Overview

Create and run evaluation suites, watch live benchmark progress, view scorecards, compare model performance, and integrate eval runs with CI workflows from the CLI.

Quick install

npm install -g nerve   # or: npx nerve
nerve --version

Subcommands

eval

Example:

nerve eval

eval suites

Example:

nerve eval suites

eval list

Example:

nerve eval list

eval get <suiteId>

Example:

nerve eval get <suiteId>

eval create

Flags:

• --file <path>

Example:

nerve eval create

eval run <suiteId>

Flags:

• -m, --model <id>
• --combo <name>
• --concurrency <n>
• --tag <tag>
• --watch

Example:

nerve eval run <suiteId>

eval list

Flags:

• --suite <id>
• --status <s>
• --since <ts>
• --limit <n>

Example:

nerve eval list

eval get <runId>

Example:

nerve eval get <runId>

eval results <runId>

Flags:

• --failed

Example:

nerve eval results <runId>

eval cancel <runId>

Flags:

• --yes

Example:

nerve eval cancel <runId>

eval scorecard <runId>

Example:

nerve eval scorecard <runId>

simulate [prompt]

Flags:

• --file <path>
• -m, --model <id>
• --combo <name>
• --reasoning-effort <level>
• --thinking-budget <n>
• --explain

Example:

nerve simulate [prompt]

Nerve — CLI Evals

Requires the nerve CLI. See CLI entry-point skill for install + global flags.

What are evals?

Evals are automated test suites that score LLM outputs against expected answers or rubrics. Nerve stores suites and run results in its local database.

Eval suites

nerve eval suites list                       # List all eval suites
nerve eval suites list --json                # JSON output

nerve eval suites get <suiteId>              # Full suite definition

Create a suite

nerve eval suites create \
  --name "code-quality" \
  --rubric "exact-match" \
  --samples-file ./samples.jsonl                 # JSONL: {input, expected_output}

Rubric options: exact-match, contains, llm-judge, regex.

--samples-file format (one JSON object per line):

{"input": "What is 2+2?", "expected_output": "4"}
{"input": "Translate 'hello' to Spanish", "expected_output": "hola"}

Run an eval

nerve eval suites run <suiteId> \
  --model claude-sonnet-4-6                      # Run suite against a specific model

nerve eval suites run <suiteId> \
  --model gpt-4o \
  --watch                                        # Live TUI progress (EvalWatch)

The run is asynchronous. Use --watch for a live terminal dashboard or poll manually:

RUN_ID=$(nerve eval suites run <suiteId> --model claude-sonnet-4-6 --output json | jq -r '.id')
nerve eval get $RUN_ID

Manage runs

nerve eval list                              # List all eval runs
nerve eval list --json

nerve eval get <runId>                       # Run details (status, model, score)
nerve eval results <runId>                   # Per-sample results
nerve eval scorecard <runId>                 # Full scorecard with pass/fail per sample
nerve eval cancel <runId>                    # Cancel a running eval

Scorecard output

nerve eval scorecard <runId> --output json

Response fields per sample:

{
  "id": "sample-1",
  "score": 0.95,
  "passed": true,
  "input": "What is 2+2?",
  "output": "4",
  "expected": "4"
}

Comparing models

Run the same suite against multiple models and compare:

for MODEL in claude-sonnet-4-6 gpt-4o gemini-2.0-flash; do
  nerve eval suites run $SUITE_ID --model $MODEL --output json | jq '{model: .model, score: .score}'
done

CI integration

# Run and fail CI if score drops below threshold
SCORE=$(nerve eval suites run $SUITE_ID --model claude-sonnet-4-6 --output json | jq -r '.score')
python3 -c "import sys; score=float('$SCORE'); sys.exit(0 if score >= 0.90 else 1)"

Errors

• suites create fails with invalid rubric → use one of: exact-match, contains, llm-judge, regex
• suites run returns model not found → verify model ID with nerve models --search <name>
• eval get shows status: failed → check nerve logs --search eval for error details
• scorecard returns empty results → the run may still be running; poll nerve eval get <runId> until status is completed