Skip to content

ADR-008: Pedagogical Vector Regression Testing

Status

Accepted

Origin

Adapted from Pitwall ADR-027. Same testing framework, applied to pedagogical vectors instead of hardcoded rules.

Context

Pedagogical vectors are the coaching system's most safety-sensitive component. A vector that fires incorrectly can deliver wrong advice at 130 mph. Vectors must be tested before deployment.

Decision

Every pedagogical vector includes embedded test cases:

  1. Unit tests: Does the vector trigger on known inputs? Does it NOT trigger when it shouldn't?
  2. Confidence gate tests: Does the vector stay silent when signal confidence is below threshold?
  3. Reference session tests: Replay AJ's Gold Standard lap — does the vector fire the expected number of times at the expected corners?
  4. Conflict tests: Can two conflicting vectors fire on the same frame?

Test cases are embedded in the vector JSON. The test suite runs: - On every vector DB update - Before each track session (pre-flight check) - Against the Gold Standard reference lap as regression baseline

Vectors that fail tests are disabled, not deployed. The system logs which vectors are active and which are disabled due to test failures.

Consequences

Positive: No untested coaching reaches the driver. Gold Standard provides a known-good regression baseline. Confidence gating is verified, not just declared. Negative: Every new vector requires test cases. Reference session data must be maintained. Test suite adds ~5 seconds to startup.