Method and limits
Every Probative report is built from dated captures. A capture records what an AI engine said in answer to a question, when it said it and the sources it cited. Judgments use a fixed seven-term likelihood scale, and our confidence in each one is stated separately. Every report sets out the limits of the method.
1. What we record
For each situation we write the questions people are likely to put to an AI engine about the company or project. A Read uses dozens of them. We put them to the main AI answer engines on at least two dates, and each Read lists the engines it covers and the dates of each run. Some engines are captured with a commercial monitoring tool and others by hand.
For every answer we keep:
- the engine
- the date
- the question
- the text of the answer
- the sources it cited
Every question is listed in the Read, so your counsel can put the same question to the same engine. We don’t edit an answer after it’s recorded.
2. How sources are sorted
We sort every source the engines cite by type: court and agency records, company filings, news, trade press, advocacy groups, academic work and sites that don’t say who publishes them. The Read shows which engines cited each source and how often. Where an answer conflicts with the public record, the Read quotes the answer and cites the record that contradicts it.
3. How likelihood is stated
Every judgment uses one of seven terms, each with a fixed range. We don’t mix them with other words for likelihood.
| Term | Likelihood |
|---|---|
| Almost no chance | 01 to 05% |
| Very unlikely | 05 to 20% |
| Unlikely | 20 to 45% |
| Roughly even chance | 45 to 55% |
| Likely | 55 to 80% |
| Very likely | 80 to 95% |
| Almost certain | 95 to 99% |
Confidence is a separate judgment about the evidence. We state it as low, moderate or high, in a sentence of its own. That sentence names the sources it depends on. A judgment can be likely and still be held with low confidence when the sources are few or disagree.
4. Where the scale comes from
The seven terms and their ranges, and the rule that keeps likelihood and confidence apart, come from Intelligence Community Directive 203, Analytic Standards. It’s a public document issued by the US Office of the Director of National Intelligence for US intelligence agencies. Probative has no connection with any government agency and isn’t endorsed by one.
We use the scale and the rule on confidence. The directive’s other standards are written for government analysts, and we don’t claim to meet them. One of them requires analysis “based on all available sources of intelligence information.” Our findings rest on public material only: the engines’ answers and the public record.
5. How reports are written
We use Anthropic’s Claude to help draft reports, on an account set so that nothing we send is used to train them. Every quotation, date and judgment is checked against the captures before a report goes out.
6. What a capture doesn’t show
A capture shows what one engine said to one question at one time. Someone asking on another day, or in other words, may get a different answer. A capture doesn’t show what any particular person was told, and a Read doesn’t explain why an engine chose one source over another. Our reports aren’t prepared for use as evidence in any proceeding.
7. When the answers match the record
If the engines’ answers match the public record, the Read says so.
8. Change log
| Version | Date | Change |
|---|---|---|
| 1.0 | September 30, 2026 | First published. |
hello@probativehq.com. We usually reply within one business day.