Run evaluations

The arcade evals command discovers and executes evaluation suites with support for multiple providers, models, and output formats.

Basic usage

Run all evaluations in the current directory:

arcade evals .

The command searches for files starting with eval_ and ending with .py.

Show detailed results with critic feedback:

arcade evals . --details

Filter to show only failures:

arcade evals . --only-failed

Multi-provider support

Single provider with default model

Use OpenAI with default model (gpt-4o):

export OPENAI_API_KEY=sk-...
arcade evals .

$env:OPENAI_API_KEY="sk-..."
arcade evals .

Use Anthropic with default model (claude-sonnet-4-5-20250929):

export ANTHROPIC_API_KEY=sk-ant-...
arcade evals . --use-provider anthropic

$env:ANTHROPIC_API_KEY="sk-ant-..."
arcade evals . --use-provider anthropic

Specific models

Specify one or more models for a provider:

arcade evals . --use-provider openai:gpt-4o,gpt-4o-mini

Multiple providers

Compare performance across providers (repeat --use-provider):

arcade evals . \
  --use-provider openai:gpt-4o \
  --use-provider anthropic:claude-sonnet-4-5-20250929 \
  --api-key openai:sk-... \
  --api-key anthropic:sk-ant-...

When you specify multiple models, results show side-by-side comparisons.

API keys

API keys are resolved in the following order:

Examples:

# Single provider
arcade evals . --api-key openai:sk-...
 
# Multiple providers
arcade evals . \
  --api-key openai:sk-... \
  --api-key anthropic:sk-ant-...

Capture mode

Record tool calls without scoring to bootstrap test expectations:

arcade evals . --capture --output captures/baseline.json

Include conversation context in captured output:

arcade evals . --capture --include-context --output captures/detailed.json

Capture mode is useful for:

• Creating initial test expectations
• Debugging model behavior
• Understanding tool call patterns

See Capture mode for details.

Output formats

Save results to files

Specify output files with extensions - format is auto-detected:

# Single format
arcade evals . --output results.md
 
# Multiple formats
arcade evals . --output results.md --output results.html --output results.json
 
# All formats (no extension)
arcade evals . --output results

Available formats

Multi-run evaluation

Run each case multiple times to measure stability:

# Run each case 5 times with random seeds
# Pass if the majority of runs pass
arcade evals . \
  --num-runs 5 \
  --seed random \
  --multi-run-pass-rule majority \
  --details \
  -o results.html

When you use --num-runs > 1, Arcade adds per-case run_stats to results. If your case has critics, Arcade also adds critic_stats.

Valid values:

• --seed: constant (default), random, or a non-negative integer
• --multi-run-pass-rule: last (default), mean, or majority

What multi-run returns

In JSON output, each case includes multi-run data:

{
  "name": "Get weather for city",
  "status": "passed",
  "score": 93.0,
  "passed": true,
  "warning": false,
  "run_stats": {
    "num_runs": 3,
    "scores": [1.0, 0.8, 1.0],
    "mean_score": 0.93,
    "std_deviation": 0.09,
    "passed": [true, false, true],
    "warned": [false, true, false],
    "seed_policy": "random",
    "run_seeds": [123, 456, 789],
    "pass_rule": "majority",
    "runs": [
      {"score": 1.0, "passed": true, "warning": false, "failure_reason": null, "details": []},
      {"score": 0.8, "passed": false, "warning": true, "failure_reason": null, "details": []},
      {"score": 1.0, "passed": true, "warning": false, "failure_reason": null, "details": []}
    ]
  },
  "critic_stats": {
    "location": {
      "run_scores": [0.7, 0.56, 0.7],
      "weight": 0.7,
      "mean_score_normalized": 0.95,
      "std_deviation_normalized": 0.07,
      "mean_score": 0.67,
      "std_deviation": 0.05
    }
  }
}

Arcade uses --multi-run-pass-rule to set the overall status, passed, and warning fields. It sets run_stats.mean_score to the average raw score across runs, and the top-level score is that aggregate score in percent.

Command options

Quick reference

--use-provider, -p

Specify a provider and optional model list. Repeat the flag for multiple providers:

--use-provider <provider>[:<model1>,<model2>]

Supported providers:

• openai (default: gpt-4o)
• anthropic (default: claude-sonnet-4-5-20250929)

Examples:

# Default model for provider
arcade evals . -p anthropic
 
# Specific model
arcade evals . -p openai:gpt-4o-mini
 
# Multiple models from same provider
arcade evals . -p openai:gpt-4o,gpt-4o-mini
 
# Multiple providers (repeat `-p`)
arcade evals . -p openai:gpt-4o -p anthropic:claude-sonnet-4-5-20250929

--api-key, -k

Provide API keys explicitly (repeatable):

arcade evals . -k openai:sk-... -k anthropic:sk-ant-...

--capture

Enable capture mode to record tool calls without scoring:

arcade evals . --capture

--include-context

Include system messages and conversation history in output:

arcade evals . --include-context --output results.md

--output, -o

Specify one or more output files. Format is auto-detected from extension:

# Single format
arcade evals . -o results.md
 
# Multiple formats (repeat flag)
arcade evals . -o results.md -o results.html
 
# All formats (no extension)
arcade evals . -o results

--details, -d

Show detailed results including critic feedback:

arcade evals . --details

--only-failed, -f

Show only failed test cases:

arcade evals . --only-failed

--max-concurrent, -c

Set maximum concurrent evaluations:

arcade evals . --max-concurrent 10

Default is 1 concurrent evaluation.

--debug

Show debug information for troubleshooting:

arcade evals . --debug

Displays detailed error traces and connection information.

Understanding results

Results are formatted based on evaluation type (regular, multi-model, or comparative) and selected flags.

Summary format

Results show overall performance:

Summary -- Total: 5 -- Passed: 4 -- Failed: 1

How flags affect output:

• --details: Adds per-critic breakdown for each case
• --only-failed: Filters to show only failed cases (summary shows original totals)
• --include-context: Includes system messages and conversation history
• --num-runs: Adds per-run statistics and aggregate scores
• Multiple models: Switches to comparison table format
• Comparative tracks: Shows side-by-side track comparison

Case results

Each case displays status and score:

PASSED Get weather for city -- Score: 100.00%
FAILED Weather with invalid city -- Score: 65.00%

Detailed feedback

Use --details to see critic-level analysis:

Details:
  location:
    Match: False, Score: 0.00/0.70
    Expected: Seattle
    Actual: Seatle
  units:
    Match: True, Score: 0.30/0.30

Multi-model results

When using multiple models, results show comparison tables:

Case: Get weather for city
  Model: gpt-4o -- Score: 100.00% -- PASSED
  Model: gpt-4o-mini -- Score: 95.00% -- WARNED

Advanced usage

High concurrency for fast execution

Increase concurrent evaluations:

arcade evals . --max-concurrent 20

Save comprehensive results

Generate all formats with full details:

arcade evals . --details --include-context --output results

This creates:

• results.txt
• results.md
• results.html
• results.json

Troubleshooting

Missing dependencies

If you see ImportError: MCP SDK is required, install the full package:

pip install 'arcade-mcp[evals]'

For Anthropic support:

pip install anthropic

Tool name mismatches

Tool names are normalized (dots become underscores). Check your tool definitions if you see unexpected names.

API rate limits

Reduce --max-concurrent value:

arcade evals . --max-concurrent 2

No evaluation files found

Ensure your evaluation files:

• Start with eval_
• End with .py
• Contain functions decorated with @tool_eval()

Next steps

• Explore capture mode for recording tool calls
• Learn about comparative evaluations for comparing tool sources