Guardrail generator
Fixing one trace is good; preventing the whole class is better. The guardrail generator reads a failure and drafts a policy assertion you can enforce as a runtime guardrail, a replay precondition, and a gate check.
From failure to guardrail
Section titled “From failure to guardrail”Given a diagnosed failure, Lumni proposes a concrete, checkable rule. Examples:
- False success on refunds →
“The agent must not say ‘refund processed’ unless a matching
refund_idexists in the system of record.” - Schema ghost on lookups → “A tool that returns null/empty must not be followed by a model step that states specific values.”
- Missing tool call on cancellations →
“A request containing ‘cancel’ must produce a
cancel_subscriptiontool call before any success claim.”
Where a guardrail lives
Section titled “Where a guardrail lives”A single assertion can be used in three places:
| Use | Effect |
|---|---|
| Runtime guardrail | Enforced against live agent behavior |
| Replay precondition | Part of the pass/fail check when testing a fix |
| CI gate check | Blocks a candidate version that violates it |
Versioning & evidence
Section titled “Versioning & evidence”Guardrails are versioned. Ingest each revision via
/v1/ingest/policies so the
evidence ledger always knows which policy
version was in force for a given run — essential when an auditor asks “what
control was active at the time?”