Not tested (C5, V0, not a core claim): the README benchmarks (LOCOMO recall@10 0.497 / QA 45.3%, LongMemEval-S QA 76%, ERPNext key-fact coverage 82.0% vs grep 70.8%). They need the vendor's harness, an LLM to answer and an LLM judge; we had no paid LLM key and did not spend the small free daily quota on it. Graphify Cloud ("80% fewer tokens", 14-day trial) was out of scope.
Also not tested: the semantic pass over docs/PDFs/images/video (needs an LLM key or an AI assistant), the /graphify skill inside an assistant, MCP server, ~38 of the ~40 languages (we used Python and TypeScript).

Query log (README vs code): the README says every query/path/explain call "is logged to ~/.cache/graphify-queries.log". In 0.9.93 nothing was written there after our queries; graphify/querylog.py says logging is opt-in (GRAPHIFY_QUERY_LOG_ENABLE=1 or GRAPHIFY_QUERY_LOG=<path>), citing issue #1797. So the code is more private than the README says; the README is out of date.

Output files: the README says "You get three files" (graph.html, GRAPH_REPORT.md, graph.json). Headless `graphify extract` wrote graph.json (+ .graphify_analysis.json, manifest, cache) and told us to run `graphify cluster-only` for the report and HTML; after that all three were there. Without an LLM key cluster-only printed "no LLM backend configured; keeping Community N placeholders", but the report still named communities after files/classes (e.g. "Cart", "tax.py", "api.ts").