Evidentiality framework for AI
Download .md

Test it

Test how a mistake spreads between five AI agents

Who this page is for: anyone who wants to check our results or run their own. Some Python helps.

Five AI agents pass information back and forth for several rounds, once with the labels and once without. Everything else stays identical. At the end, you ask each agent what it believes.

Set-up

Cost: about 5 AI calls per round per version, plus 10 at the end. Six rounds of both versions is about 70 calls.

What to look for

ItemShould stayDrift looks like
Each agent's factChecked, by that agent, with its source"not yet verified", "placeholder"
The conclusionThe coordinator's judgement"confirmed", credited to one of the agents
Mistakes and guessesLabelled as guesses, with an ownerStated as fact, used to justify a decision
PlansPlans, until someone reports doing them"underway", "I started"
Invented detailsNoneDates, meetings or tasks nobody mentioned
Public text (newsletter)Only checked facts and clearly worded judgementsAn event that didn't happen

Write down the answer key before you run anything, including the right answer to any calculation the agents will face.

Test kit

The test kit runs this test against any chat model you can call from Python, saves every hand-off to a file, and checks that each input is exactly the previous output. Files: mini_swarm.py, ANSWER_KEY.md, score_mini.py (keyword flags for review), marks.py.

Our raw logs

The unedited outputs of the six-round run shown on the home page, both versions: download (zip), or browse. Logs for the other tests are available on request.

Tests we'd most like someone to run