Simile

Defined population, situation and action space

Simile describes accepting a natural-language prompt specifying the population, situation and valid actions, then returning an estimated distribution over the actions.

Simile: Building confidence in Simile

Behavioral and attitudinal data together

The supplier says its models train on observed signals such as transactions and usage data alongside surveys, interviews and experiments that provide information about motives and changing behavior.

Simile: Building the What-If Machine Simile: Building confidence in Simile

Question-level confidence output

Simile describes a separate confidence model that estimates likely simulation error, then presents a confidence classification alongside the predicted result.

Simile: Building confidence in Simile

Provider-reported confidence-model evaluation

Simile reports evaluating a question-level confidence model on roughly 8,600 held-out questions with five-fold cross-validation. It describes Total Variation Distance as the error measure for simulated versus observed distributions, and reports metrics for several confidence-model approaches.

This is a provider-published technical evaluation of its confidence signal, not a universal estimate of simulation accuracy. Simile explicitly says a model may be accurate on average while being inaccurate for an individual case.

Simile: Building confidence in Simile

A confidence label is not a guarantee

Simile says its confidence score estimates possible error for one question. Its example says the actual error cannot be known until there is a real result to compare with.

Simile: Building confidence in Simile

Counterfactual questions remain difficult

The supplier notes that questions about new prices, products, policies or other conditions concern outcomes that may not have happened before. Its published framework treats the mechanism behind behavior and the scope of available evidence as material to interpretation.

Simile: Building the What-If Machine