# ModelDial

> ModelDial publishes verified first-party AI model Radar results. Use the compact JSON endpoints for current retrieval and the publication indexes plus linked snapshots for audits.

## Agent data

- [Latest compact Radar](https://modeldial.com/api/v1/radar/latest.json): Current published overall ranking with all three axis scores and weights, plus the backward-compatible backend ranking.
- [Agent configuration profiles](https://modeldial.com/api/v1/radar/agent-profile.json): Advisory balanced, quality, and value assignments for main-agent and worker-agent roles, derived from publisher decision tags.
- [Latest protocol-aware changes](https://modeldial.com/api/v1/radar/changes.json): Backend rank, score, elapsed-time, and recommendation changes between the latest publication and its nearest earlier protocol-compatible publication; this is not overall-ranking history, and reference-cost deltas require a matching pricing snapshot.
- [Backend & Testing publication index](https://modeldial.com/api/v1/radar/index.json): Backend batch IDs, publication times, protocol identity, hashes, and complete snapshot links.
- [Capability publication index](https://modeldial.com/data/benchmark-snapshots/index.json): Latest verified batch identity and archive link for each capability and the weighted overall ranking.
- [Complete Backend & Testing snapshot](https://modeldial.com/data/reference-snapshots/latest.json): Validated Backend & Testing publication with question-level and provenance fields.
- [OpenAPI document](https://modeldial.com/openapi-v1.json): Machine-readable schema for the compact endpoints.

## Reading rules

- Use overallRankings as the default current ranking. Use rankings only for the backend axis. Preserve publisher order in either array.
- The overall ranking combines the latest verified result from each capability. Its three source publications may have different batch IDs and publication times.
- Agent profiles project independently evaluated configurations across the latest published axis results. They are not paired-agent benchmarks and do not authorize automatic configuration changes.
- Treat changes as comparable only when the changes endpoint returns comparison.comparable=true. A single score drop is a signal, not proof that a provider changed a model.
- Explain a reference-cost delta only when comparison.costComparable=true; otherwise treat the cost delta as unavailable without discarding comparable rank, score, or elapsed-time changes.
- publishedAt is the verified publication time in UTC, not the evaluation start time.
- estimatedReferenceCostUsd is an estimate for the controlled batch, not a user bill.
- Detect overall updates by overallBatch.id and backend updates by batch.id. A five-minute polling interval is sufficient around scheduled publication windows.

## Human documentation

- [Agent guide](https://modeldial.com/agent)
- [ModelDial Skill](https://modeldial.com/modeldial-skill/SKILL.md)
- [Public Radar](https://modeldial.com/radar)
- [Evaluation method](https://modeldial.com/method)
- [Data license](https://modeldial.com/data-license)

The published Radar dataset is licensed under CC BY 4.0. Attribute ModelDial and identify the batch when practical.