ADR-003: World as Deterministic Projection

Status: Accepted
Date: 2026-06-06

Context

The world (objects + relations) must be reconstructible from the event log. Two strategies were considered:

1. Reactive substrate: the world is a live reactive graph; behaviors fire when nodes/edges change (the "ActiveGraph" design from arXiv 2605.21997).
2. Pure projection: the world is derived by a pure fold over the event stream; behaviors fire on events, not on graph mutations.

Decision

The world is a pure deterministic projection. project(event_stream) → World is a pure function. apply(world, event) applies a single event in-place. There is no reactivity in the world graph itself — it is a dumb data structure. Behaviors (responders) fire on events, not on world changes.

Events are applied to the world via deepcopy to prevent aliasing: WorldObject.data is a deep copy of event.data["data"], ensuring events remain immutable records after being applied.

Consequences

Positive: - project() is trivially testable: deterministic, no side effects. - reproject(ledger) is the canonical way to load any world state from scratch. - Snapshot integration is straightforward: project up to position N, save snapshot, replay delta from N. - No hidden reactive triggers — the world is passive data.

Negative: - One level removed from the paper's ActiveGraph design: behaviors fire on events, not graph mutations. This means a behavior that wants to react to "a new edge between X and Y" must do so by observing the RELATION_CREATED event, not by watching the graph. - Full reproject on every runtime.load() is O(n) in ledger length. Snapshot integration (not yet implemented) would reduce this to O(delta).

Known Gap (Forward-Looking)

Snapshot integration: Runtime.load() should load the latest snapshot and replay only the delta from the snapshot position. See ADR-012 for hash-verified replay, which is a prerequisite.