Skip to content

World Evolution Engine

Status: Running today (this overview describes the platform framing). For the runtime contract depth (admitted but not yet shipped as a runnable engine) see WEE Execution.

The World Evolution Engine (WEE) is the runtime-owned machinery that makes a world feel alive — characters acting, scenes progressing, events accruing — under contracts that are auditable, replayable, and fail-closed.

What WEE Solves

A world that has rules but no progression is just a database. A world where any code can write canonical truth at any time is just chaos. WEE is what sits between: a typed pipeline that turns participant inputs and scheduled events into proposed mutations, validates them, stages them as commit requests, and either commits to Realm or fails closed under explicit reasons.

WEE runs inside Runtime. Realm remains the truth authority; WEE is the engine that produces well-formed mutation requests for Realm to admit. WEE does not bypass Realm; it composes against Realm.

The Nine Stages

WEE has its own execution stage taxonomy, distinct from Workflow:

StageOrderPurpose
INGRESS1Receive an event proposal
NORMALIZE2Canonicalize the proposal shape
SCHEDULE3Order the work in the engine's queue
DISPATCH4Hand the work to the right handler
TRANSITION5Compute the typed state transition
EFFECT6Compute downstream effects (presence, social, economy)
COMMIT_REQUEST7Stage the proposed mutation as a typed commit request
CHECKPOINT8Snapshot the engine's intermediate state for replay
TERMINAL9The work has reached a terminal outcome

Each stage has typed inputs and outputs. A stage that produces malformed output fails closed; the engine does not silently fall back to a generic stage handler.

Recorded Replay Only

WEE V1 replays from recorded events, checkpoints, and commit-request outcomes only. It doesn't re-infer; it does not make fresh route selections during replay; it does not call models again.

This is a deliberate choice. Replay is for understanding what happened, not for re-executing. Re-inference at replay would mean the engine could produce different answers in replay than in production; that breaks audit reconstruction.

If a system later needs deterministic re-execution, it must be admitted as a separate execution mode under its own contract.

How WEE Differs From Workflow

Both WEE and Runtime's Workflow surface execute multi-step work. They are separate by design.

PropertyWorkflowWEE
OwnerRuntimeRuntime
PurposeGeneral AI execution graph (text/image/etc.)World evolution: turning events into Realm commits
Stage taxonomy15 typed nodes (AI_*, TRANSFORM_*, CONTROL_*)9 stages (INGRESS → ... → TERMINAL)
State machine`ACCEPTED → QUEUED → RUNNING → COMPLETEDFAILED
ReplayRecorded replay; no re-inferenceRecorded replay; no re-inference
Output targetStream / artifact / typed resultRealm commit request

A workflow is the right tool for "generate this image with these parameters." WEE is the right tool for "advance the world by one event under typed contracts."

There is an explicit hardcut: workflow partial-reuse is forbidden as a shortcut for WEE work. Building world evolution out of workflow nodes by side effect is a shadow truth pattern; WEE has its own contract precisely so this does not happen.

Reader Scenario: A World Event Becomes A Realm Commit

A participant performs an action in a world running inside WEE — say, a scripted scene transition.

  1. INGRESS receives the event proposal: "scene S advances to phase P".
  2. NORMALIZE canonicalizes the proposal: the scene id is resolved, the phase is validated against scene rules.
  3. SCHEDULE orders the proposal in the engine's queue. If other proposals are pending for the same scene, ordering is typed.
  4. DISPATCH routes the proposal to the scene-progression handler.
  5. TRANSITION computes the typed state transition: scene moves from phase P-1 to phase P. The transition is bounded by admitted state-transition rules; an undefined transition fails closed.
  6. EFFECT computes the downstream effects — presence updates, possible social-state shifts, economy events that the scene's rules entail.
  7. COMMIT_REQUEST stages the proposed mutation as a typed commit request. The request carries the commit envelope required by Realm: worldId, appId, sessionId, effectClass, scope, schemaId, schemaVersion, actorRefs, reason, evidenceRefs.
  8. CHECKPOINT snapshots the engine's intermediate state. If a later replay needs to reconstruct what happened up to this point, this checkpoint is the anchor.
  9. TERMINAL records the outcome. If Realm admits the commit, the terminal outcome is committed; if Realm refuses (validation failure, authorization failure), the terminal outcome is failed with a typed reason.

This entire sequence is recorded. A future replay can step through it without making any new model calls.

Reader Scenario: Replay Cannot Append History

Suppose an auditor wants to understand what happened when a specific world event went wrong yesterday.

  • Replay can read. The recorded events, checkpoints, and commit-request outcomes are sufficient to reconstruct what the engine did.
  • Replay cannot append. A replay run is not a CANON_MUTATION run; world history is append-only and only canon-mutating runs may append.
  • Replay cannot re-infer. Replay does not re-call models or re-run route selection; it reuses recorded outcomes.

This is what makes replay a real audit tool. If replay could append or re-infer, it would not be reading the past — it would be overwriting it.

Source Basis

Nimi AI open world platform documentation.