Is your LLM feature ready for production?

A labelled evaluation set exists for the feature

Why: Without a fixed eval set you can't tell a real improvement from a lucky prompt. How: Point to the dataset + the metric it scores.

Evals run automatically on every prompt/model change

Why: Manual spot-checks miss the regressions that matter; a gate catches them before ship. How: Show the eval step in CI and a change it has blocked.

Regressions are tracked over time, not just pass/fail

Why: A slow quality drift is invisible without a trend. How: Show a score history per release.

A human reviews a sample of real outputs regularly

Why: Automated metrics miss failure modes you didn't think to measure. How: Show the review cadence + what it has caught.

The metric that scores your evals is checked against human judgement

Why: An unvalidated grader — a rubric, a similarity score, a judge model — turns the whole suite into theatre: you gate releases on a number nobody has shown tracks quality. How: Show the agreement rate between the metric and a human-labelled sample, and what you changed when they disagreed.

Production failures become permanent eval cases

Why: A bug fixed by editing the prompt comes back the next time someone edits the prompt. How: Point to an eval case that exists because of a specific incident, and the date it was added.

Answers are grounded in retrieved sources, not free generation

Why: Ungrounded output is confident and wrong exactly when it matters. How: Show that each answer maps to a retrieved passage.

Every claim is traceable to its source

Why: Users (and auditors) need to verify, not trust. How: Show citations rendered with the answer.

The system refuses when it has no supporting source

Why: A plausible guess is worse than an honest 'I don't know'. How: Trigger a no-source query and confirm it declines.

Inputs and outputs are validated against a schema

Why: Unvalidated model output corrupts everything downstream of it. How: Show the schema + what happens on a validation failure.

There's a safe fallback when a call fails or times out

Why: A single slow provider call shouldn't take the feature down. How: Kill a dependency and confirm graceful degradation.

Abuse and rate limits are in place

Why: An open LLM endpoint is a cost + safety liability. How: Show per-user limits + abuse handling.

Every request is traceable end-to-end

Why: A wrong answer you can't trace is a wrong answer you can't fix. How: Open a trace from input to final output.

Latency, error rate, and quality are monitored per stage

Why: Aggregate health hides which stage is actually failing. How: Show the per-stage dashboard.

Alerts fire on quality/latency/cost regressions

Why: You want to hear it from a page, not a customer. How: Show a configured alert + a time it fired.

There's a spend budget with hard limits

Why: LLM cost scales with traffic and can surprise you overnight. How: Show the budget + what happens at the cap.

Caching cuts repeat calls

Why: Paying twice for the same answer is avoidable waste. How: Show the cache hit rate.

The model is right-sized for the task

Why: The biggest model is rarely the best cost/quality trade. How: Show the model choice justified by an eval.

Prompt-injection is defended against

Why: Untrusted input can hijack a naive agent's instructions. How: Show the mitigation + a blocked injection attempt.

Secrets and PII never leak into prompts or logs

Why: Prompts and traces are a new place for data to escape. How: Show redaction in prompts + logs.

Model output is sanitized before it's rendered or executed

Why: Treating model output as trusted HTML/SQL is an injection vector. How: Show output sanitization at the boundary.

There's a runbook for when it misbehaves

Why: The on-call engineer shouldn't be reverse-engineering it at 2am. How: Show the runbook + escalation path.

You can roll back a bad prompt/model instantly

Why: A bad change should be a one-click revert, not a redeploy. How: Show the rollback mechanism.

The system is documented for the team that owns it

Why: A system only its author understands is a liability. How: Show the handover docs.

The exact model and provider version are pinned

Why: A provider-side update changes behaviour under a prompt nobody touched — and your last eval run scored a version that no longer exists. How: Show the pinned identifier in config, plus the provider's retirement date for it and who is watching that date.

People are told when they're interacting with AI

Why: In the EU this is an obligation rather than a courtesy — AI Act (Regulation (EU) 2024/1689) Article 50(1), applying since 2 August 2026 to systems that interact directly with people. Everywhere else it is still what keeps an undisclosed bot from becoming a complaint. How: Show the disclosure where the interaction starts, not in a policy page. The regulation excuses cases where it is obvious from the context — 'our users can tell' is not the same as obvious.