Editorial method

Evidence before narrative.

This guide is designed to make uncertainty inspectable—not to make every row look equally settled.

Last reviewed: September 3, 2026

01 / Review process

Scope the claim

Break broad statements into architecture facts, experiment results, interpretations, and commercial-product claims.

Prefer primary evidence

Use proceedings, official papers, preprints, repositories, and official technical statements before summaries or news coverage.

Verify identity and version

Record title, authors, venue or repository, date, identifier, and the version reviewed.

Attach the caveat

Describe the setting, model, task, comparison, or uncertainty that limits each source’s reach.

Separate evidence classes

Do not turn reputable reporting into official confirmation, or a preprint into a peer-reviewed result.

Re-review material claims

Update the claims ledger and changelog when direct evidence changes a public conclusion.

02 / Claim labels

Status is part of the claim.

Supported: Directly established by the cited architecture or evidence.
Supported with limits: Established within a stated setting but unsafe to generalize without qualification.
Mixed evidence: Relevant studies or measurements do not point to one clean conclusion.
Unsupported: The available evidence does not establish the claim.
Open question: No adequate public answer was located.
Reported, unconfirmed: Published by a reputable outlet but not confirmed in a direct technical source.

03 / Review boundary

Cutoff and corrections

The current research review covers material located through September 3, 2026. “No source located” is a statement about this review, not proof that a source cannot exist.

Substantive corrections are logged and reviewed against direct evidence. See the corrections process or inspect the public changelog.

This is an independent educational reference, not a peer-reviewed publication or a statement from a model developer.