Agency

Appearance

Judging eval records ​

agency eval judge compares two .eval.json records from agency eval extract or two eval run directories from agency eval run, sends final responses to an LLM judge, and writes a verdict JSON showing which side better meets your goal.

Synopsis ​

bash
agency eval judge --goal <text> [--out <path>] <recordA.eval.json> <recordB.eval.json>
agency eval judge [--out <path>] <runA> <runB>

Options:

  • --goal <text> — plain-English description of what success looks like. Required for record comparison; rejected for run directories, which carry their own goals in each input's input.json.
  • --samples <n> — judge samples per input. Defaults to 3.
  • --confidence-threshold <n> — minimum input confidence counted as a suite win. Defaults to 50.
  • --margin-threshold <n> — suite win margin required. Defaults to 0.
  • --position-bias <swap|none> — alternate A/B positions across samples or keep original order. Defaults to swap.
  • -o, --out <path> — output verdict JSON path. Defaults to <recordA-stem>.vs.<recordB-stem>.verdict.json in the current working directory.

Example ​

bash
agency eval judge \
  --goal "give the correct capital of India (New Delhi) with no additional text" \
  lib/eval/judge/fixtures/india-A.eval.json \
  lib/eval/judge/fixtures/india-B.eval.json

Compare two run directories against an input suite:

bash
agency eval judge runs/baseline runs/candidate

The command prints a short summary and writes the full verdict:

text
Winner: A (medium)
Reasoning: Response A gives the capital directly, while Response B wraps it in JSON.

Wrote verdict to india-A.vs.india-B.verdict.json

Verdict shape ​

typescript
type PairwiseVerdict = {
  verdictVersion: 1;
  goal: string;
  inputs: [
    { path: string; response: string | null; truncated?: true },
    { path: string; response: string | null; truncated?: true },
  ];
  winner: "A" | "B" | "tie";
  confidence: "low" | "medium" | "high";
  reasoning: string;
  generatedAt: string;
};

inputs[*].response is the exact string sent to the judge, or null when the eval record had no recorded output. truncated: true mirrors eval extract metadata when the source response was already capped at extraction time.

Limitations ​

Behavioral assertions and single-record scoring are not supported yet.