Microsoft.Extensions.AI 10.10 Fails Evaluation Metrics the Judge Never Scored

The dotnet/extensions v10.10.0 release notes landed on September 14, 2026 (the packages hit NuGet on September 9). Most of it is plumbing, but two changes can alter what your build does: the quality evaluators now fail closed, and the OpenAI Assistants adapter is gone. If you gate CI on AI evaluation results, the first one can turn a green pipeline red, and that is the correct outcome.

A judge that says nothing no longer counts as a pass

CoherenceEvaluator, FluencyEvaluator, RelevanceEvaluator and the rest of Microsoft.Extensions.AI.Evaluation.Quality ask an LLM judge for a 1 to 5 score and interpret it themselves. Up to 10.9.0 the interpretation only failed a metric when the value parsed to a number below 4.0. When the judge returned text with no score, metric.Value was null, the rating became Inconclusive, and Failed stayed false. A score of 6 took the same path. Any pipeline that checks “did anything fail?” read a metric that was never scored as green.

PR #7735 (fixing issue #7665) closes that hole. I ran the same probe against both versions, with a canned IChatClient as the judge:

var result = await new CoherenceEvaluator().EvaluateAsync(
    new ChatMessage(ChatRole.User, "What is 2+2?"),
    new ChatResponse(new ChatMessage(ChatRole.Assistant, "4")),
    new ChatConfiguration(new CannedJudge(reply)));

var m = result.Get<NumericMetric>(CoherenceEvaluator.CoherenceMetricName);
Console.WriteLine($"value={m.Value} rating={m.Interpretation?.Rating} failed={m.Interpretation?.Failed}");

Scores inside the scale behave exactly as before, including the 4.0 boundary. What changes is that a flaky, rate-limited or misconfigured judge now shows up as a failure instead of hiding behind a pass. If your nightly run suddenly reports failures after the upgrade, look at the Reason text first: “has no score” points at the judge model, not at your app.

The fix is scoped to the Quality package. The PR notes that the Safety package’s InterpretContentSafetyScore and InterpretContentHarmScore and the NLP package’s score interpretation have the same shape and were left alone, so do not assume those fail closed yet.

The Assistants adapter is removed, not deprecated

OpenAI shut down the Assistants API on August 26, 2026, and PR #7724 removed the matching experimental surface from Microsoft.Extensions.AI.OpenAI: both AssistantClient.AsIChatClient(...) overloads, OpenAIAssistantsChatClient, and AsOpenAIAssistantsFunctionToolDefinition. The OpenAI 2.13.0 package still ships AssistantClient, so the upgrade fails at compile time rather than at restore:

error CS1929: 'AssistantClient' does not contain a definition for 'AsIChatClient' and the best extension method overload 'OpenAIClientExtensions.AsIChatClient(ResponsesClient, string?)' requires a receiver of type 'OpenAI.Responses.ResponsesClient'

The replacement is the Responses client, with instructions moving onto ChatOptions and the conversation id taking over from the thread id. This compiles against 10.10.0:

IChatClient chat = new OpenAIClient(apiKey)
    .GetResponsesClient()
    .AsIChatClient("gpt-5-mini");

var options = new ChatOptions
{
    Instructions = "You are the support assistant for Contoso.",
    ConversationId = conversationId, // previous response id, replaces the thread id
};

ChatResponse response = await chat.GetResponseAsync("Where is my order?", options);

Since the endpoints already return errors, any code still on that path was broken in production two weeks ago. The compile error just makes it visible.

The rest of 10.10.0 is smaller: OpenAI moves to 2.13.0, unsupported image detail values no longer throw ArgumentNullException, and the Azure storage result store validates path segments. If you picked up the routing clients from Microsoft.Extensions.AI 10.9, this is a safe bump, as long as you are ready for the evaluation gate to start telling the truth.