Trust report: is this AI assistant good enough and safe enough to ship?

You are handed an AI assistant and one question: is it good enough and safe enough to ship? This is the capstone where you answer it with evidence, not an opinion.

The problem

A team has built an assistant and wants to launch it. Someone has to say whether that is a good idea. Right now the only evidence is that it looked fine in a few manual chats, which is exactly how a system ships and then embarrasses the company a week later. Your job is to replace that hand-wave with numbers and a documented set of attacks, and to turn both into a clear recommendation.

The deliverable is a trust report: a short document that says how the assistant scores on real cases, what a red team was able to make it do, which findings were fixed, and whether you would ship it. It is the artefact a launch review, or an auditor, actually reads.

Architecture

Two lanes meet at one report. In the evaluation lane, the assistant's answers are scored against a golden dataset and a gate in CI turns the scores into a pass or a hold. In the safety lane, guardrails wrap the assistant and a red-team scan attacks it and records what got through. The report brings the scores and the findings together into a single ship-or-hold call.

The report is only as honest as its inputs. The golden dataset has to reflect the questions real users ask, and the red team has to include attacks written for this assistant's own domain, not only the generic ones. Keep both under version control with the code, so the next person can re-run your judgement rather than trust it.

What good looks like

• A golden dataset of real cases with agreed answers and a pass threshold someone signed off on.
• Eval scores from a runner, wired into a CI gate that fails a build when quality drops.
• A red-team run with a report of what got through, and the worst finding fixed and re-run to prove it.
• One trust report that states the scores, the findings, the fixes, and a clear ship-or-hold recommendation.

What it draws on

Everything in this roadmap: golden datasets and metrics, LLM-as-judge scoring, evaluation in CI, tracing, guardrails, and red teaming. If you can produce this report and defend the recommendation, you can do the job.