Auditable AI research: what makes research checkable

Oryx Centre for Strategic Futures is a two-person research initiative founded 2 October 2026. It is a Udyam-registered micro enterprise (UDYAM-UP-56-0189301) — NOT a registered organisation and NOT a registered non-profit. Every number below from Oryx’s own work is labelled as a pilot result, not a universal rate.

An AI system can produce a polished research report in minutes. The report looks finished. The question that matters is a different one: can someone else check it? If a reader cannot trace each claim back to a source they can open, read, and judge for themselves, the report is not auditable — however confident it sounds.

Auditable AI research is research built so that checking it is possible. This page explains what that means in practice: five properties a report must have, how to test for them in about fifteen minutes, and what to do when a report fails the test. You need nothing from Oryx to use any of this.

What “auditable” means

A report is auditable when a competent reader, given the report alone, can answer three questions about every material claim in it:

1. What is the source? The claim points at a specific, named source — a paper, a document, a dataset — not at a vague authority.

2. Can I open it? The source is retrievable: the link resolves, the DOI exists, the document can be found.

3. Does it actually support the claim? The source says what the report says it says — not something adjacent, not the opposite.

If any of those three fails, the claim is unverified, and the report’s conclusion rests partly on trust rather than evidence. Auditability is not a quality score. It is a property of how the work was done and recorded.

The five properties of auditable research

1. Every claim has a source. Each checkable statement in the report is tied to at least one source. A claim with no source is an assertion; it may be true, but nobody can tell from the report. This applies to numbers, quotations, and factual statements — not to the report’s own reasoning or recommendations, which should be labelled as the author’s judgement, not as sourced fact.

2. Every source resolves. A citation is not a source until it can be opened. A DOI that does not exist, a link that returns an error, or a document that cannot be located is a failed citation, not a minor flaw. In a published pilot audit of 66 AI-generated citations, 31 (47%) had broken links (Paper 2, study 1 — a pilot, not a universal rate). The lesson is not that AI always breaks links; it is that resolvability must be checked, because it fails often enough to matter.

3. Every source is judged on support, not just existence. A real source that says something else is a failed check, not a pass. For each claim, mark the verdict plainly: supports, partially supports, or does not support. In the same pilot audit, 7 of the 66 citations (11%) mismatched the claim they were attached to, and only 5 both resolved and fully supported the claim (Paper 2, study 1 — a pilot, not a universal rate). Existence and support are two different checks, and only the second one tells you whether the claim is true.

4. A named human reviewed it before use. A human review gate is a named person who checked the output before it was used, with the time and authority to stop it — not a disclaimer at the foot of the page. The gate’s job is the three checks above, claim by claim. If no one is named, the report is unreviewed, whatever the footer says. (See Oryx’s guide, “Human review gate AI: what a human must check” — oryxai.in.)

5. The check is recorded and re-runnable. The audit trail is kept with the report: which claims were checked, what each source said, what the verdict was, who checked it, and when. A reader should be able to re-run the check and reach the same verdicts. If the record is missing, the report asks you to trust the process without any way to verify it — which is the opposite of auditable.

What auditable research is not

• Not a badge. No label, seal, or “AI-verified” stamp makes a report auditable. Only the five properties above do, and a reader can test them without trusting the label.

• Not a plagiarism score. Originality and accuracy are different things. A report can be entirely original and still cite sources that do not support its claims.

• Not a disclaimer. “AI-generated, please verify” transfers the work to the reader without doing it. An auditable report does the checking first and shows its workings.

• Not perfection. An auditable report can still contain errors. The point is that errors can be found, named, and corrected — because the evidence trail exists.

How to test a report in 15 minutes

Take the report and a copy of this list. You do not need the author’s tools or access.

1. Pick five claims at random — not the ones the author highlights. If any of the five has no source, the report fails property 1.

2. Open every source for those five claims. Follow each DOI or link. A dead link or a source you cannot locate fails property 2.

3. Read the relevant passage of each source you opened. Ask: does it say what the report claims? Mark supports, partially supports, or does not support. Anything other than “supports” fails property 3 for that claim.

4. Look for the reviewer’s name. Who checked this report, and when? If no one is named, the report fails property 4.

5. Look for the check record. Is there a list of what was checked and what the verdicts were? If there is no record, the report fails property 5.

Five claims, fifteen minutes, five verdicts. A report that passes all five is auditable. A report that fails any of them is not — and the failure tells you exactly what is missing, so you can ask the author for it before you rely on the report.

When auditability matters most

Not every document needs this standard. A brainstorming note does not. Auditability matters when the report will be used to decide something: a purchase, a policy, a publication, a clinical or safety judgement, or any commitment that is hard to undo. The higher the cost of being wrong, the less acceptable it is to rely on a report nobody can check.

If you commission research, ask for auditability in the brief: sources per claim, resolvable citations, support verdicts, a named reviewer, and the check record delivered with the report. If a provider cannot offer those, that is useful information about what you are buying — and it should be priced accordingly.

Sources and further reading

• Oryx Working Paper v3 (Zenodo DOI 10.5281/zenodo.23718472) — the auditable-research method, including human review gates.

• Paper 2 v2, Trust, But Verify (Zenodo DOI 10.5281/zenodo.23731235) — the pilot citation audit: 66 AI-generated citations, four verdicts, all numbers labelled pilot.

• Oryx’s guide, “How to check AI-generated citations in your own paper” (oryxai.in) — the five-step check for individual citations.

• Oryx’s guide, “How to check an AI vendor’s claims before you sign” (oryxai.in) — the same claim-and-source discipline applied to vendor proposals.

---

Oryx AI offers claim-level evidence audits for research drafts and vendor claims — ₹7,500 per pilot report, payable only on acceptance, currently invite-only beta (contact: amartya@oryxai.in).

Oryx Centre for Strategic Futures is a two-person research initiative working on auditable AI.