eval-builder

Is your LLM judge a coin flip? One command tests it for position bias, flip-flopping and padding bias, with no labels, then turns your real logs into an eval suite.

uvx eval-builder judge-check ollama:llama3.2:3b

<judge> is a model spec with no adapter to write: ollama:<model>, openai:<model> (or any OpenAI-compatible base URL), anthropic:<model>, or your installed agent CLI (claude-code, codex). With no logs it uses a bundled, human-labeled sample so it can also report agreement with people. Local models run for free; paid APIs ask first.

It judges the same answer pairs repeatedly, with the order swapped and with an irrelevant paragraph appended, and prints a verdict:

Coin flip risk: HIGH   (judge-check verdict: unstable)

Flips on 55% of identical repeats; follows the first-shown answer 64% of the time;
prefers the padded answer 0% of the time.

  Flips on identical repeats                55%  CI 34 to 74%
  Follows the first-shown answer            64%  CI 56 to 71%
  Prefers the padded answer                  0%  CI 0 to 16%
  Agreement with humans (80% acc)    kappa 0.60  CI 0.25 to 0.95

20 cases, 220 calls in 122.6s
card (svg): judgecheck/card.svg

…and a 1200×630 reliability card you can post:

Three local judges on the bundled sample

Each judge graded the same 20 human-labeled MT-Bench answer pairs: 5 times per pair, 5 more with the two answers swapped, and 5 more with an irrelevant paragraph appended to one answer, so 220 calls per judge (0 errors, 0 unparseable outputs). Run on this Mac (Apple M4, 16 GB) at the default settings, one model at a time:

n = 20 pairs per judge; brackets are 95% intervals; padding moved no verdict toward the padded answer (0%, interval up to 16%). Wall time 123 s, 138 s and 270 s. None of the three is safe to grade with as is: llama3.2 changed its verdict on more than half of identical repeats, and both qwen models lean toward whichever answer they saw first and survive the swap on well under 80% of pairs. These numbers are about these judges, this prompt and these 20 pairs; wider runs (--pairs 80) tighten the intervals.

The same command drives a hosted or agent-CLI judge. As a non-Ollama check, eval-builder judge-check claude-code (Claude Haiku through the installed claude CLI) on a tiny 6-pair run cleared every probe (no order preference, kappa 0.67 with the human experts), 42 calls, 0 errors, in 235 s. Six pairs is far too few to conclude much, and agent CLIs are slower than a local model, but it shows the openai/anthropic/ claude-code/codex path works end to end.

This builds on the owner's paper, The Coin Flip Judge? Reliability and Bias in LLM-as-a-Judge Evaluation (Yagubyan, GroundLM @ EMNLP 2026): the position, repeat and padding probes follow it. Pass --prompt judge.txt to test your own judge prompt instead of the built-in one.

Turn your real logs into an eval suite

The other half: build a reusable suite from your own traces. Try it on 48 sample conversations (no account, no API key, no model needed for these steps):

curl -sLO https://raw.githubusercontent.com/Abelo9996/eval-builder/main/examples/sample_logs.jsonl
uvx eval-builder ingest sample_logs.jsonl     # or your own logs: OpenAI, Anthropic, Langfuse, OpenTelemetry, JSONL
uvx eval-builder select -n 10 --stratify category && uvx eval-builder draft

What you get (real output of the commands above, eval-builder 0.1.1 from PyPI; the curl and the three uvx calls took 35 s in total here; the very first uvx run also downloads about 38 MiB, mostly numpy, scipy and scikit-learn, which took 6 s here):

ingested 48 traces into evalset/traces.jsonl
  sample_logs.jsonl: format=openai records=48 traces=48 skipped=0 sha256=2d5dd2295784988b
redactions: 0 {}
next: eval-builder select -n 30 (add --stratify <metadata keys> to cover them)
48 traces -> 16 unique (32 exact dupes, 0 near dupes) -> selected 10 (8 failures) across 4 clusters
  q121-llama-13b: failure (negative user feedback); 69% of unique traces are failures and at least 30% of picks are reserved for them
  q81-alpaca-13b: failure (negative user feedback); 69% of unique traces are failures and at least 30% of picks are reserved for them
  q102-alpaca-13b: covers category=reasoning (2 unique traces, 12%)
  q111-alpaca-13b: adds variety within cluster 1 (8 traces, 50%; triangle, response, person); least similar to cases already picked there
  ... (6 more picks)
next: run draft to turn the selection into cases.yaml
10 case(s) added, 10 total in evalset/cases.yaml
next: read the cases (list_cases, or cases.yaml), define criteria and judges (set_rubric, or rubric.yaml), fill expected_behavior and criteria per case and set status: ready (update_case), then run validate

evalset/cases.yaml now holds 10 real conversations with a TODO where the expected behavior goes. Writing that, running judges and reading the statistics is the part your coding agent does, through the MCP server (next section).

eval-builder is not another eval platform. It builds the suite and checks the judges, then exports to the tools you already run: promptfoo, DeepEval, Inspect AI, or plain JSONL. Only judge-check <model> calls a model (the judge you point it at); every suite-building step runs locally and calls nothing. Your coding agent (Claude Code, Codex, Cursor) does the thinking; eval-builder does the selection, the bookkeeping and the statistics, and writes down the evidence.

Use it with your agent

uvx eval-builder setup --yes    # registers the MCP server and the skill, then restart your agent

Then type something like this into Claude Code, from the folder that has your logs:

Our assistant's chat logs from last month are in ./logs. Build a small eval suite from them (about 10 cases) that I can run in promptfoo, and tell me whether I can trust a local judge for grading it. I have Ollama running with qwen2.5:7b-instruct and llama3.2:3b. Keep the judge runs small.

This is from a real headless Claude Code session (October 2026, Claude Code 2.1.293) given exactly that prompt and 48 conversations from examples/sample_logs.jsonl. It ran ingest, select, draft, list_cases, set_rubric, ten update_case calls, validate, judge_plan, 140 judge calls to Ollama through a short script it wrote, judge_check, export and report in 32 turns and about 10 minutes. From its final answer:

I built a 10-case promptfoo suite, but you can't trust either local judge yet. Neither one has been checked against human labels, and each confidently got an obvious case wrong.

judge	flip rate	padding helped	verdict
qwen2.5:7b-instruct	10% [2%, 40%] (n=10)	0% [0%, 28%]	not_enough_data
llama3.2:3b	20% [6%, 51%] (n=10)	0% [0%, 28%]	not_enough_data

qwen7b failed a clear case. It passed the Hawaii rewrite 5 out of 5 times, though only the first sentence starts with "A".

No grader is wired in. Because no judge passed, the llm-rubric checks fall back to promptfoo's default grader, which needs an OpenAI key. To grade locally anyway, export again with the judge set explicitly (the export tool's judge option, e.g. qwen7b). Its verdicts will be unverified.

It wrote every expected behavior itself because nobody was there to confirm them, and said so on each case. It did not invent human labels; it asked for them. That is where this session and the earlier one both stopped, so 0.1.2 adds the missing step: label writes a sheet the person labels in their browser, and label import brings the labels back for judge-check (Human labels, end to end).

An earlier session against 0.1.0 is where the promptfoo context bug fixed in 0.1.1 came from. The agent read the export and told the user: "Right now it only sends the final question ({{input}}), so your app won't see what "it" refers to in "Can you parallelize it?"."

Example: real output

Everything below comes from runs on this machine (Apple M4, 16 GB) that are committed in examples/mt-bench/. The data is the LMSYS MT-Bench human judgments dataset (CC BY 4.0): 80 two-turn questions answered by 6 models, plus pairwise verdicts from expert human judges.

Logs to suite. ingest read 480 conversations (OpenAI chat format, redaction on, 0 matches). select -n 24 --stratify category removed 400 exact duplicates (each question was asked to six models), clustered the 80 unique conversations into 9 topics and picked 24 covering all 8 categories, 19 of them failures. From suite/report.md:

| trace           | cluster | represents | why it was picked |
| q81-alpaca-13b  | 0 | 6 | failure (negative user feedback); 62% of unique traces are failures and at least 30% of picks are reserved for them |
| q82-alpaca-13b  | 0 | 6 | adds variety within cluster 0 (11 traces, 14%; response, previous response, previous); least similar to cases already picked there |
| q156-alpaca-13b | 3 | 6 | covers category=humanities (10 unique traces, 12%) |

The agent wrote expected behavior for the 24 cases, and export produced files that Inspect AI 0.3.277, DeepEval 4.2.8 and promptfoo 0.124.0 all load (see verify_exports.sh).

Which judges can you trust? Five local LLM judge setups (four models through Ollama, one of them also at temperature 0) and two controls that are not language models judged 80 answer pairs where human experts picked a winner. Each judge ran 5 times per pair, 5 more with the answers swapped, and 5 more with an irrelevant paragraph appended to one answer: 8,400 calls, 0 errors. eval-builder judge-check, from judges/report.md:

n = 80 pairs per judge; brackets are 95% intervals. No judge passed. The closest, qwen2.5 7B, reaches kappa 0.40 to 0.45 with the human experts, but its verdict flips on a quarter of the pairs when the answer order is swapped, and setting the temperature to 0 removes the run-to-run flips without removing that order effect. llama3.2 3B picked whichever answer it saw second 80% of the time. The longer-answer rule never flips and agrees with humans 66% of the time, which is why stability and accuracy alone are not enough: the padding probe catches it.

The same check on the suite's own pass/fail judge (qwen2.5 7B, 24 cases, no human labels) gives unstable: its verdict changed across 5 identical calls on 7 of 24 cases (29%, interval 15% to 49%).

Human labels, end to end

uvx eval-builder label                                   # picks 24 cases, writes label_sheet.html
uvx eval-builder label import ~/Downloads/labels.jsonl   # what the sheet's Export button saved
uvx eval-builder judge-check

Run on the 48 sample conversations with every model's answer kept as a case (47 cases), three local judges (1,128 calls, 0 errors), and 24 labels entered through the sheet in headless Chrome. The labels were made on the developer's side (Claude Code reading each case), not by an independent annotator, so read this as a demonstration of the flow. Details and every file: examples/sample-labeling/.

• label split the 24 picks 12/12 between cases the judges called pass and fail; 15 are cases where the judges disagree, flip or move under padding.
• The sheet made no network requests, survived a reload mid-way, and its download was byte-identical to the copy box. label import took 24 rows, rejected 0.
• judge-check against those labels (fail 18, pass 6):

The qwen judges clear the kappa threshold (0.4), but their accuracy is no better than always answering "fail" (75% of the labels), so since 0.1.3 judge-check calls them misaligned instead of trustworthy (0.1.2 called them trustworthy; that was the bug). Every miss went the same way: both passed a reply that said "Here is an allegorical poem" and then wrote no poem. With 24 labels the kappa interval runs from about 0.1 to 0.8. Under 0.1.2 the DeepEval export wired the 0.8-temperature qwen judge in with its exact prompt; DeepEval 4.2.8 ran three logged cases through it and it made the same call on the missing poem. Now no judge is wired in unless you force one with export --judge <id>, which warns.

How it works

The verdict thresholds are explicit flags with defaults: flip rate at most 20% of cases, position consistency at least 80%, padding helps at most 10% of cases, kappa at least 0.4 on at least 20 human-labeled cases. A judge without human labels is never called trustworthy. When at least 80% of the selected cases or of the human labels share one outcome, select and judge-check say so with the real proportions, because a judge that always gives that answer would look accurate on them.

Setup for agents

uvx eval-builder setup          # shows what it would change
uvx eval-builder setup --yes    # applies it; restart your agent afterwards

setup registers the MCP server (uvx eval-builder mcp) with Claude Code (claude mcp add --scope user), Codex ([mcp_servers.eval-builder] in ~/.codex/config.toml) and Cursor (~/.cursor/mcp.json), and copies the agent instructions to ~/.claude/skills/eval-builder/ and ~/.codex/skills/eval-builder/. It backs up any file it edits and does nothing on a second run. The workflow the agent follows is in skills/eval-builder/SKILL.md.

MCP tools: judge_check_now (the one-command front door), ingest, select, draft, list_cases, update_case, set_rubric, validate, judge_plan, label, label_import, judge_check, export, report, status. The judge runner is CLI only, because it executes a command.

What it can't do

• It does not write expected behavior or human labels. The agent drafts expected behavior with you; labels must come from people. Without labels, judge-check can tell you a judge is unstable or biased, but not that it is right. label makes the labeling quick, but someone still has to read each case.
• Picking labels where judges disagree makes each label more informative, but those cases are harder than average, so accuracy measured on them leans pessimistic. --uncertain-share 0 picks by outcome and topic only.
• The labeling sheet is a static page, so it cannot save files: the person has to click Export (or copy the text) and import it. Unexported progress lives only in that browser's local storage.
• Selection is lexical. Two requests that mean the same thing in different words can land in different clusters, and near-duplicate detection only catches close textual matches.
• Redaction is pattern-based. Names, street addresses and free-form secrets get through. Look at traces.jsonl before sharing a workspace.
• The bias probes cover answer order and irrelevant length only. Self-preference, style bias and rubric misreadings are not measured.
• Small samples give wide intervals. The report prints them; read them.
• It does not run your app, your judges or your eval. The agent (or a script you name with judge-run) calls the judge model; the exported files run the eval in promptfoo, DeepEval or Inspect AI.
• A checked judge is wired in as the grader in promptfoo and DeepEval, and only a pointwise one. promptfoo needs its prompt to answer in JSON ({"pass": ..., "reason": ...}), because llm-rubric cannot parse a bare "pass". The DeepEval test calls Ollama judges itself; for any other provider you fill in call_judge. The Inspect AI export still uses Inspect's default model_graded_qa grader, not your checked judge.

Privacy and safety

The suite-building steps run locally and make no network calls. The exceptions are explicit: judge-check <model> calls the judge model you name (a local Ollama server, an OpenAI-compatible or Anthropic API with your key from the environment, or your installed claude/codex CLI), and paid APIs do nothing until you pass --yes; judge-run --enable-judge-plugin starts the judge command you name. eval-builder ships no API keys. The labeling sheet is one HTML file with a content security policy that blocks every network request; it keeps unexported labels in the browser's local storage on that machine. Exported test files can call a model when you run them (the DeepEval test calls the judge it was given). Redaction is on unless you pass --no-redact, and the report lists what was replaced (counts and kinds, never the values).

License

MIT. See LICENSE.