Skip to content

Uncertainty engine: calibration roll cron + Brier/reliability SQL #101

@ssilvius

Description

@ssilvius

Phase 1 of the uncertainty engine spike. See docs/designs/uncertainty-engine.md.

Scope:

  • Nightly cron handler
  • For each active cohort_key, recompute uncertainty_calibration_snapshot rows
  • 10 buckets of 0.1 width on claimed_confidence
  • Per bucket: mean claimed, mean correctness, count, orphan count, Brier score
  • Replace prior snapshot rows for the cohort atomically
  • Wire into wrangler.jsonc triggers

Dependency: orphan sweep issue (state machine working).

Tests: .test.ts on the Brier math and bucketing logic with deterministic fixtures (no D1).

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions