[API Proposal]: Expose evaluator chain of thought as structured data on EvaluationMetric #7805

Background and motivation

The quality evaluators in Microsoft.Extensions.AI.Evaluation.Quality ask the judge model for (chain of thought), (reason) and (score). Reason and Value are first-class properties on the metric. The chain of thought is only available as an Informational diagnostic, prefixed with a hardcoded English label (EvaluationMetricExtensions.cs#L76-L80):

We show evaluation results to end users in a localized product, and we instruct the judge to answer in the user's language. The chain of thought comes back localized, but the only way to get it is to string-match the English label. That breaks silently if the wording changes, and it mixes developer diagnostics with user-facing content.

API Proposal

Store the parsed chain of thought in EvaluationMetric.Metadata under a documented

public static class EvaluationMetricMetadataKeys
{
    public const string ChainOfThought = "evaluation-chain-of-thought";
}

API Usage

if (metric.Metadata?.TryGetValue(EvaluationMetricMetadataKeys.ChainOfThought, out string? cot) == true)
{
    ShowReasoning(cot);
}

Alternative Designs

Perhaps:

• A dedicated string? ChainOfThought property on EvaluationMetric.
• Expose the label as a public constant so callers can strip it reliably.

Risks

Minimal