What's My AIHow the model catalog tracks provenance, runtime coverage, and refresh ownership.

guide

Every American local-model claim needs a policy trail before it can influence a recommendation.

WhatsMy.AI does not publish an American local-model page from memory alone. Each tracked model has to point back to an official release page, declare the evidence behind the origin claim, and keep runtime coverage tied to the exact paths shown on the site.

Policy snapshot

Current policy version: 2026-03-12-us-origin-v2.

Every catalog entry must ship with a source URL, release date, license, modality, runtime availability map, and last-verified timestamp.

The release date is normalized to the precision the source actually gives us, usually a month and year, so the catalog does not invent a day that the source never published.

US-origin eligibility requirements

Entity provenance. A model only qualifies as an American local model when the stewarding release entity can be traced to an official US-controlled company or institute page. Stored at provenance.originEvidence.entityProvenanceUrl.

Official release channel. The canonical source URL must resolve to an official provider-controlled release page for the exact model line, not a community mirror or repackaging page. Stored at provenance.originEvidence.officialReleaseChannelUrl.

Downloadable local-runtime path. Every American entry needs at least one current downloadable local-runtime path with a reviewed install or artifact URL. Guided setup docs can supplement that evidence but cannot replace it. Stored at runtimeSupport[*].actionUrl.

License requirements. The displayed license label must trace back to an official page that states the governing terms for the release being recommended. Stored at provenance.originEvidence.licenseUrl.

Reviewer timestamp. Each entry carries a per-entry review timestamp. CI treats that timestamp as the reviewer stamp for freshness and policy enforcement. Stored at provenance.lastVerifiedAt.

Recorded evidence fields

`provenance.originEvidence.entityProvenanceUrl` records the official US-entity evidence for the stewarding organization.

`provenance.originEvidence.officialReleaseChannelUrl` must mirror the catalog source URL for the exact release page.

`provenance.originEvidence.licenseUrl` links the license terms the catalog is relying on.

`runtimeSupport[*].actionUrl` stores the reviewed download or install path for each tracked local runtime, and at least one US runtime path must be directly downloadable.

`runtimeSupport[*].verification.lastVerifiedAt` records when that runtime path was last rechecked.

`runtimeSupport[*].verification.checks[*].target` records whether the evidence is a compatibility review or a distribution review.

`runtimeSupport[*].verification.checks[*].sourceUrl` points to the exact reviewed runtime page, and `caveat` stores any constraints that need to stay visible.

When a runtime page is a catalog or guide rather than the raw artifact, the drift check also derives the tracked download path from the reviewed runtime command.

`lastVerifiedAt` is the reviewer timestamp carried by each entry and reused by the freshness window.

Verification gate

`npm run verify:model-catalog` is the focused metadata gate for this catalog.

`npm run codex:verify` runs the same provenance check before the broader test and build flow, so missing metadata or mismatched runtime availability fails the project validation path.

The CI gate now blocks any recommended entry that is missing entity evidence, official release evidence, license evidence, or a current downloadable local-runtime path.

The scheduled drift workflow re-checks official release URLs, license evidence, download paths, and runtime availability pages against live provider/runtime sources.

When a field changes, the workflow writes a public downgrade status, uploads a diff report, and opens a single actionable GitHub failure until a reviewer re-verifies the entry.

Recommendation graduation criteria

Public evidence bundle. A family only graduates into the visible recommendation track when the provenance guide links its public evidence bundle and decision history. Threshold: Required for every published family

Downloadable runtime path. At least one tracked runtime path must stay directly downloadable and reviewable so the recommendation does not lean on generic setup copy alone. Threshold: At least 1 current downloadable local-runtime path

Memory fit threshold. Recommendation pages only graduate when the benchmark and catalog can point to a concrete minimum and comfortable memory band for the family. Threshold: Tier ceilings published in the provenance guide

Freshness SLA. Reviewer timestamps must stay within the active review window so stale origin or runtime evidence drops out before it can influence the shortlist. Threshold: 30-day review cadence

Freshness is enforced with the same 30-day SLA that drives the catalog review window.

Tier memory ceilings

3B. Minimum memory must stay at or below 2.5 GB and comfortable memory must stay at or below 4.0 GB.

7B. Minimum memory must stay at or below 6.5 GB and comfortable memory must stay at or below 8.0 GB.

13B. Minimum memory must stay at or below 11.0 GB and comfortable memory must stay at or below 14.0 GB.

34B. Minimum memory must stay at or below 19.5 GB and comfortable memory must stay at or below 24.0 GB.

70B. Minimum memory must stay at or below 40.0 GB and comfortable memory must stay at or below 48.0 GB.

120B. Minimum memory must stay at or below 67.0 GB and comfortable memory must stay at or below 80.0 GB.

frontier. Minimum memory must stay at or below 245.0 GB and comfortable memory must stay at or below 288.0 GB.

Ownership and cadence

Owner: What'sMy.AI benchmark maintainers.

Refresh cadence: every 30 days, or sooner when a source page, runtime package, or license changes.

An entry is stale when its `lastVerifiedAt` timestamp falls outside the review window.

An entry is drifted when the scheduled check records a live source, license, download path, or runtime evidence mismatch in the checked-in drift status file.

What gets checked

The verification check confirms the source URL and runtime action URLs are HTTPS, the release date and last-verified timestamp are parseable, and the display labels still match the source metadata block.

It also requires the runtime availability map to stay in lockstep with the tracked Ollama, LM Studio, and llama.cpp paths so the public copy cannot drift away from the actual evidence.

Each tracked runtime path must declare whether the reviewed evidence was compatibility or distribution, plus the exact source URL and any recorded caveat.

For `origin: US` entries, the source URL must match the recorded official release channel and at least one runtime path must be a reviewed downloadable path rather than a generic setup note.

Model-family decision changelog

This public log uses the same format for additions, removals, upgrades, downgrades, and reclassifications, including the compact-model refresh that replaced older starter defaults.

Upgraded2026-03-12

Registry operations

Added scheduled drift checks, field-level diff alerts, and automatic public downgrade whenever a tracked source, license, or runtime path drifts from the verified snapshot.

Last verified by Codex on SYM-38 on 2026-03-12. Public reviewer labels stay at the workspace or team level so private contact details never appear here.

gpt-oss evidence bundle · Llama frontier line evidence bundle · Official release channel rule · License requirements rule · Downloadable local-runtime path rule · Community GGUF and runtime repackaging edge case

Added2026-03-12

Granite 4.0

Added Granite 4.0 to the public American-model shortlist after IBM-controlled release evidence, Apache terms, and direct local-runtime paths were reviewed together.

Last verified by Codex on SYM-39 on 2026-03-12. Public reviewer labels stay at the workspace or team level so private contact details never appear here.

Granite 4.0 evidence bundle · Entity provenance rule · Downloadable local-runtime path rule

Upgraded2026-03-12

OLMo 3

Upgraded OLMo 3 into the strongest fully open small and 32B-class branch once the Ai2 release pages and Apache licensing evidence were aligned with maintained runtime coverage.

Last verified by Codex on SYM-39 on 2026-03-12. Public reviewer labels stay at the workspace or team level so private contact details never appear here.

OLMo 3 evidence bundle · Official release channel rule · License requirements rule

Upgraded2026-03-12

Phi-4

Promoted Phi-4-reasoning to the clearest 13B-class text-first recommendation because its official Microsoft release and direct runtime path are both legible and current.

Last verified by Codex on SYM-39 on 2026-03-12. Public reviewer labels stay at the workspace or team level so private contact details never appear here.

Phi-4 evidence bundle · Official release channel rule · Downloadable local-runtime path rule

Added2026-03-12

gpt-oss

Added gpt-oss at the 20B and 120B tiers once the family had official release evidence, public license terms, and downloadable local-runtime packaging for the catalog.

Last verified by Codex on SYM-39 on 2026-03-12. Public reviewer labels stay at the workspace or team level so private contact details never appear here.

gpt-oss evidence bundle · Official release channel rule · License requirements rule

Reviewer history

Every public decision entry carries a reviewer label and date. The reviewer history below summarizes the latest public verifier for each currently tracked family without exposing private contact data.

reviewer history

Granite 4.0

Tracked releases: Granite 4.0 Micro, Granite 4.0 H-Small.

Latest public entry: Added on 2026-03-12. Last verified by Codex on SYM-39 on 2026-03-12.

reviewer history

OLMo 3

Tracked releases: OLMo 3 Instruct 7B, OLMo 3.1 Instruct 32B.

Latest public entry: Upgraded on 2026-03-12. Last verified by Codex on SYM-39 on 2026-03-12.

reviewer history

Phi-4

Tracked releases: Phi-4-reasoning.

Latest public entry: Upgraded on 2026-03-12. Last verified by Codex on SYM-39 on 2026-03-12.

reviewer history

gpt-oss

Tracked releases: gpt-oss-20b, gpt-oss-120b.

Latest public entry: Upgraded on 2026-03-12. Last verified by Codex on SYM-38 on 2026-03-12.

reviewer history

Gemma 3

Tracked releases: Gemma 3 12B.

Latest public entry: Reclassified on 2026-03-12. Last verified by Codex on SYM-39 on 2026-03-12.

reviewer history

Llama frontier line

Tracked releases: Llama 3.1 8B, Llama 3.3 70B, Llama 4 Scout, Llama 3.1 405B, Llama 4 Maverick.

Latest public entry: Upgraded on 2026-03-12. Last verified by Codex on SYM-38 on 2026-03-12.

Evidence bundles

Each bundle below rolls up the release pages, entity proof, license terms, and at least one downloadable runtime path that the decision changelog points back to.

evidence bundle

Phi-4

Phi-4-reasoning remains the text-first 13B-class Microsoft option because its official release, license terms, and direct Ollama path are all public and reviewable.

Phi-4-reasoning. Official release · Ollama runtime path

evidence bundle

Gemma 3

Gemma 3 remains in the American shortlist, but its public bundle keeps the runtime caveat visible because only the Ollama path is directly downloadable today.

Gemma 3 12B. Official release · Ollama runtime path

Policy rule changelog

2026-03-12 · 2026-03-12-us-origin-v2. Published the US-origin eligibility policy, added machine-readable evidence URLs plus per-runtime verification metadata to the catalog, and made the verification gate fail when a recommended entry is missing policy evidence or a current downloadable runtime path.

2026-03-11 · 2026-03-11-american-local-v2. Standardized source-normalized release dates, runtime availability parity checks, and the 30-day provenance review cadence for the catalog.

Edge-case decisions

Official Hugging Face org pages. Counts as the official release channel when the provider controls the org page directly. Several US-origin labs publish their weights through a verified Hugging Face organization. That page is accepted as the release evidence only when it is clearly owned by the stewarding entity.

Community GGUF and runtime repackaging. Counts for runtime evidence after review, but never for release-channel evidence. Community GGUF builds can back LM Studio or llama.cpp coverage once a reviewer verifies the path. They cannot be used as the official release source for origin eligibility.

Guided runtime docs without a download path. Not sufficient on their own. A generic setup guide can explain compatibility, but a US-origin entry still needs at least one downloadable local-runtime path before it can be published.

Distributed contributors and global hosting. Origin follows the stewarding release entity, not every contributor or CDN endpoint. The policy classifies the entry by the organization that officially releases and licenses the model line. Mirrors, contributors, or inference hosts do not change the catalog origin on their own.