> ## Documentation Index
> Fetch the complete documentation index at: https://evalgate.com/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Create a scorer

> Create deterministic, code, composite, or model-backed scorers.

# Create a scorer

A scorer turns evaluation evidence into a score or decision. Start with the
cheapest deterministic check that captures the requirement, and add model-backed
judgment only where semantic interpretation earns its cost.

## Pick the right kind

| Kind | When to use it | Example |
| - | - | - |
| **Deterministic** | Exact, rule-based checks | Output contains a required identifier |
| **Code** | Logic too complex for a simple rule | Parse, normalize, and compare structured output |
| **Composite** | Combine several scorer signals | Weighted deterministic + judge evidence |
| **LLM judge** | Semantic judgment is required | Is the answer safe and on-topic? |

## Step 1 — Open Measurement → Scorers

Open `/measurement?view=scorers` and choose **New scorer**.

The old `/scorers` list route redirects into this Measurement view. Existing
scorer detail links such as `/scorers/{scorerId}` remain valid.

## Step 2 — Define the scorer

Choose the scorer kind and configure the behavior it owns. Keep model/provider,
prompt, thresholds, mappings, examples, and other decision inputs versioned
where the scorer type exposes them.

Code-backed execution is governed by the scorer runtime boundary; do not assume
your local Node or Python environment is identical to production execution.

## Step 3 — Add and review examples

Add labeled examples that represent both acceptable behavior and important
failures. Examples are measurement evidence, not a substitute for independent
qualification.

## Step 4 — Test and qualify

Run the scorer against its examples and inspect disagreements. Where the
workflow requires evaluator qualification, complete that evidence before using
the scorer as release authority.

## Step 5 — Approve, publish, and bind the exact version

Publishing/binding uses immutable scorer versions. A later edit creates new
evidence; it does not silently rewrite the version used by an earlier run.

## Up next

[Calibrate the scorer](https://evalgate.com/docs/help/scorers/calibrate-scorer.md) when scores need cross-version or
cross-model comparability.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.