0002: Event-sourced core

Context

Chronicle's state (beliefs with provenance, rumor content, grudges, reputation) is exactly the kind of state where how you got here matters as much as the current value — beliefs need an evidence chain, rumors need a mutation history, grudges need an originating incident. The debug dashboard's causality timeline and the scenario-based regression suite both depend on being able to reconstruct that history.

Decision

All state is derived by folding over an append-only EventLog (chronicle/events.py). Nothing in the simulation mutates state directly; everything happens by appending an event and re-deriving.

Rationale

• Provenance for free. Talk of the Town's belief model (facets with value, predecessor, evidence, strength) is naturally an event fold — each new belief facet's predecessor link is just "the prior derived state." See docs/research/ for the full citation once filed.
• Headless testability. Scenarios seed the log and assert on derived state — no running game, no mocked engine calls.
• Debuggability. The dashboard's causality timeline is "which events produced this belief," which is a log query, not a bespoke feature.
• Replay. Any world state is reproducible by replaying the log, which makes bug reports and regression scenarios exact rather than approximate.

Consequences

• Derivation functions must stay pure (log → state), which constrains how the LLM tiers can be wired in: LLM output has to become a new event, not a direct state mutation, before it's folded in.
• Log size/performance at ~1,000 NPCs over long play sessions needs attention eventually (snapshotting or compaction), but is out of scope for the initial skeleton.