What we do not claim

We do not save you 79× tokens

The category’s number. Independent 2026-08 session bench: nobody hit 60%. JetBrains on rtk: advertised 60–90%, measured more expensive. Us: −0.2% Claude Code, +3.0% Codex. A thin mini harness did cut tokens (−36%, 4 questions). That is not “saves tokens.” What moves every time is 15 tool calls vs 43.

We are not the fastest query at kernel scale

CodeGraph answers in 0.79 s, ripgrep in 1.59 s, we in 3.70 s. They get the line (or a type name) sooner. We get the symbol. On gather we are ahead of those graphs (118 s vs 290 s vs 528 s; GitNexus did not finish). “Fastest code graph” is not a sentence we use.

We do not replace grep

Exact strings, comments, config values stay ripgrep. Locate questions: grep 47%, us 42%. The store is callers, blast, what this talks to.

The agent will not always call us

Claude often never looks at an MCP graph (Graphify 3/15, code-review-graph 0/15 on the same bench Codex used every time). We do not ship MCP. We do not block Grep. If it ignores the store, say this.

The store is not always fresh

Autosync default is off. A watcher you never invoke is not freshness. After you edit: ctx-optimize sync (no-change ~0.25 s). "autosync": "lazy" is a key, not the default.

We do not invent callers

A name declared more than once is AMBIGUOUS — a shortlist, filtered out of affected / change-plan / path unless you ask. Blast radius is a floor. Match eight times in mux returns the file, not a guessed caller.

We have not beaten Serena or potpie on a clock

No row. Empty stays empty. Serena’s LSP edges are more precise than our name-resolve. That is their column.

query is not a concept engine and not embeddings

“What does it shell out to” is boundaries, not a smarter embedding. Ranking sees label + path, not doc bodies. Neighbours are shown, not re-scored. The next retrieval ADR does not add a vector index.

Standing counter-weight in the repo: CRITIQUE.md.