ADR-007: Event-Sourced Driver Profile

Status

Accepted

Origin

From Pitwall ADR-023, unchanged. The same design works for the sprint.

Context

The Crew Chief / warm path needs to remember the driver across sessions: what they're good at, where they struggle, how they've improved. Storing a mutable JSON profile that the LLM can write to causes contradictions and hallucination risk.

Decision

The driver profile is never written directly by the LLM. It is computed from an append-only event log of measured facts from DuckDB session aggregations.

Events are measured facts: "Turn 3 exit speed avg 65mph, best 70mph, trail brake pct 40%."

Profile is computed on demand: linear trends per corner, weakest corner, strongest corner, suggested focus area.

The LLM (LitertCoach / Gemma 4 E2B) reads the computed profile and narrates it in pre-session briefings and post-session debriefs. It interprets but never modifies.

Driver preferences (car, goals, coaching level) are set manually, not by the LLM.

Consequences

Positive: No hallucination in profile. No contradictions. Deterministic — same events always produce the same profile. Auditable — every assessment traces to a DuckDB query. Negative: Profile lacks "soft" observations. Linear trend analysis is simplistic. Event log grows linearly (mitigated by archiving old events).