How to verify an AI-written logframe before it reaches a donor
Oct 06 2026 · Templates
AI drafts a logframe in seconds and invents a baseline in the same breath. Here is the check that catches fake sources and false certainty.
To verify an AI-written logframe, work in three passes. First, trace every baseline and target to a document you can open; if none exists, replace the number with 'to collect'. Second, confirm every indicator names a real means of verification — a system, register or survey, not the word 'reports'. Third, scan for causal verbs such as 'caused' or 'drove' and downgrade them unless the design includes a comparison group. Then read the source material the model was given and ask what it left out, especially negative findings. Anything you cannot source is inference, and inference must be labelled as such before the logframe reaches a donor.
Why do AI-written logframes fail quietly?
A logframe is a grid of promises: indicators, baselines, targets and means of verification. A language model is fluent in the shape of that grid and careless with its contents. It will produce a plausible percentage, a plausible source and a confident causal verb in the same paragraph, and the prose will read perfectly well.
That is the problem. Errors in a logframe are not stylistic, so they do not look wrong. They sit in cells that a reviewer skims.
What does a four-minute logframe check actually look at?
Four passes, in this order. Do them before you edit anything else.
- Numbers — highlight every baseline and target. If it is not in a document you can open, delete it or write 'to collect'.
- Sources — every indicator needs a named means of verification: a system, a register, a survey. Never accept 'reports'.
- Causation — search for 'caused', 'drove' and 'led to'. Without a comparison group, change the verb to 'was followed by'.
- Omissions — ask what the source material contained that the draft left out, especially negative findings.
The order matters. Fixing the numbers first stops you polishing a target that should not exist.
Which parts of a logframe does AI get wrong most often?
| Logframe element | What AI tends to produce | What to check |
|---|---|---|
| Baseline | A specific figure with no source | Whether the figure exists in a document you can open |
| Target | A round, ambitious number | Whether the target follows from the baseline and the timeframe |
| Indicator | A restatement of the activity | Whether it measures change, not effort |
| Means of verification | A vague noun such as 'reports' | Whether a named system, register or survey exists |
| Causal claim | Confident 'caused' or 'drove' language | Whether the design includes a comparison group |
How do you verify a baseline or target you cannot source?
You do not verify it. You mark it. Write 'to collect' in the baseline cell and note who will collect it and when. A donor can work with a gap that is declared. A donor cannot work with a number that collapses under one question.
The same applies to targets. A target is a decision, not a finding, so it needs a stated basis — a previous round, a sector norm, or a funder's requirement. If the model supplied that basis, check it before you keep it.
How do you check means of verification?
A means of verification is a place you can go. Name the system, the register, the survey or the file. If the answer is 'the implementing partner's reports', that is a process, not a source, and it will not survive a review.
- Name the system or document, not the category.
- Say who holds it and how often it is updated.
- Check it covers the indicator's unit and period.
- If it does not exist yet, say when it will.
How do you stop an AI logframe claiming causation?
Search the draft for causal verbs and rewrite them. 'The training caused attendance to rise' becomes 'attendance rose over the same period'. Without a comparison group, a logframe can describe change but cannot attribute it.
If the design does include a comparison group, say so explicitly in the assumptions or the methodology note, so the claim is defensible rather than accidental.
What should you do about what the AI left out?
Ask the model what it excluded. Then read the source material yourself and look for the findings that complicate the story — the group that did not improve, the indicator that moved the wrong way, the assumption that failed.
A logframe that contains only good news is a logframe a reviewer will distrust.
Can you use AI to check its own logframe?
Yes, with rules. Give the model a short instruction file that requires it to tag every sentence as [EVIDENCE] or [INFERENCE], and to list the source behind each evidence tag. Then audit the inference lines yourself — that is where the invented numbers live. If you are still choosing a model for this work, /which-ai walks through the trade-offs.
This keeps the speed and removes the false certainty. It also leaves you an audit trail you can show a donor.
What do you hand the donor when they ask how you know?
A ledger: claim, source, type. One line per indicator, saying where the number came from and whether it is measured or assumed. It takes minutes to build from a verified logframe and it answers the question before it is asked.
If you want the grid, the prompts and the ledger format ready to use, start with the free template at /blog/logframe-template-google-sheets. When you need to build a full set, /store/builder covers the logframe and /store/indicator covers the indicator reference sheets that sit behind it.
Written for M&E professionals. Not a guarantee of accuracy — always verify every claim against its source, and let a named human sign off. Get the free handbook.