TBV docs

Runtime verification for non-deterministic AI agent actions and outputs. Submit each call to an independent verification agent; route low-confidence cases to human reviewers; close the loop without changing your primary model.

Quickstart →

Install the SDK and submit your first verification in under 5 minutes.

API reference →

All endpoints, request/response shapes, and error codes.

Task lifecycle →

How a task moves from pending to completed — and what happens in between.

Examples →

Refund classification, fraud screening, content moderation, tool-call validation.

How it works

1. Your code calls TBV with a prompt, the context, and the set of allowed verdicts.
2. Your verification agent runs — an LLM judge constrained by your policy plus authoritative typologies. It returns a verdict and a confidence score.
3. If confidence ≥ threshold, the verification completes. Your code gets the verdict.
4. Otherwise, it routes to a human reviewer at review.tobeverified.com. Verdicts attribute to the reviewer; the data feeds future calibration.

Where TBV fits

Use TBV where unit tests can't reach — anywhere you ship LLM output to a user or a downstream system and "wrong" has a real cost:

· Refund / billing classification
· Fraud screening & AML decisions
· PII redaction & content moderation
· Intent routing for support tickets
· Tool / function-call validation
· Anything else where confidence-routed second opinions matter

These docs are versioned alongside the SDK. Current version: 0.5.1 (developer preview).