From protocol stages to researcher decisions

The ten-step protocol organises responsibilities, not a mandatory sequence of forms. Reason provenance is cross-cutting: preserve a proportionate contemporaneous trace when a consequential proposal, specification, challenge or reversal occurs. Stage 10 assesses the limits of later attribution; it is not the first moment to collect provenance.

Pre-use boundary across all stages: verify whether source materials may be disclosed to the selected AI service. Research ethics, consent, institution and collaborator rules, confidentiality, contractual restrictions and service arrangements may prohibit transfer irrespective of whether material would later be published.

Stage	Researcher-controlled decision	Evidence or limit	Possible record
1. Substantive problem	Which question matters independently of AI?	Disciplinary rationale; a generated topic does not establish importance.	Research Charter
2. Proposals	Which candidate deserves further investigation?	Proposal is not evidence; preserve source and selection when consequential.	Decision Log
3. Specification	Which units, categories, exclusions and kind of conclusion are intended?	Preserved design choices and revisions; record whose constraints changed the course.	Research Charter
4. Verification	Which outputs, documents or interpretations actually passed scrutiny?	Design-appropriate execution, source criticism or interpretive testing; no form replaces them.	Evidence Ledger
5. Challenge	Which plausible alternative could weaken the claim?	Comparison, contrary case or critique able to change the conclusion; retain consequential challenge history.	Claim Challenge Record
6. Inference revision	What is retained, narrowed, reframed or withdrawn?	Separate sound evidence from unsupported inference; retain the decision and its trigger.	Inference Revision Record
7. Claim authorisation	What exact wording does the checked research permit?	Explicit acceptable and prohibited formulations with necessary conditions.	Claim-Preservation Contract
8. Manuscript revision	Does an edit preserve evidence, qualifications and warranted novelty?	Distinguish evidence-led narrowing from stylistic convergence or loss of salience.	Post-Revision Sufficiency Audit
9. Final inspection	Does the exported object actually convey the supported claim?	Check figure meaning, citations, caveats, rights and document metadata.	Final-object case
10. Bounded attribution	What later claim about contribution does the already retained trace warrant?	Distinguish contemporaneous records from ex-ante rules, frozen artefacts and recollection.	Reason-Provenance Record

Three separate judgements

Consider an AI-proposed comparison that enters a paper. Verification asks whether the comparison was performed and checked; claim authorisation asks which inference the result licenses; reason provenance asks what trace warrants saying who proposed, selected or rejected it. Validity and attribution cannot substitute for each other.

Stage 3’s inferential target simply means the kind of conclusion being sought: a description, an association, a causal effect, a theoretical interpretation or something else. Each requires the appropriate basis of support.

Qualitative research has analogous but not identical gates. An interview theme proposed by AI remains provisional until analytic memos, extracts, contextual material and discordant readings have been considered in ways consistent with the study’s epistemology. Historical source criticism may instead require checking provenance, dating, omissions and the author’s position. Neither case can be settled by code execution or a generic reliability threshold.

Return upstream when necessary

Warning sign	Return to	Decision rule
Generated programme never executed or supposed source never read	Evidence production	Remove the claim until independent verification occurs.
Cross-sectional association rewritten as a causal effect	Interpretation and claim authorisation	Do not supply identification through rhetorical polish.
An edit omits an essential caveat	Manuscript governance	Restore it; reopen the claim boundary only if new evidence or theory warrants a change.
Repeated simulated reviews flatten an established contribution	Researcher judgement and risks	Defend warranted novelty; stop style-only rounds without a specific validity gain.
Origination asserted from a final manuscript or later memory	Attribution assessment	Report the supported narrower history and what remains unknown.

Use the crosswalk to navigate, not as a compliance measure or claim of methodological superiority.