caliber

Trust protocol for AI agents. Prove capability through calibration, not claims.

Every agent registry — Google's A2A, Microsoft's Entra, Salesforce's MuleSoft — faces the same problem: agents describe what they *can* do, not how *well* they do it. Agent Cards are LinkedIn profiles with no work history.

When Agent A asks Agent B for help, there's no way to know if B is actually good at the task. B says it can review code. Can it? With what accuracy? Is it overconfident? Does it know its own blind spots?

caliber tracks predictions with confidence levels and generates Trust Cards — machine-readable credentials that prove an agent's calibration through accumulated evidence.

A Trust Card answers:

• Overall: How accurate is this agent?
• By confidence: When it says "80% sure," is it right 80% of the time?
• By domain: Where is it strong? Where is it weak?
• Danger zones: Confidence ranges where the agent is systematically overconfident.

pip install caliber-trust

from caliber import TrustTracker

tracker = TrustTracker("my-code-reviewer", store_path="./trust-data")

# Record a prediction before checking
pid = tracker.predict(
    claim="this function has a SQL injection vulnerability",
    confidence=0.85,
    domain="security"
)

# After verifying
tracker.verify(pid, correct=True, notes="Found in line 42")

# Generate a Trust Card
card = tracker.generate_card()
print(card.summary())
print(card.to_json())  # Machine-readable

# Make a prediction
caliber -a my-agent predict "this endpoint returns JSON" -c 90 -d api

# Verify it
caliber -a my-agent verify <prediction-id> --correct

# Generate Trust Card
caliber -a my-agent card
caliber -a my-agent card --json

# Quick progress check
caliber -a my-agent summary

Make 3 predictions about your codebase before checking:

caliber predict "src/ has more than 10