ADR-005: Pedagogical Vector Retrieval from Ross Bentley Curriculum¶
Status¶
Accepted
Origin¶
New for sprint. Replaces Pitwall's hardcoded DECISION_MATRIX with structured driving curriculum.
Context¶
Pitwall's coaching rules are hardcoded threshold checks:
This fires at the right time but doesn't explain why, can't adapt to skill level, and doesn't connect to a coherent curriculum. A beginner hearing "trail brake" doesn't know what that means or why it matters.
Ross Bentley's Speed Secrets curriculum provides a structured, pedagogically sound framework for teaching performance driving. Concepts like weight transfer, traction circle, slip angle, late apex, and vision have specific telemetry signatures that can be detected in real time.
Decision¶
Encode Ross Bentley's curriculum as pedagogical vectors — structured JSON objects mapping telemetry conditions to driving concepts, with coaching text calibrated per skill level.
Each vector has: - Trigger conditions: telemetry thresholds that activate the vector - Confidence requirements: minimum signal confidence to activate - Physics explanation: the underlying vehicle dynamics concept - Coaching by level: beginner, intermediate, advanced messaging - Anti-patterns: common mistakes detectable from telemetry - Gold Standard reference: what AJ does at this moment
Vectors are matched to incoming frames by the pre-filter. Matched vectors are injected into Gemma 4's prompt as context, enabling the LLM to generate physics-grounded coaching that references the curriculum.
This is context injection, not RAG. The vector DB is small (~50 vectors covering the full curriculum). All vectors are loaded at startup. Matching is simple condition evaluation, not embedding similarity search.
Consequences¶
Positive: Coaching is grounded in a proven curriculum. Explanations reference real physics. Skill-level adaptation is built into each vector. Anti-patterns catch specific mistakes. Gold Standard provides concrete targets. Negative: Vectors must be curated and tested (ADR-008). The curriculum is opinionated (Ross Bentley's approach, not the only valid approach). Vectors don't cover every possible driving situation — edge cases get generic coaching from the LLM.