// AI RELIABILITY · SELF-ASSESSMENT CHECKLIST

Find Your AI's Failure Points Before Your Users Do

A 50-point self-assessment checklist for teams shipping agents, RAG, and AI copilots. Built to surface the exact reliability gaps that traditional testing and APM can't see, before they become a production incident.

Get the Checklist — $49
Instant PDF download
Secure checkout via StripeInstant downloadNo spam ever
01Evaluation & Testing
You have a labeled eval set, not just anecdotal examples.
A collection of real or representative inputs with known-good expected behavior, versioned like code.
Eval runs are part of your CI pipeline, not a manual pre-launch ritual.
If evals only run when someone remembers to run them, they will eventually not run.
You test against adversarial and edge-case inputs, not just happy-path examples.
Malformed input, contradictory context, and off-topic requests all belong in your eval set.
LLM-as-judge scoring, where used, has been validated against human judgment.
A judge model that hasn't been checked against real human ratings is an unverified assumption, not a metric.
You have regression tests for previously-fixed failure modes.
A bug that was fixed once and never turned into a test case is a bug that will recur silently.
Eval scores are tracked over time, not just checked at a single point.
A model or prompt change that quietly degrades quality months later is invisible without a trend line.
Someone other than the system's builder has reviewed the eval set for blind spots.
The person who built the system is the worst-positioned person to notice what it's not being tested for.
The 50-Point AI Reliability Auditmleg.tech · 01 / Evaluation & Testing
// who this is for

Built for the team that already shipped, not the one still planning.

You've shipped an LLM feature and have a nagging feeling you don't actually know where it's fragile. The demo worked. Production is a different question, and nobody's answered it yet.

You found out about an AI failure from a customer, a Slack screenshot, or a support ticket — not from your own monitoring. If that's happened once, it will happen again in the same blind spot.

You're about to ship an agent, RAG pipeline, or AI copilot and want a real gap-check before launch — not a checklist that stops at "did we write a good system prompt."

// what's inside

Fifty checks. Eight categories. Zero fluff.

Evaluation & Testing

Whether your eval set actually catches adversarial input, or only ever sees the happy path your demo used.

Guardrails & Safety

The fail-open-vs-fail-closed decision most teams never make on purpose, and what happens the first time it matters.

Grounding & Retrieval

The exact citation-checking bug that makes a system confidently cite a fact a document explicitly says doesn't exist.

Observability & Monitoring

Why "the request succeeded" and "the answer was right" are different metrics, and why most dashboards only track the first one.

Plus four more categories covering prompt/model management, human-in-the-loop escalation, deployment, and incident response.

Michael Legemah

Michael Legemah is a Principal AI Engineer who has spent over a decade building production systems for AWS, the U.S. Army, and U.S. Space Force. The last several years focused specifically on agentic AI, RAG pipelines, and the evaluation infrastructure that keeps them honest. This checklist comes directly from failures he's diagnosed and fixed inside real production systems, not from a summary of someone else's blog posts.


Fifty checks. One afternoon. A concrete list of what to fix first.

Get the Checklist — $49
Secure checkout via StripeInstant downloadNo spam ever