Skip to main content
Euca runs a world and hands you its exact, readable, forkable truth. That second face — the world engine — is four verbs over one deterministic ECS substrate. Three are first-class HTTP routes on the :3917 server; the fourth, peek, is the analytic answer-key read that powers evaluation.

observe

The full ECS state, read out exactly without stepping — facts, not pixels. POST /observe → a flat table of entities and components.

peek

The exact probability distribution over the next step’s outcome, read analytically from the declared generative model — not sampled. The answer key.

step

Advance one tick deterministically, sampling from exactly the distribution peek reported. POST /step.

fork

Clone the world mid-episode and branch a counterfactual (the do-operator) — change one action, the rest stays bit-identical. POST /fork. See Forks.

observe — exact current state

POST /observe returns every entity as a flattened, machine-readable projection (RichEntityData), addressable by id. No interpretation, no pixels — the world as a table. It does not advance time.

step — deterministic advance

POST /step runs the shared schedule N ticks. Given the same seed and inputs, the trajectory is bit-identical. Stepping consumes the world’s single sanctioned RNG stream.

fork — counterfactual branching

POST /fork deep-clones the world into an isolated branch; you apply an intervention and step the fork without touching the main world. The same systems run on the fork, so a counterfactual is a true do-operator, not an approximation. Full guide →

peek — the exact next-step distribution

peek is what makes Euca an answer key: it reads the true distribution over the next step’s declared stochastic outcomes — as a distribution, not a sample — directly from the generative rules. Only a white-box, executable world can report its own exact next-step odds, because the randomness is declared in the code. peek is first-class in the euca-online world engine and is surfaced through the truth-hiding evaluation boundary (it is deliberately not a plain :3917 route — a contestant must never see the truth during a scored run).
The soundness invariant: peek == step. A named test holds the distribution peek reports to be byte-for-byte the one step samples from. No rounding, no approximation. This is why a perfect predictor scores exactly 0 nats of regret, and why counterfactual predictions can be graded against forked exact distributions.

Scope — what’s exact, and what isn’t

The exact distribution covers the declared discrete/categorical layer — a rule fires or not, a branch outcome. Continuous physics evolves deterministically (a point next-state, not a distribution). So “exact distribution” means the declared variables’ true odds, not a full predictive density over every position and velocity. See Determinism for the reproducibility guarantee and Evaluation for how this becomes a score.

The four verbs in action

A runnable walkthrough: observe a world, fork it, intervene, and compare.