Your name is on the report. Make every claim defendable.
You already use ChatGPT or Claude for transcripts, coding and drafts. Evaluation AI adds the part that's missing: a 20-minute check that catches invented quotes, wrong numbers and quiet omissions — plus an audit trail you can hand to the donor. It runs inside your own account.
No code · Works with ChatGPT, Claude and Gemini · Built from 208 M&E ads, 264 Amazon reviews and 1,000+ practitioner threads
The checking step is the part that doesn't scale.
Your AI drafts in minutes. Verifying every claim still takes days — and that is where the risk lives.
It quoted something that wasn't there
Confident summaries, invented citations. You found one. How many did you not find?
Manual re-reading doesn't scale
300 transcripts, four reports, one you. The check that worked on one project breaks on three.
They can smell AI prose
Unlabelled inference and “not X but Y” phrasing undermine the evidence — even when the evidence is sound.
Client data can't touch the cloud
Transcripts and beneficiary data are confidential. The check has to run in your account, not ours.
Four files. One loop. Every claim labelled.
The harness is four plain-text files that make your AI check its own work — no code, no new platform, no data leaving your account.
Load the rules
A single instruction file turns your ChatGPT / Claude / Gemini into a strict verifier: label every claim, never invent, ask when unsure.
Run the labels
Every sentence becomes [EVIDENCE] (traceable to a source) or [INFERENCE] (the model's reasoning) — with a source and confidence level.
Hunt the failures
The checklist sweeps numbers, citations, causal language, consensus claims and the quiet omission — the most dangerous failure of all.
Log and sign
The provenance ledger records claim → source → confidence. You sign; the client-facing verification statement goes with the report.
Steal the three prompts right now
This is the free version of the checking step. Paste your draft and sources into your own AI.
1. "List every claim in the text above as [EVIDENCE] or [INFERENCE] with a source." 2. "What does the source contain that this draft leaves out?" 3. "Where does this draft assert causation, consensus or magnitude the source does not support?"
These three catch most failures. The pack adds the full Skill, six golden sets with traps, per-artifact checklists and the ledger template — the difference between a prompt and a system.
The page that wins the next contract.
Anyone can say “AI-assisted”. You hand the donor a statement with numbers on it.
VERIFICATION STATEMENT Harness version: v1.2 · Skill: Evaluation Narrative + Citation Audit Golden Set: 10 items · pass rate 9/10 Claims extracted: 214 Traceable to source: 199 Labelled inference: 15 Omission audit: passed Causal claims reviewed: 6 (2 rewritten to "contributed to") Verification fee: included / $1,500 Reviewer: [your name] · Date: [date] Known limitations: two sources undated; one indicator series interpolated.
The pack ships this as a fill-in template. The paid book shows how to price it: 10–20% verification premium, or a flat $500–2,000 per deliverable.
Start with the Pack. Add the agents when you need them.
No subscription required for the check itself. Membership adds the builder, the practice and the monthly benchmark.
The system in four files.
- Skill files for ChatGPT, Claude and Gemini
- Six golden sets with traps
- Per-artifact checklists
- Provenance ledger + verification statement
- Evaluation Plan workbook
For consultants shipping client work every month.
- Everything in the Pack
- Harness Builder + workflows
- Donor Reviewer Simulator: 10 min/mo
- Monthly model benchmark
- Commercial licence
Per deliverable.
- We run the standard harness on your report
- Claim-by-claim ledger returned
- Signed verification statement
- 48-hour turnaround
- Sprint from $4,500 for a full build
Rehearse the hostile question.
Before the donor asks “how do you know?”, practise answering. Live sessions run on prepaid minutes.
Donor Reviewer Simulator
A skeptical donor reviewer reads your findings and hunts for unsupported claims, missing evidence, causal leaps and quiet omissions. You defend. You get a structured score — and a report that survives the real review.
Run a donor review →Monthly model benchmark
We run M&E verification golden sets against every major model and publish the pass rates. You know which AI is reliable this month — and when to switch.
See the benchmark →Pro includes 10 agent minutes every month. Every session is clearly labelled as AI.
Client data never has to leave your account.
Your account, your keys
Run everything in your own ChatGPT, Claude or Gemini account, BYOK, or a local/self-hosted setup. We never need to see the documents.
AI is always labelled
The same disclosure discipline we recommend for your reports. If AI touched it, the reader knows — including our own video and agents.
Humans sign off
The agent never issues a verification claim. The ledger and the signature belong to you — that is precisely what the donor is buying.
The honest answers.
I already prompt carefully. Why do I need this?
Careful prompting reduces errors; it does not catch them. The checking step — re-reading 300 pages to verify every claim, number and citation — is what doesn't scale. The checklist and ledger turn it into a 20-minute routine.
Can I use my client's confidential data?
Yes. Run the check inside your own AI account, BYOK, or a local/self-hosted environment. Nothing has to leave your machine or your account, which is also what your NDA wants to hear.
Will clients pay more for verification?
That is the pitch in our paid book: add 10–20% as a verification premium, or charge a flat $500–2,000 per deliverable for the check and the statement. The statement is the reason you get re-hired.
How long does it take to run?
About 20 minutes for a typical report once the Skill and checklist are loaded. The first run takes longer while you adapt it to your client's format.
Does this work if my donor has a custom format?
The check is format-independent — it verifies claims, not layouts. Format-specific checklists cover FCDO, EU, UN and World Bank; custom donor templates can be adapted once and reused.
Isn't this just ChatGPT?
ChatGPT writes; this checks. The output is a labelled ledger and a statement with numbers on it — the page a donor reads when they ask “how do you know?” We sell the verification layer, not the generation.
Sign nothing you can't defend.
Read the free edition to see the method, or get the Pack and run your next report through the check this week.
Free. The Bridge, one issue a week — AI translated into evaluation work. Unsubscribe anytime.