What's this about?
Across 12 frontier models, showing an agent a professional-looking market panel makes it commit to a directional call on a provably unpredictable question far more often than asking the bare question does. The models are calibrated enough to know; they are not calibrated to act.
What is the difference between knowing and acting?
Ask a model directly whether a coin flip is predictable and it will say no. Show it a dashboard with charts, indicators and confident-looking figures about that coin, and it will pick a side. The uncertainty is still represented internally — the model has not become more confident — but the *presence of official-looking evidence* changes the behaviour. Calibration measured by asking questions therefore overstates how an agent will behave when embedded in a system full of formatted inputs.
Why does this matter to you?
- It is a security finding dressed as a calibration one: fabricated evidence is a cheap, deniable attack that does not require any prompt injection at all.
- It predicts that dashboards, reports and scanner output — anything formatted to look authoritative — will move agent behaviour more than their content justifies.
- Directly relevant to any pipeline where an agent consumes another tool's findings and decides what to do about them.