Sources & editorial standards

Read a claim at the level its evidence supports.

LLMSX combines original project writing, agent-assisted research and third-party reference material. Those sources have different limits. A format score, a citation and a measured result are not interchangeable.

Original guides and research

An experiment should name its questions, environment, comparison and observed results. Readers should be able to inspect the underlying artifacts. A result from two repositories or one model does not establish how every system behaves. Negative results and failed setups matter as much as improvements.

An implementation or incident report describes a particular system and what happened there. An architecture paper explains a design; it does not demonstrate an effect size unless it includes a measurement. Publication dates identify when an article was published. A date in the body may identify an older implementation or dataset.

For an example with public methods and raw results, see the documentation cost experiment. For a design report that states its measurement gaps, see State That Survives the Session.

Agent-assisted references

Concept references organize material into summaries, facets and source-linked facts. Some were assembled with AI-assisted research and extraction. Citations can name upstream documents or archived reports. Archived reports are no longer hosted here. Check the primary evidence before using a claim.

A source link makes a claim inspectable. It does not establish that the source is authoritative, current or correctly represented. Check version-sensitive commands and technical limits against current primary documentation before applying them.

Source availability

Concept references retain their source citations. Archived research reports are no longer hosted here. Follow primary references before relying on a claim.

Incomplete work and corrections

Topic entries without a published reference remain browsable as part of the map. They are labelled as topic entries and excluded from Search. That exclusion is a content availability rule, not evidence that the remaining references passed a fact check.

If an article's public evidence does not support its measured-result framing, flag that gap prominently and exclude it from promotion and Search until the evidence is resolved. Do not silently turn an unsupported measurement into an estimate.

Report a correction with a page URL and the specific evidence through GitHub issues or the contact address. Keep the original context available when documenting a correction so readers can understand what changed.

AI assistance and reader judgment

AI assistance is part of this project, including research and writing. It does not replace source inspection or reproduce an experiment by itself. Treat examples as examples, architecture as architecture, and only supported observations as measurements.