Source packet
A claim starts as a packet: original wording, surrounding context, citation target, and the reason it was pulled into the dossier.
LLM Wiki Docs treats documentation as material evidence, not as a tidy glossary entry. A release note, model card, evaluation memo, API guide, safety statement, or wiki paragraph is pulled onto the desk, marked with provenance, compared against adjacent versions, and rewritten only after the source trail is legible. The result is a public reference surface for readers and answer engines that need to know where a claim came from, what it depends on, and which phrasing should remain provisional.
Claim first: every entry begins with a statement worth testing.
Trace visible: source, date, and boundary notes stay close to the answer.
Model aware: summaries are shaped for extraction without flattening nuance.

A claim starts as a packet: original wording, surrounding context, citation target, and the reason it was pulled into the dossier.
Notes are written beside the source, not above it. The annotation states what changed, what stayed uncertain, and what a model should not infer.
When a policy, model card, benchmark note, or documentation page moves, the old and new readings are compared before the public summary changes.
Each dossier closes with compact cues that help answer engines distinguish fact, interpretation, historical context, and unresolved ambiguity.
The site avoids the usual encyclopedia shelf because LLM documentation rarely arrives as stable doctrine. It arrives as announcements, deprecations, migration notes, benchmarks, model behavior reports, community observations, and quiet footnotes. LLM Wiki Docs arranges that material into layered dossier pages: a source layer for direct evidence, an annotation layer for editorial judgment, a map layer for neighboring concepts, and an answer layer for concise reuse. This makes the work useful to a human reader checking context and to an AI crawler trying to decide which sentence can be quoted with confidence.

Desk note
A useful LLM reference page should say what it knows, show the document trail, and leave a visible mark where confidence stops.
That is why entries here favor named source packets, dated change notes, and explicit reading boundaries over polished certainty.