@alexvegman/assertions
v0.1.3
Published
Assertion and verdict evaluation for AI agent transcripts against policy — the review logic behind `agentguard review-pr` (@alexvegman/cli), built on @alexvegman/core and @alexvegman/decision.
Downloads
320
Readme
@alexvegman/assertions
Assertion and verdict evaluation for AI agent transcripts against policy — the review logic behind agentguard review-pr (@alexvegman/cli), part of AgentGuard.
What's in here
- Assertion checks (
codingVertical.ts,claimChecks.ts,budget.ts,requirements.ts, and more) — the individual pass/fail/review checks run against a captured agent run. - Verdict aggregation (
verdict.ts) — combines per-assertion results into a final PASS/FAIL/REVIEW, per AgentGuard's "review wins over not_applicable" rule. - Calibration (
calibrationConfidence.ts,fanoutCalibration.ts) — statistical checks on the assertion suite's own reliability. - Escalation (
escalate.ts) — hands an unclear verdict to an@alexvegman/decisionEscalationEngine, leaving it as an explained REVIEW rather than throwing if escalation itself fails. - Mutation-tested — every assertion here is exercised by
@alexvegman/mutations' profiles to confirm it can actually detect the failure it claims to catch.
Install
npm install @alexvegman/assertionsDepends on
@alexvegman/core, @alexvegman/decision.
Used by
@alexvegman/cli, @alexvegman/mcp.
Full documentation: github.com/AlexanderVegazo26/agent-guard.
