with nowhere to request is what a reader is entitled to read as "not available"); an author with no contribution.
• BAD_DISCLOSURE — the statement is malformed: an unknown field, a value outside its vocabulary, a role that is not CRediT, or an AI declaration that contradicts itself (used: false with a scope).

The difference matters when you explain it to a user: the first means finish the sentence, and the second means you have typed a field this package does not read.

Anonymity is the one thing the manuscript must not carry

The disclosure names authors, an approving body and a funder, so it is written beside the manuscript — as disclosure.md next to the title page — and never into it. The submission folder is built with no title block for the same reason, and a check asserts that not one author name from the record appears in the anonymized manuscript. Say this to the user when you build a package: the two files that identify the authors are the title page and the disclosure, and both are uploaded separately.

The question you must NOT answer

When a user asks "do I need ethics approval for this?", the answer is not a classification. It is:

This package will not decide that, and neither should I. Whether your study needs approval, needs consent, or is exempt is your institution's determination — the categories are usually exempt, expedited and full review, and which applies turns on things I cannot see: your jurisdiction, your institution's rules, your funder's terms and who the participants are. What I can do is record what they told you, and make sure the record names the body and the reference number so a reader can check it.

Then help with what IS answerable: what to collect, what the statement has to contain, what the venue asks for, and where a weak record would be visible. reference/ethics.md carries the detail — consent and assent, vulnerable populations, data protection, secondary use, animals, dual use, conflicts, ICMJE authorship and CRediT, funding, data and code availability, AI disclosure, and the integrity failures a statement exists to make visible.

Authorship, and what CRediT is for

The fourteen CRediT roles are the vocabulary: conceptualization, methodology, software, validation, formal analysis, investigation, resources, data curation, writing (original draft), writing (review and editing), visualization, supervision, project administration, funding acquisition.

• Every author needs at least one, and the matrix in disclosure.md shows who did what. A name with no role is the one entry a reader cannot interpret.
• ICMJE's four criteria decide who is an author at all (contribution, drafting or revising, final approval, accountability for the whole work) — a contributor who meets none of them is acknowledged, not listed, and one who meets all four cannot be dropped for convenience.
• Order is a convention, not a fact. It differs by field and by venue: say what the field does and check the venue's guidance rather than importing another field's rule.
• An AI system cannot be an author. It cannot take responsibility, cannot consent to be one, and every policy surveyed says so. Its use is disclosed, not credited.

Before the run goes out

Run rs_build slug=… format=package and read CHECKLIST.md: it says which statements the run HAS and which it does not, one line each, in those words. A line that says "Funding: not recorded" tells the author exactly what is missing, which is worth more than a line telling them to check something.

Then say the honest thing about what a submission is: the builder can check that the statements are there and complete, and it cannot check that any of them is true. That is what the author's signature means.