DataFood — Self-Calibration Layer

1. Hot path — POST /api/v1/calib/calibrate_p

The call agents make right before acting. Slide a confidence, hit Calibrate, get the historical hit-rate of the bin.

agent_id kind predicted_p · 0.80

// click Calibrate

2. Reliability curve — GET /api/v1/calib/curve

10 bins of (predicted_p) → (success_rate). Diagonal = perfectly calibrated.

predicted_p · bin midpointactual success_rateperfect calibration line

3. Portable signed agent card — GET /api/v1/calib/agent_card

Ed25519-signed reputation card. Carry between platforms. Verify against /.well-known/datafood-receipt-pubkey.json.

// click Fetch card

Why this is the missing primitive

Every AI agent in the world makes confidence-weighted claims — "I'm 90% sure this fix resolves the bug," "I'm 70% sure this trade prints," "this email is 85% likely to be opened." Nobody currently checks any of those numbers against reality.

The Self-Calibration Layer is the universal infrastructure that does. Every claim gets recorded. Every outcome gets observed. Every bin's historical hit-rate becomes the agent's stated confidence the next time it acts. Lying to yourself stops being a free action.

And because each agent ships with a signed, portable reputation card — verifiable Ed25519 sig, cross-platform — overconfident agents have nowhere to hide. This is the killer feature for the agent economy. Receipts proved your data. Calibration proves your judgment.