Editorial method · version 1
Claims stay attached to their evidence.
This site ranks sources by proximity to the claim, dates every review, and separates vendor statements, independent methodology, community experience, and editorial interpretation.
01 · Source hierarchy
Closest credible source wins.
Official product and research material
Used for identity, release status, model IDs, pricing, availability, architecture, and vendor evaluation claims.
Independent methodology and measurements
Used to interpret benchmark design, known task defects, configuration differences, and reproducibility.
Practitioner discussion
Used for questions, failure modes, terminology, and early sentiment, never as the sole source for controlled facts.
Editorial interpretation
Explicitly labeled synthesis, such as the first workload a team should run before adoption.
02 · Claim labels
| Label | Meaning | What it cannot prove |
|---|---|---|
| Confirmed release fact | A directly observable identifier, date, price, or availability statement in current first-party material. | Long-term stability or product quality. |
| Meta says | A capability, training, benefit, or benchmark claim made by Meta. | Independent reproducibility. |
| Independent methodology | An external benchmark or documented measurement process. | Performance on a different agent, budget, environment, or repository. |
| Community | A dated anecdote or discussion thread. | Representativeness, causality, or benchmark quality. |
| Editorial interpretation | A clearly marked recommendation derived from the evidence. | A universal best choice. |
03 · Benchmark reading
Every score needs its configuration.
A benchmark comparison is accepted only when the benchmark version, model, agent or harness, effort setting, environment, run date, and cost basis are visible. Cross-version or model-only comparisons are not presented as an agent ranking.
Current limitation
This initial site snapshot does not publish a Muse Spark 1.2 comparison table. The release is less than a day old, and a stable, clearly versioned independent aggregate was not available during review.
04 · Corrections and updates
Change the claim, keep the trail.
When a source changes, the site updates the material claim, review date, and qualification. A vendor correction is linked to the corrected first-party source. A community report is not deleted merely because it conflicts with a launch claim; the conflict is described and weighted by provenance.
The current site has no named individual editor and does not imply one. Until a public issue tracker or contact route is published, the correction policy is documented here but inbound submissions are not available.