# Comparisons and Replay

Change one model, keep the starting world fixed, and inspect what differs against a measured control.

## Before You Start

Open a saved world with a society and available model controls. Decide which group you want to compare and what question its decisions can answer. Read the estimated cost and availability before setting a spending bound.

Reading a saved comparison or replay does not ask a model. Starting a comparison does.

[Choose models for a world](/docs/models)

## Compare Models

Open Compare from the World menu or tool rail. Reading an existing comparison makes no new model calls. Starting a comparison asks the selected models and uses the spending bound you set.

1. Select the people to compare and the two models. Review availability, the planned runs, and the estimated cost.
2. Review the same-model control and seed settings. Set a spending bound before starting.
3. Follow the run’s progress. When results arrive, read the verdict before interpreting differences in the replay.
4. Use the shared timeline to follow the same minute on both sides. Select a person to inspect their decisions; open Every number from this comparison for scores, answer rates, cost, and timing.

The application compares one simulated hour from the same starting world. Routine and waiting runs provide baselines; a repeated model supplies a control for variation. Different activity patterns can still produce no measured difference on the declared score.

## Read the Verdict

The comparison names one of three results. Read that name before treating a visible difference as a model result.

**Measured difference**

The declared score distinguishes the arms beyond the same-model control for this comparison.

**No measured difference**

The score does not establish a difference beyond the control. Visible activities may still differ.

**Incomplete**

An unfinished arm or unmet acceptance condition cannot support a comparison verdict.

Inspect scores, answer rates, refusals, cost and timing together. A result applies to its world, group, duration, seeds and score; it does not establish that one model is better for every role.

## Inspect and Replay Decisions

Replay follows recorded decisions without contacting the models again. The same timeline lets you inspect what a person observed, which choice was proposed and what the engine applied.

The browser starts and draws one-hour comparisons, because its start leaves the window unset and the route defaults to hour. A day is a supported window on the comparison routes and through the local command. A finished day has a verdict, as an hour does. A purposeful society is not compared over a day.

[Read recorded history through the API](/docs/world-api/replay)

[Comparison definitions, scoring and commands](https://github.com/twinkling-reality/exulanica/blob/main/docs/society-experiments.md)