Merit Metrics APIs

Decorator

@metric

@metric(
    fn: Callable | None = None,
    *,
    scope: Scope | str = Scope.SESSION,
)

Parameters:

Name	Type	Default	Description
`fn`	`Callable \| None`	`None`	Generator or async generator function
`scope`	`Scope \| str`	`Scope.SESSION`	Lifecycle scope: `"case"`, `"suite"`, or `"session"`

Returns: Decorated function registered as a metric resource Generator Requirements:

Must yield a Metric instance first (this gets injected)
Optionally yield a final calculated value (becomes MetricResult.value)
Can assert on metric properties after all data is collected

Example:

import merit
from merit import Metric, metrics

@merit.metric
def accuracy():
    """Track accuracy across all test cases."""
    metric = Metric()
    yield metric  # This gets injected into merits

    # After all merits run, compute final value
    final_accuracy = metric.mean
    yield final_accuracy  # Captured in MetricResult

    # Assert on aggregate quality
    assert metric.mean > 0.8, f"Accuracy too low: {metric.mean}"

@merit.metric(scope="suite")
def latency_ms():
    """Track latency within a suite."""
    metric = Metric()
    yield metric

    # Check 95th percentile
    assert metric.p95 < 500, f"P95 latency too high: {metric.p95}ms"

# Use in merit functions
def merit_test(accuracy: Metric, latency_ms: Metric):
    result = model.predict("input")

    with metrics(accuracy, latency_ms):
        assert result == "expected"

Classes

Metric

Thread-safe class for recording data and computing statistical properties. Attributes:

Name	Type	Description
`name`	`str \| None`	Metric name (auto-set by decorator)
`metadata`	`MetricMetadata`	Collection metadata (scope, contributors, timestamps)

Methods:

Method	Parameters	Description
`add_record`	`value: int \| float \| bool \| list[int \| float \| bool] \| tuple[int \| float \| bool, ...]`	Record one or more data points (numeric/bool only)

Statistical Properties: All properties are computed lazily and cached. Recording happens when you call add_record(...) (manually) or when you use with metrics(...) to record assertion pass/fail outcomes.

Property	Type	Description
`raw_values`	`list[int \| float \| bool]`	All recorded values
`len`	`int`	Number of values
`sum`	`float`	Sum of all values
`min`	`float`	Minimum value
`max`	`float`	Maximum value
`mean`	`float`	Arithmetic mean
`median`	`float`	Median (50th percentile)
`std`	`float`	Sample standard deviation
`variance`	`float`	Sample variance
`pstd`	`float`	Population standard deviation
`pvariance`	`float`	Population variance
`p25`	`float`	25th percentile
`p50`	`float`	50th percentile (median)
`p75`	`float`	75th percentile
`p90`	`float`	90th percentile
`p95`	`float`	95th percentile
`p99`	`float`	99th percentile
`percentiles`	`list[float]`	All 99 percentiles (p1 to p99)
`ci_90`	`tuple[float, float]`	90% confidence interval (lower, upper)
`ci_95`	`tuple[float, float]`	95% confidence interval
`ci_99`	`tuple[float, float]`	99% confidence interval
`counter`	`Counter[int \| float \| bool]`	Frequency count of each unique value
`distribution`	`dict[int \| float \| bool, float]`	Share of each unique value (0-1)

Example:

from merit import Metric, metric, metrics

@metric
def response_quality():
    metric = Metric()
    yield metric

    # Access statistical properties
    print(f"Mean quality: {metric.mean}")
    print(f"Std deviation: {metric.std}")
    print(f"95th percentile: {metric.p95}")
    print(f"95% CI: {metric.ci_95}")

    # Distribution analysis
    print(f"Pass rate: {metric.distribution.get(True, 0)}")
    print(f"Fail count: {metric.counter[False]}")

    # Quality gate
    assert metric.mean > 0.8

@merit.parametrize("input,expected", test_cases)
def merit_quality_check(input, expected, response_quality: Metric):
    result = model.predict(input)
    is_correct = result == expected

    # Record to metric
    with metrics(response_quality):
        assert is_correct

MetricMetadata

Metadata tracking metric lifecycle and contributors. Attributes:

Name	Type	Description
`last_item_recorded_at`	`datetime \| None`	Timestamp of most recent value
`first_item_recorded_at`	`datetime \| None`	Timestamp of first value
`scope`	`Scope`	Lifecycle scope (SESSION, SUITE, CASE)
`collected_from_merits`	`set[str]`	Names of merit functions that contributed
`collected_from_resources`	`set[str]`	Names of resources that contributed
`collected_from_cases`	`set[str]`	Case IDs that contributed

Example:

@metric
def my_metric():
    metric = Metric()
    yield metric

    # Inspect metadata
    meta = metric.metadata
    print(f"Scope: {meta.scope}")
    print(f"Contributing merits: {meta.collected_from_merits}")
    print(f"Case count: {len(meta.collected_from_cases)}")
    print(f"First recorded: {meta.first_item_recorded_at}")
    print(f"Last recorded: {meta.last_item_recorded_at}")

MetricResult

Result captured when a metric resource completes. Attributes:

Name	Type	Description
`name`	`str`	Metric name
`metadata`	`MetricMetadata`	Snapshot of metadata at completion
`assertion_results`	`list[AssertionResult]`	Assertions evaluated in metric teardown
`value`	`CalculatedValue`	Final yielded value (or `NaN` if none)

Note: MetricResult instances are automatically collected and included in merit run reports. Example:

@metric
def accuracy():
    metric = Metric()
    yield metric

    final_accuracy = metric.mean
    yield final_accuracy  # Becomes MetricResult.value

    assert metric.mean > 0.8  # Captured in assertion_results

# MetricResult is created automatically after metric completes
# and includes:
# - name: "accuracy"
# - value: final_accuracy (the second yielded value)
# - assertion_results: [assertion about metric.mean]
# - metadata: snapshot of metric collection info

Context Manager

metrics()

Attach metrics to assertions for automatic data collection. Signature:

@contextmanager
def metrics(*metrics: Metric) -> Iterator[None]

Parameters:

Name	Type	Description
`metrics`	`Metric`	Metrics to record assertion outcomes into

Returns: Context manager that captures assertion results Behavior:

When an assertion passes inside the context, records True to all metrics
When an assertion fails, records False to all metrics
Multiple assertions in one context each contribute a data point
Works with both standard assertions and predicate assertions

Example:

from merit import Metric, metric, metrics

@metric
def accuracy():
    metric = Metric()
    yield metric

    # After all tests: mean should be > 80%
    assert metric.mean > 0.8

@metric
def hallucination_rate():
    metric = Metric()
    yield metric

    # After all tests: less than 5% should be False
    false_rate = metric.counter[False] / metric.len if metric.len > 0 else 0
    assert false_rate < 0.05

@merit.parametrize("question,context", test_cases)
async def merit_bot_quality(question, context, bot, accuracy: Metric, hallucination_rate: Metric):
    answer = bot.ask(question, context=context)

    # Multiple metrics in one context
    with metrics(accuracy, hallucination_rate):
        # Each assertion records True/False to all metrics
        assert len(answer) > 0
        assert not await has_unsupported_facts(answer, context)

Usage Patterns

Basic Metric Collection

from merit import Metric, metric, metrics

@metric
def latency_ms():
    metric = Metric()
    yield metric

    # Quality gates
    assert metric.p95 < 500
    assert metric.mean < 200

def merit_performance(api_client, latency_ms: Metric):
    import time

    start = time.time()
    result = api_client.call()
    duration_ms = (time.time() - start) * 1000

    # Manual recording
    latency_ms.add_record(duration_ms)

    assert result.ok

Assertion-Based Collection

from merit import Metric, metric, metrics

@metric
def accuracy():
    metric = Metric()
    yield metric
    assert metric.mean > 0.8

@merit.parametrize("input,expected", test_cases)
def merit_classifier(input, expected, classifier, accuracy: Metric):
    result = classifier(input)

    # Automatic recording based on assertion outcome
    with metrics(accuracy):
        assert result == expected  # Records True/False

Multiple Metrics

from merit import Metric, metric, metrics

@metric
def precision():
    metric = Metric()
    yield metric
    assert metric.mean > 0.85

@metric
def recall():
    metric = Metric()
    yield metric
    assert metric.mean > 0.80

def merit_evaluation(model, precision: Metric, recall: Metric):
    predictions = model.predict(test_data)

    for pred, actual in zip(predictions, ground_truth):
        # Track precision
        if pred == "positive":
            with metrics(precision):
                assert pred == actual

        # Track recall
        if actual == "positive":
            with metrics(recall):
                assert pred == actual

Distribution Analysis

from merit import Metric, metric, metrics

SENTIMENT_TO_SCORE = {"negative": -1, "neutral": 0, "positive": 1}

@metric
def sentiment_score_distribution():
    metric = Metric()
    yield metric

    # Analyze distribution
    total = metric.len
    dist = metric.distribution

    print(f"Positive: {dist.get(1, 0) * 100:.1f}%")
    print(f"Neutral: {dist.get(0, 0) * 100:.1f}%")
    print(f"Negative: {dist.get(-1, 0) * 100:.1f}%")

    # Ensure balanced distribution
    for score in (-1, 0, 1):
        rate = dist.get(score, 0)
        assert 0.2 < rate < 0.5, f"score {score} rate out of range: {rate}"

@merit.parametrize("text", test_texts)
def merit_sentiment(text, classifier, sentiment_score_distribution: Metric):
    label = classifier.classify(text)  # e.g. "positive" | "neutral" | "negative"
    score = SENTIMENT_TO_SCORE[label]

    # Record numeric values
    sentiment_score_distribution.add_record(score)

Percentile Analysis

from merit import Metric, metric, metrics

@metric
def response_time():
    metric = Metric()
    yield metric

    # SLA checks
    assert metric.p50 < 100, f"Median too high: {metric.p50}ms"
    assert metric.p95 < 500, f"P95 too high: {metric.p95}ms"
    assert metric.p99 < 1000, f"P99 too high: {metric.p99}ms"

    # Final value for reporting
    yield metric.p95

@merit.repeat(100)
def merit_latency_test(api, response_time: Metric):
    import time

    start = time.time()
    api.call()
    duration_ms = (time.time() - start) * 1000

    response_time.add_record(duration_ms)

Confidence Intervals

from merit import Metric, metric, metrics

@metric
def success_rate():
    metric = Metric()
    yield metric

    # Check confidence interval
    lower, upper = metric.ci_95

    print(f"Success rate: {metric.mean:.2%}")
    print(f"95% CI: [{lower:.2%}, {upper:.2%}]")

    # Lower bound must exceed threshold
    assert lower > 0.75, f"95% CI lower bound too low: {lower}"

@merit.repeat(50, min_passes=40)
def merit_reliability(llm, success_rate: Metric):
    response = llm.generate("Hello")

    with metrics(success_rate):
        assert "hello" in response.lower()

Get Started

Usage

Concepts

API Reference

Examples

Decorator

@metric

Classes

Metric

MetricMetadata

MetricResult

Context Manager

metrics()

Usage Patterns

Basic Metric Collection

Assertion-Based Collection

Multiple Metrics

Distribution Analysis

Percentile Analysis

Confidence Intervals

​Decorator

​@metric

​Classes

​Metric

​MetricMetadata

​MetricResult

​Context Manager

​metrics()

​Usage Patterns

​Basic Metric Collection

​Assertion-Based Collection

​Multiple Metrics

​Distribution Analysis

​Percentile Analysis

​Confidence Intervals

Decorator

@metric

Classes

Metric

MetricMetadata

MetricResult

Context Manager

metrics()

Usage Patterns

Basic Metric Collection

Assertion-Based Collection

Multiple Metrics

Distribution Analysis

Percentile Analysis

Confidence Intervals