# Metkagram Intent-to-Pattern Retrieval Benchmark

This is a public editorial regression benchmark, not an independent model leaderboard or a language-learning efficacy study.

## Task

Given only the natural-language query, return a ranked intent, one reasoning move, and up to three canonical Metkagram pattern IDs.

## Metrics

- intent top-1 accuracy
- move top-1 accuracy
- pattern hit@3 against the editorially acceptable pattern set

## Current public suite

- cases: 54
- locales: en=36, ru=18
- dataset version: 1.0.0+5b4307882dc4

## Reproducibility

Report the benchmark version, run date, system/model version, prompt or retrieval configuration, and whether the evaluated system had access to the Metkagram taxonomy/API/corpus. Report all three metrics and misses or abstentions.

The bundled deterministic resolver currently reports: intent top-1 100%, move top-1 100%, pattern hit@3 100%. This is an internal consistency signal because the benchmark and resolver are curated inside the same project.

Canonical benchmark JSON: https://metkagram.github.io/data/reasoning-benchmark.json

Baseline report: https://metkagram.github.io/data/reasoning-evaluation.json

Current terms: https://metkagram.github.io/en/licensing/