clef-evals v0.2.0: production quality

Production-grade rewrite of the calibration-first Clef evaluation toolkit.

Core

• Typed public API: ClefJudge / AsyncClefJudge over the real Clef schema (noul, choice, score questions with per-option probabilities)
• ClefClient: retries with exponential backoff + jitter, Retry-After support, configurable timeouts, structured ClefError hierarchy, sync + async (httpx)
• Config from environment with one-shot validation (reports all problems at once)
• Metrics: accuracy, ECE, binary + multiclass Brier, latency p50/p95/p99, cost per 1k calls from token usage
• Real CLI (clef-eval run) with JSON output and gate exit codes

Quality

• 122 unit tests, >90% coverage enforced, integration tests excluded by default
• CI matrix 3.10/3.11/3.12, ruff clean, py.typed shipped
• Regression-gate reusable GitHub Action (current vs committed baseline)
• evals/ benchmark harness + committed datasets + published reference numbers
• Kaggle kernel for cloud reproduction, calibration walkthrough notebook, showcase video renderer

clef-evals v0.1.0

Calibration-first evaluation toolkit for Cloudflare's Clef decision models.

• ClefJudge: run Clef as LLM-as-judge over eval sets
• ece() / brier_score(): calibration metrics
• clef-eval CLI entry point