Multi-Dimension Evaluation Suite

This guide covers building multi-dimension evaluation pipelines — the actual way to compose multiple evaluations into a unified model comparison in umwelten.

Instead of a single monolithic evaluation, you run independent evaluations (each using EvalSuite), then combine their results with eval combine to produce leaderboards and narrative reports.

Overview

The workflow has three steps:

1. Run independent evaluations — each one tests a specific capability (reasoning, coding, instruction-following, etc.)
2. Define an EvalDimension[] suite config — maps evaluation names to labels, max scores, and score extractors
3. Combine with eval combine — loads results from disk, builds per-model scorecards, and generates reports

Step 1: Run Individual Evaluations

Each dimension is a standalone EvalSuite script. Run them independently:

dotenvx run -- pnpm tsx examples/model-showdown/reasoning/eval.ts --all
dotenvx run -- pnpm tsx examples/model-showdown/instruction/eval.ts --all
dotenvx run -- pnpm tsx examples/model-showdown/coding/eval.ts --all

Results are written to output/evaluations/{eval-name}/runs/{NNN}/ as JSON files — one per model per task.

Step 2: Define the Suite Config

Create a TypeScript file that exports an EvalDimension[]. Each dimension maps to one evaluation's output directory and tells the combiner how to extract scores.

import type { EvalDimension } from '../../src/evaluation/combine/types.js';

export const SHOWDOWN_SUITE: EvalDimension[] = [
  {
    evalName: 'model-showdown-reasoning',
    label: 'Reasoning',
    maxScore: 20,
    perTaskMaxScore: 5,
    extractScore: (r) => r.judge?.reasoning_quality ?? r.score ?? 0,
    hasResultsSubdir: true,
  },
  {
    evalName: 'model-showdown-knowledge',
    label: 'Knowledge',
    maxScore: 30,
    perTaskMaxScore: 1,
    extractScore: (r) => r.correct ? 1 : 0,
  },
  {
    evalName: 'model-showdown-instruction',
    label: 'Instruction',
    maxScore: 30,
    perTaskMaxScore: 5,
    extractScore: (r) => r.score ?? 0,
  },
  {
    evalName: 'model-showdown-coding',
    label: 'Coding',
    maxScore: 126,
    perTaskMaxScore: 7,
    extractScore: (r) => r.totalScore ?? r.score ?? 0,
    hasResultsSubdir: true,
  },
];

EvalDimension fields

Field	Description
evalName	Directory name under output/evaluations/
label	Human-readable name shown in reports
maxScore	Total max score for the full evaluation
perTaskMaxScore	Points added to denominator per result file (for partial runs)
extractScore	Function to pull a numeric score from each result JSON file
hasResultsSubdir	Whether task dirs contain a results/ subdirectory
runNumber	Pin to a specific run (default: latest)

Step 3: Combine and Report

# Console table (default)
dotenvx run -- pnpm run cli -- eval combine \
  --config examples/model-showdown/suite-config.ts

# Markdown report
dotenvx run -- pnpm run cli -- eval combine \
  --config examples/model-showdown/suite-config.ts \
  --format md

# Full narrative report with methodology and analysis
dotenvx run -- pnpm run cli -- eval combine \
  --config examples/model-showdown/suite-config.ts \
  --format narrative --output report.md

# Focus on a specific model
dotenvx run -- pnpm run cli -- eval combine \
  --config examples/model-showdown/suite-config.ts \
  --format md --focus gemini

What the Combiner Does

1. Loads results — scans output/evaluations/{evalName}/runs/{latest}/ for each dimension
2. Extracts scores — calls your extractScore function on each JSON result file
3. Builds scorecards — aggregates per-model scores across all dimensions into ModelScorecard objects
4. Calculates rankings — computes combined percentage (mean of dimension percentages), total cost, total duration
5. Generates reports — outputs leaderboard tables, cost efficiency analysis, speed comparisons, and per-dimension breakdowns

Report Formats

Format	Description
console	Structured tables to stdout
md / markdown	Markdown with leaderboard and dimension tables
json	Structured JSON for programmatic use
narrative	Full prose writeup with methodology, analysis, and judge explanations

Programmatic Usage

You can also use the combine system directly in code:

import { loadSuite } from '../../src/evaluation/combine/loader.js';
import { buildSuiteReport } from '../../src/evaluation/combine/report-builder.js';
import { buildNarrativeReport } from '../../src/evaluation/combine/narrative-report.js';
import type { EvalDimension } from '../../src/evaluation/combine/types.js';

const dimensions: EvalDimension[] = [ /* ... */ ];

const suite = loadSuite(dimensions);

// Structured report
const report = buildSuiteReport(suite);

// Narrative markdown
const narrative = buildNarrativeReport(suite);

Full Working Example

See examples/model-showdown/ for the complete working example including:

• suite-config.ts — 5-dimension suite config
• run-all.ts — orchestration script that runs all dimensions then combines
• Individual evaluations — reasoning/, knowledge/, instruction/, coding/, mcp-tool-use/
• shared/models.ts — shared model list

For a step-by-step walkthrough, see the Model Showdown walkthrough.

Next Steps

• Start with EvalSuite Examples to learn how to write individual evaluations
• See the Pairwise Ranking Example for Elo-based head-to-head comparisons
• Read the Model Evaluation guide for full API reference