Knowledge Distiller — AI skill for Claude Code

Knowledge Distiller v4.0

Extract knowledge, not text — and keep the evidence attached.

Knowledge Distiller is a local, file-based knowledge-graph toolchain and compiler contract. Its canonical `.knowledge.json` graph separates concepts, claims, relationships, facts, source evidence, time, spatial roles and derivations. Deterministic tools validate, merge and render that graph without requiring a daemon or database and without outbound network access or telemetry. An authenticated loopback API is an optional facade, not a runtime prerequisite.

The repository currently implements bounded source normalization and deterministic post-compilation validation, merge, evaluation, rendering and run receipts. It does **not** contain a Segment-to-Graph semantic compiler: an external agent/compiler must author the evidence-linked graph first. The extractor and post-compilation runner are separate tools; there is no automatic end-to-end handoff and the runner neither invokes a model nor applies a compiler profile.

Spec 1.1 is an additive trust layer: existing Spec 1.0 files remain valid.

Why this version is different

Weak shortcut	Implemented contract
Plausible statement with a source name	Reusable evidence with an exact selector/locator
One opaque “quality” number	conformance_score plus scoped golden evaluation; semantic truth is never implied
“The AI reasoned that …”	Explicit origin, evidence, short derivation summary and review status; no chain-of-thought storage
A place string with unclear meaning	Role-qualified spatial context such as jurisdiction, market or mentioned location
Conflicting value overwritten	Both facts retained and linked through an explicit conflict
Count-only merge guard	Identity- and payload-aware merge with exact prior archive plus mandatory JSON and Markdown diffs
CDN-backed viewer	One dependency-free offline HTML file with an interactive SVG graph
IDs interpolated into paths	Pre