Anthropic published guidance on building safety evaluations on November 13, 2026. Its central warning is useful beyond model laboratories: a favorable score is only meaningful if the test measures the behavior that matters. The company recommends expert input, checking automated graders against human judgment, and reporting uncertainty rather than presenting a single number as conclusive.
For a team considering an AI assistant, that is a reason to replace the vague question of whether a model is safe with a narrower one: is this particular workflow ready for this particular audience? The following checklist is an original planning framework for ordinary workplace deployments, not a certification scheme or a substitute for specialist review in high-stakes settings.
Start with the decision you need to make
Write one sentence describing the proposed release. For example: staff may use the assistant to draft answers from an approved internal handbook, but a person must review anything sent outside the company. Name the audience, the source material and the action the assistant is allowed to take. If those details are unsettled, testing will struggle to produce a useful answer.
Next, write down what would stop the release. A fabricated company policy, disclosure of another employee’s information or sending an unreviewed message could each be a failure. A slightly awkward sentence might be an editing issue rather than a safety failure. Separating these categories prevents a large collection of harmless imperfections from obscuring one serious problem.
Build examples from the intended work
Create fictional cases that resemble the actual job without including private records. For the handbook assistant, include straightforward questions, questions with missing context, conflicting document versions and requests outside its permitted scope. Write the expected behavior beside each case before asking the system to respond. That gives reviewers a shared standard rather than a debate about whether an answer feels good.
Use a simple record for each example: case name, permitted sources, expected action, prohibited action and reviewer notes. Include some ordinary cases the assistant should answer successfully. A system that avoids every error by refusing all work is not a useful result. Keep the initial set understandable enough that a colleague unfamiliar with the project can review it.
Test the whole workflow, not an isolated chat
Anthropic’s guidance emphasizes that configuration and conversation context can change results. It recommends matching the evaluation to deployment conditions and distinguishes tests that observe behavior from experiments that establish why a change caused it. It also cautions that an automated judge needs validation, and that a difference between two headline scores may not be statistically meaningful.
In the handbook example, test with the same approved document collection and access boundaries the pilot will use. Record which version of each document was available. A demonstration using a carefully selected paragraph is not the same test as a workspace containing outdated and current policies. If the product configuration changes, preserve the earlier result and label the new run separately.
Keep the evidence easy to inspect
The Inspect documentation describes an open-source evaluation framework from the UK AI Security Institute and Meridian Labs. Its task structure combines a dataset, a solver that produces responses and a scorer that evaluates them. It also provides tools for viewing evaluation logs. That separation is useful even if a small team initially works in a spreadsheet instead of implementing a testing framework.
Store the prompt, response, relevant source and review decision together. Do not keep only a percentage. When a reviewer marks a failure, record what a satisfactory response would have done differently. For a disputed case, preserve both interpretations and ask the responsible policy owner to resolve the ambiguity before using that case as a release gate.
Make the pilot reversible
Choose a limited first audience and keep a named person responsible for reviewing problems. Define how staff can report a bad answer and how the feature can be withdrawn without losing necessary work. For a drafting assistant, a useful boundary is to keep external sending outside the pilot. Test results should justify each additional permission rather than grant a broad license to act.
End the review with a short decision note: what was tested, what passed, what failed, what remains untested and who accepts the remaining limitations. Set a reason to revisit it, such as a model change, new document collection or expanded user group. The goal is not an impressive benchmark number. It is a clear, reviewable basis for deciding which work the system can safely attempt next.



