AI pilot evaluation checklist: agree success first

In short: most AI pilots do not fail on the technology. They fail on the argument afterwards, because success was never written down. Agree the decision, the written pass criteria, the evidence check, the human gate, and the exit before any work starts — then judging the pilot is a checking exercise, not a negotiation.

Oryx Centre for Strategic Futures is a two-person research initiative founded 2 October 2026. Pilot engagements are contracted through Riya Nalwa’s Udyam-registered proprietorship (UDYAM-UP-56-0189301). Every number below from Oryx’s own work is labelled as a pilot result, not a universal rate.

Most AI pilots do not fail on the technology. They fail on the argument afterwards: the vendor says the pilot worked, the buyer says it did not, and nobody can settle it because success was never written down. A pilot evaluation checklist exists to prevent that argument. It is agreed before any work starts, it names who judges the result, and it says what happens when the result falls short.

This page teaches the checklist itself. You need nothing from Oryx to use it.

Start with the decision, not the demo

A demo shows what a system can do under ideal conditions. A pilot should answer a narrower question: does this system help with one real decision, on real inputs, to a standard we set in advance? If you cannot name the decision the pilot informs, you do not have a pilot yet — you have a demo with a longer timeline.

Write the decision down first, in one sentence: “We need to know whether X can produce Y to standard Z, so that we can decide W.” Everything in the checklist below hangs off that sentence.

The checklist

A. The question and the inputs

• The pilot question is written in one sentence, with the decision it informs named.
• The inputs are real: the pilot runs on the buyer’s own material, not the vendor’s showcase examples.
• The scope is bounded: one question, one deliverable, one deadline — not “explore what the tool can do.”
• What is out of scope is written down too, so scope creep is visible when it happens.

What good looks like: anyone reading the pilot brief cold could say what is being tested and what decision depends on it.

B. Success criteria, agreed in writing before work starts

This is the section that decides whether the pilot ends in a verdict or an argument.

• Pass criteria are written, specific, and measurable — “accurate” is not a criterion; “every material claim traced to a retrievable source” is.
• Criteria are agreed and signed by both sides before work starts, not drafted after results arrive.
• Each criterion names how it will be checked: who checks, against what evidence, by when.
• Partial success is defined: which gaps mean “fix and re-check” and which mean “fail.”
• A named human on the buyer’s side accepts or rejects the result. “The team will review it” is not a name.

What good looks like: on delivery day, judging the pilot is a checking exercise, not a negotiation.

C. Evidence and traceability

A pilot report that cannot show its workings cannot be evaluated — it can only be believed or disbelieved.

• Every material claim in the pilot output is tied to a specific source the buyer can open.
• Sources resolve: links, DOIs, or documents are real and retrievable, not plausible-looking references.
• Sources support the claim: a real source saying something else is a failed check, not a pass. Mark each claim supports / partially supports / does not support.
• Uncertainty is stated, not smoothed over: where evidence is thin or conflicting, the report says so.
• The check record is kept with the pilot file, so the verdict can be re-examined later.

Why this section matters is visible in published pilot evidence, labelled as pilot evidence, not a universal rate: in Oryx’s published pilot audit of 66 AI-generated citations (Paper 2, study 1), 31 citations (47%) had broken links, 7 (11%) mismatched the claim they were attached to, and 11 (17%) could not be verified — only 5 both resolved and fully supported the claim. The point is not that every AI output fails at that rate. It is that “looks right” is a weak test, and a pilot that skips the evidence check has skipped the part most likely to fail.

D. Human review gate

• A human review step is built into the pilot, not added as an afterthought: someone accountable reads the output before it is judged.
• The reviewer is named, and is not the same party that produced the output.
• The reviewer’s verdict — pass, conditional, or fail — is recorded in writing with reasons.

E. Failure, remedy, and exit

• What happens on a fail is written before work starts: is there a revision round, a re-check, or does the engagement simply end?
• Payment terms follow the verdict, not the effort: the contract says what is owed on pass, on conditional, and on fail.
• Exit is clean: the buyer keeps the pilot output and the check record whatever the verdict, and there is no obligation to continue into a larger contract.
• Data handling on exit is stated: what the vendor retains, what is returned or deleted, and by when.

F. Cost and what happens after

• The full pilot cost is stated upfront, including any cost of a re-check or revision round.
• Continuation is a separate decision with separate terms — the pilot agreement does not silently roll into a subscription or retainer.
• A date is set for the go / no-go decision after delivery, so a pilot cannot drift indefinitely.

How to score the pilot

Score each section, not just the pilot as a whole:

• Pass: every item in the section is met, with evidence on file.
• Conditional: named gaps, each with a fix and a re-check date. A conditional is not a pass, and it does not become one by waiting.
• Fail: any section fails on a criterion agreed in advance, or the evidence check cannot be completed at all.

Record the section scores and the overall verdict in one page, attach it to the pilot file, and keep it with the contract. Six months later, that page is the difference between “we tried that and it didn’t work” and “we tried that, it failed on evidence traceability, and here is the record.”

Red flags

• Success criteria proposed after results are already in.
• “Accuracy” or “quality” stated without a checkable definition.
• No named human acceptor on the buyer’s side.
• Sources that cannot be opened, or claims that cannot be traced to a source.
• Payment owed on delivery regardless of the verdict.
• A pilot agreement that rolls into a larger contract unless the buyer opts out.
• No written remedy for a failed pilot — the only options are “accept it” or “argue about it.”

A worked example: an acceptance-gated pilot

One way to run a pilot so that the checklist enforces itself is to gate payment on acceptance. Oryx’s own pilot process works this way, and it is described here as an illustration of the method, not as a claim about any other provider:

1. The buyer sends the research question and the decision it informs.
2. Pass criteria are written and agreed in writing before any work starts.
3. The work runs, and a human review gate checks the report before delivery.
4. The buyer checks the report against the agreed criteria. Pass means an invoice is issued; fail means nothing is charged.

The mechanism matters more than the provider: because the criteria are fixed before the work, and payment follows the verdict rather than the effort, the argument the checklist exists to prevent has nowhere to start.

Questions to send before a pilot starts (copy-paste list)

1. What decision will this pilot inform, in one sentence?
2. What are the written pass criteria, and who signs them before work starts?
3. How will each criterion be checked, and by whom?
4. How is every material claim in the output traced to a source I can open?
5. Who is the named human reviewer, and is that person independent of the producer?
6. What happens on a fail — revision, re-check, or clean exit — and what is owed in each case?
7. What do I keep whatever the verdict, and what happens to my data on exit?
8. What does continuation cost, and is it a separate decision from accepting the pilot?

Sources and further reading

• Oryx Working Paper v3, From Answers to Auditable Intelligence (Zenodo DOI 10.5281/zenodo.23718472) — the six-layer framework behind the evidence and human-gate checks above.
• Paper 2 v2, Trust, But Verify (Zenodo DOI 10.5281/zenodo.23731235) — the published pilot citation audit quoted in section C.

---

Offer: 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@oryxcentre.dpdns.org).

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