Trust & provenance

LLM output is fallible. Core LLM Wiki records how each fact got into memory and how much to trust it, so your app can decide what to show and what to inject into prompts.

The signals on every fact

Field	Values	Meaning
confidence	certain · inferred · tentative	How sure the writer (model or user) was
source_type	immutable_document · librarian_inferred · user_stated · user_confirmed	Where the fact came from (details)
lifecycle_status	draft · stable · deprecated	Review state
source_ref / source_hash	document identifier and SHA-256 hash	Which document the fact was extracted from
okf_verified	list of { by, at }	Who verified it. human:* verifiers give the human-reviewed trust tier.
valid_from / valid_to	epoch ms	When the fact was true (temporal facts)

Drafts

A fact with lifecycle_status: 'draft' hasn’t been reviewed. Drafts are visible by default. Hide them per call or globally with excludeDrafts:

await wiki.read('user-1', 'deploy process', { excludeDrafts: true });

Promote a draft after review with promoteDraft(). See Review drafts & grounding.

Grounding

When grounding is on, chosen writers must quote the source text they were shown. A deterministic check (no LLM involved) verifies the quotes:

• All quotes found → the fact is stored stable with trust tier machine-confirmed.
• Quotes missing or not found → the fact is stored as a draft rather than rejected.

Grounding is not a prompt-injection defense

Grounding proves a fact quotes its source. It doesn’t prove the source is honest. Injected text in a document can be quoted like any other text. See Security.

Lint

lint(entityId) is a read-only health report. It counts dangling edges, manifest violations, untyped facts, drafts, and unverified inferences. It never repairs anything.