# AI Pareto Explorer

> An auditable model-selection dataset and interactive Pareto-frontier explorer spanning roughly 75 AI benchmarks. It preserves exact model revisions, benchmark configurations, source dates and links, uncertainty, reported costs, execution metrics, and current token prices.

The site has no chatbot. The primary artifact is a versioned, validated JSON snapshot suitable for crawlers, agents, notebooks, and applications. Missing values mean the source did not publish the field; they are never imputed.

## Machine-readable resources

- [API index](https://pareto.ehrlich.dev/api/v1): Endpoint directory, snapshot time, schema version, and counts.
- [Complete JSON snapshot](https://pareto.ehrlich.dev/api/v1/snapshot): All benchmarks, exact models, observations, provenance, schema glossary, and embedded quality report.
- [Benchmarks](https://pareto.ehrlich.dev/api/v1/benchmarks): Benchmark catalog. Filter with `?id=deepswe`.
- [Models](https://pareto.ehrlich.dev/api/v1/models): Paginated model catalog. Filters: `q`, `org`, `open`, `offset`, `limit`.
- [Observations](https://pareto.ehrlich.dev/api/v1/observations): Paginated observations. Filters: `benchmark`, `model`, `offset`, `limit`.
- [Quality report](https://pareto.ehrlich.dev/api/v1/quality): Counts, field coverage, quarantines, conflicts, and validation status.
- [OpenAPI](https://pareto.ehrlich.dev/openapi.json): API contract.
- [Field and methodology guide](https://pareto.ehrlich.dev/llms-full.txt): Semantics, examples, and interpretation cautions.
- [Source repository](https://github.com/ehrlich-b/pareto): Pipeline and viewer source.

## Interpretation

- `scores` contains observations, not one canonical score per model. Multiple configurations are intentionally retained.
- `m` is an exact normalized model-revision id; `sm` is the source's original model id.
- Reported benchmark cost (`c`) retains the benchmark's denominator and must be interpreted with its benchmark `cost_label` and `cost_basis`.
- Token prices are current OpenRouter prices and are separate from historic benchmark execution cost.
- Pareto membership depends on the selected axes and filtered cohort. The browser computes it client-side.
- Benchmark data is attributed to Epoch AI's AI Benchmarking Hub under CC-BY 4.0. Additional provenance is embedded in the snapshot.