Tools
Inspection Report
Generate a structured PDF audit of any .ckf package directly from the Viewer. Two tiers: a heuristic Basic report that always works, and an Advanced LLM report that adds reasoning over consistency, composition-hallucination risk and source coverage.
Where to find it
Open /viewer, load any .ckf file (and optionally the original source). The Inspection Report card appears above the knowledge grid with two actions: Generate Basic Report and Generate Advanced Report (AI).
Basic report (heuristic, always works)
Pure client-side, schema-stable, no API key. Mirrors the philosophy of the heuristic compiler — it is the trustworthy default, with zero hallucination surface. Includes:
- Cover — title, domain, language, source file, generation timestamp.
- Package inventory — counts per section + atomic-unit type distribution.
- Quality card — human readability, AI utility, confidence distribution.
- Traceability — % of items with
source_refsorsource_excerpts, worst sections. - Source coverage — % of declared
source_excerptsthat literally appear in the linked source. - Compression diff — source bytes vs CKF bytes (only when source is linked).
- Executable rules catalog — if-then rules, decision rules, exceptions.
- Knowledge limits — declared assumptions, missing context, weakly-supported claims.
Advanced report (LLM, BYOK or admin Advanced AI)
Adds derived analysis on top of the heuristic baseline, using a single tool-call generation against your chosen provider. Each section's footer logs model, provider, tokens_in/out. Sections:
- Executive summary — 1-page synthesis of the compiled knowledge.
- Consistency findings — conflicts across principles, rules and exceptions.
- Composition-hallucination risk — flags rule combinations likely to produce contradictory or fabricated chains during retrieval.
- Coverage gaps — topics present in the source but missing from the CKF.
- Suggestions — actionable improvements per section.
Access
Admins and users in the compiler_allowlist can run the Advanced report on Lovable Advanced AI for free. Everyone else uses BYOK — keys live in request arguments only and are never stored or logged.
Why two tiers
The same rationale as the MCP ckf.compile / ckf.compile_llm split: heuristic is structurally consistent and always works; LLM upgrades the inference quality and reduces composition hallucination at the cost of requiring a provider. See Heuristic vs LLM and Schema changes inference for worked PDF-vs-CKF examples.
Privacy
- Basic report runs entirely in the browser — your
.ckfnever leaves the page. - Advanced report sends a trimmed digest (counts + sample IDs + up to ~8KB of source head) to your selected provider; nothing is persisted on the server.
- BYOK keys are passed through per-request and discarded.