ADR 0002 — Fan-out reduces via an LLM state, not built-in aggregators

Status: Accepted

Context

Fan-out (sample/over) produces a list of results. Something must collapse the list to one value (self-consistency vote, map-reduce combine, debate synthesis). Two options: built-in aggregator keywords (aggregate: majority|best|concat) on the fan-out state, or an ordinary downstream state that reads the list and reduces it via prompt + gates.

Decision

Reduction is an ordinary downstream state. No built-in aggregators. The fan-out state deposits a list; a normal generative state reads {{list}} and votes / selects / merges in prose.

Consequences

• Keeps the model coherent: everything is states + gates + prose. No second, non-LLM evaluation semantics to specify, validate, or teach.
• The schema stays simpler (no aggregator grammar).
• Reduction is non-deterministic (it's an LLM call) — acceptable, and consistent with the LLM-as-runtime premise (ADR 0004). Deterministic reduction, when needed, can use a code-hook gate on the reducer (ADR 0006) without built-in aggregators.