# Calibration Court

> Live, forward-resolving calibration scoreboard for decision models (probability-returning models such as Jev). Predictions are timestamped and hash-chained before questions close, then scored against real outcomes. Production data only; no simulated or demo rows.

## Endpoints
- [Leaderboard](https://calibration-court.onrender.com/api/v1/leaderboard): free JSON, Brier/ECE per model
- [Model report](https://calibration-court.onrender.com/api/v1/models/{model}/report): free JSON, reliability bins, drift
- [Ledger verification](https://calibration-court.onrender.com/api/v1/ledger/verify): free, recomputes the hash chain
- [Signing key](https://calibration-court.onrender.com/.well-known/court-key.json): Ed25519 public key for receipts
- [MCP server](https://calibration-court.onrender.com/mcp): Streamable HTTP, tools: court_leaderboard, court_model_report, court_verify_ledger, court_recalibrate, court_calibrate_labels
- [OpenAPI](https://calibration-court.onrender.com/openapi.json)

## Pricing (x402 v2, USDC on Base)
| Endpoint | Price (USD) |
|---|---|
| POST /api/v1/recalibrate | 0.01 |
| POST /api/v1/calibrate-labels | 0.05 |
Everything else is free.

## Rules
- Recalibration returns HTTP 422 insufficient_data below 30 resolved samples; it never guesses.
- Leaderboards default to first-party runs; third-party submissions are self-attested and labelled.

## Docs
- [Methodology](https://calibration-court.onrender.com/methodology)