Agent model evidence, not leaderboards.

This guide starts from 31 Leading and Notable model records. Those tiers reflect general catalog curation—not demonstrated tool calling, autonomous task success or production readiness.

1. Selection universe

The curated selection contains 12 Leading and 19 Notable entries in data/model-curation.json. Curated status is about the existing registry, not a new agent benchmark. Media generators, retrieval encoders and decision scorers are labeled as components rather than autonomous agent brains.

2. Distinct evidence levels

Repository/config/license evidence establishes particular artifact fields. Publisher statements remain attributed. Independent benchmark results require exact task set, model revision and evaluation method. Open Model Weights measured results exist only when an actual reproducible OMW run is recorded. Everything else remains unknown.

3. Tool-call and system readiness

Neither a chat template nor a runtime's tool API proves correct tool selection. Tool-use reliability, structured argument fidelity, long-horizon completion and computer-use success require model-plus-runtime-plus-agent-system tests. We do not manufacture boolean support or global readiness scores.

4. Memory and licenses

Weight-only decimal GB is a formula based on total parameter count and precision. It is not observed VRAM, achievable quantization or a published checkpoint. Commercial-use classifications are not legal opinions; custom terms and territorial restrictions must be read from original licenses.

5. Why results are not ranked

The Explorer filters documented metadata and category-derived evaluation candidates. It does not sort by agent quality. Two models with different tokenizers, prompts, serving parsers or benchmark versions must not be declared comparable merely because a score exists.

6. Freshness and correction

The underlying registry receives source-linked verification updates. Unchanged source checks are not new article modifications. Any newly verified agent capability needs a field-specific source, revision, observation date, scope and review before publication.

Evidence taxonomy

Primary artifact

Weights, configuration, model card and license establish specific documented fields.

Publisher claim

Attributable to the publisher, never silently presented as an OMW result.

Independent evaluation

Versioned external benchmark evidence; task, hardware and evaluator must match.

OMW measured / Unknown

OMW measured only with retained reproducible evidence; otherwise explicitly unknown.

COMMON QUESTIONS

Frequently asked questions

Does Leading mean agent-ready?

No. Leading and Notable are general catalog-curation tiers; agent-specific reliability is unverified unless separately evidenced.

Are missing tool-support facts negative results?

No. Unknown is not false, and unsupported claims are not filled from model-family expectations.

READ THE ORIGINALS

Documentation and primary sources

Source-specific capability evidence

The agent evidence ledger attributes each numeric result to its publisher, records evaluation conditions and keeps missing benchmark revisions explicit. The model selector never converts an absent entry into a negative claim. A local harness can check tool-call formatting but does not independently reproduce BFCL or SWE-bench scores.

Evaluate before deployment.

Model metadata is source-linked; agent task success, tool-call reliability and production readiness require separate tests.

Review the methodology → Return to the explorer →