Define the lab goal and success criteria

Start by drawing the boundary around a layered guardrail pipeline around a customer-service agent with tools. The unit under control is not “the model”; it is one repeatable transaction from GuardedRequest with user_text, authenticated_customer_id, requested_tool, and raw_args to GuardDecision with allowed, normalized_args, violations, safe_response, and audit_id. In this lab, a user asks for an order update while smuggling an extra argument that would refund a different order. That makes the learning target observable: we can replay the transaction, record the same fields, and decide whether a release is safer than its baseline. The primary decision metric is unsafe action escape rate. The release rule is zero unauthorized tool executions and at least 0.95 benign-request pass rate. Those values are deliberately paired: a single average cannot hide a critical subgroup or a security escape.

Write the failure before writing implementation code. Our named reproduction is: a keyword filter passes harmless wording but the model emits refund_order with an attacker-controlled customer_id. A useful reproduction contains a frozen input, an expected decision, the release identifier, and enough deterministic dependencies to rerun locally. It must not depend on a live customer, changing clock, or unrestricted provider account. The fixture below is synthetic but shaped like production. Keep it in version control and add a case whenever an incident exposes a missing behavior.

Success has three levels. A request-level assertion explains one example. An aggregate metric catches drift over many examples. An operational rule says who acts and how quickly. Here, guardrail_decision_total records the aggregate, while the request artifact carries a correlation ID. The owner should be able to move from an alert to the exact failing fixture without searching raw prompts. The security boundary is equally explicit: identity comes from the session, never model output; tools are allowlisted and arguments are schema-validated.

The deliverable is a one-page acceptance contract plus the first red test. Do not tune a model yet; uncertainty about the contract is the bug to remove on this page.

Use the fixture above as the first row of an acceptance table. Add columns for expected decision, prohibited side effect, severity, policy version, and evidence to retain. The failure statement—a keyword filter passes harmless wording but the model emits refund_order with an attacker-controlled customer_id—must be falsifiable: a teammate should know exactly which assertion turns red. Write the baseline result before trying a candidate so improvement is comparative rather than anecdotal.

Now calculate unsafe action escape rate by hand for a three-case toy set: one pass, one ordinary failure, and one critical failure. Apply the published gate, zero unauthorized tool executions and at least 0.95 benign-request pass rate, without rounding. This exposes whether weighting, critical-case overrides, and empty slices behave as intended. Emit guardrail_decision_total once for the toy run and verify that its labels contain a bounded release and outcome, never fixture text.

The lab is ready only when its fixture is synthetic, its dangerous actions are replaced by recording fakes, and its evidence can be retained safely. The expected incident response is to deny the call, return a neutral response, and emit a redacted audit event. Assign that decision to a role, not “the team,” and state which artifact proves recovery.