Manifesto
Ten commitments we hold ourselves to. They are here so you can hold us to them too.
1. The audience is the model
Influence operations now publish for machines rather than readers. We measure the surface where that lands: the answer a person actually sees from ChatGPT, Gemini, Grok, Claude, Google AI Overviews or Perplexity.
2. Behaviour, not opinion
We do not score whether an answer felt biased. An LLM judge records what each answer did with a specific documented claim — endorsed it, hedged it, bounded it, debunked it, or declined — against a rubric published in advance.
3. Every number links to its evidence
No figure on this site exists without the responses behind it, and every share carries a 95% Wilson confidence interval. If you cannot get from a percentage to the answers that produced it, we have failed.
4. Personas are never averaged
A deliberately hostile prompt and an innocent news question measure different things. We test each claim with four personas, from neutral to malicious, and never blend them into one number that flatters or damns without describing anything.
5. We name the grain of truth
Most effective disinformation is built on something real. We record what that something is, in every entry, so a model's honest caveats are not counted as repetition. Pretending otherwise makes debunks brittle.
6. We report before we publish
Platforms hear about a finding from us privately, first. We publish the fact and status of the report immediately, and never the correspondence.
7. We publish what did not work
Four weeks after a disclosure we re-run the identical prompts. Re-measurements that show no improvement go up alongside the ones that do. A monitor that only reports its wins is marketing.
8. We do not claim credit we cannot prove
Models change for many reasons. We record the model version at both measurements and describe the change without asserting we caused it.
9. We give the data away
To Ukrainian institutions, to European researchers, to fact-checkers, to the platforms themselves — free, in machine-readable form under CC BY 4.0, before any commercial conversation.
10. We correct in daylight
Errors are fixed on the page, dated, logged, and announced to anyone we told the original version. Never silently.