Simulator measurement · After artifact and callback verification
From an explanation
to a testable choice.
Water on · basalt · amber light · warm
Energy 40/100 · Did not survive · 2 internal experiment credits.
A learner can see how evidence changes an explanation and why the next experiment is worth spending a limited experiment budget on.
The next decision becomes inspectable: what evidence the agent saw, what it changed in its explanation, and what its choice ruled out. That is the learning benefit; a better score than a local algorithm is not established.
Your turn: what should we test next?
Choose a controlled test, explain your reasoning and inspect how its possible outcomes would divide the ten remaining models. No experiment or payment is made.
Compare against the local choices frozen before this response
| Selector | Experiment | Expected remaining models |
|---|---|---|
| Real IMD | Water on · basalt · amber light · warm | 7.86 |
| ada | Water off · basalt · amber light · warm | 7.86 |
| kepler | Water on · basalt · amber light · warm | 7.86 |
| lin | Water off · basalt · amber light · warm | 7.86 |
Baselines know the 24-model family and response formula. The IMD request contains observations and previous agent feedback, not that family or formula. This is not an equal-information benchmark.
These expectations are calculated from the 14-model prior, not separate measurements of the local choices.