Point Cartographer at a repo, a book, a corpus you've never seen. Three calls later you know what it is, how it's shaped, and what it can answer — without loading a single document. Ask your question, and the answer comes back at a line address you can re-fetch forever. No embeddings. No chunks. No model in the loop.
Retrieval pipelines — chunk, embed, top-k, as commonly deployed in 2026 — assume you already know your question. They're answer machines with no glance. Their strongest case is answering at scale, and we grant it. But the real problem with an unknown resource is that you don't know what's in it, and chunking it into an embedding index destroys the one thing that would tell you: its structure. Cartographer inverts this. The map comes first; questions come from the map.
Seven tool calls against a private research repository on a one-hour token. Zero model inference. Every number below is from the run.
27 documents, 4 states, 42,000 lines. A research knowledge base with a surface/archive pattern and its own retrieval contract. The coverage ledger reported 5 frontmatter parse failures — mapped anyway, at a lower rung, and it said so.
18 titles and their tag vocabulary: oral-theology, theodrama, communal-hermeneutics, phronesis. Now you know what to ask. The question isn't yours; the map handed it to you — before a single document body was read.
The question — how do oral communities validate theology without written texts? — routed to an 8,628-line thesis, into its chapter 7, to line 7053. The verbatim answer, at a citable address:
"The canon is the foundation of oral theology. The community is the locus of oral theology. Improvisation is the active participation in the theo-drama of the canon in community." L7053–L7058
Source: Stephen Stringer, Faith, Living, Understanding (thesis), archived in the corpus mapped above — quoted verbatim at its line address.
The map routes. The source testifies. Re-fetch the address and get the same bytes, every time.
Also measured the same day, same posture: a 6,662-document entity database legible in two calls · Moby-Dick and Pride & Prejudice (texts courtesy of Project Gutenberg, gutenberg.org) as raw flat text, no markup · a 728-document translation manual with YAML sidecar titles. Five corpus shapes, all oriented in ≤3 calls, $0.
Every document carries mechanical signal — frontmatter, links, headings, paths. Cartographer reads the highest rung each document offers and tells you which rung it stood on. Rich repos map rich; bare repos map honest. Nothing is rewritten, summarized, or interpreted to become navigable. The map is a derived sidecar; your documents never move.
S0 descriptors → S1 frontmatter →
S2 links → S3 headings → S4 paths
Degrades gracefully. Lies never. A flat file never claims a structural rung it lacks.
Markdown maps natively. Plain text just works — Moby-Dick's 21,937 raw lines mapped with an entity card and labeled regions, no markup needed; structure only makes the map prettier. Since this page was first drafted, the ladder grew twice — exactly the way it's supposed to: plain text now infers its own chapter spine from its whitespace convention (Moby-Dick's 135 chapters, banded), and .tsv/.csv rows joined the map as documents. A format with no adapter yet — HTML, reStructuredText, EPUB — gets a census-refusal, never an empty map: counted, named by extension, refused with the reason. New adapters land the way those two did, each reporting exactly what it read.
The glance, the search index, the skeletons — mechanical, milliseconds, $0 by construction.
Every row points at source lines. Checking an answer is a read, not an act of faith.
What was mapped, what wasn't, and why — by extension, by rung, on every answer.
Content-addressed to a commit. Re-consult to follow head; the same sha returns the same map with zero upstream calls.
Every other tool flattens your repository to "whatever main says right now." Cartographer treats the ref as a first-class coordinate. Name the ref to follow head. Name the sha to stay put. Probe with resolve_ref to decide which — one cheap call answers "has the world moved?" without rebuilding anything. Consult a branch and the answer carries its receipt: resolved: { ref, sha, observed_at }. You always know whether you're reading the present or a pin.
A review pinned to the exact commit under discussion. Same sha, same map, forever — and free after the first consult.
Track head; the map re-cuts only when head actually moves, and the receipt names the sha it resolved to.
Because maps pin to commits, point Cartographer at the sha from before the refactor and walk the repo as it was — same glance, same skeletons, same line addresses, against bytes that no longer exist on any branch. measured
Coming up the rail: map-to-map diff — two shas in, and the answer is which states grew, which documents moved, which headings appeared, mechanically, at $0. The changelog nobody wrote, derived from the structure that was always there. roadmap
Cartographer maps private repos the way a contractor enters a building: with a badge you issued, scoped to one room, expiring on its own. Pass a GitHub token on the call — a fine-grained PAT or a short-lived app token — and it's used for exactly one archive fetch. Not stored. Not logged. Not needed again: the map and its document archive are content-addressed under the commit, so every later glance, zoom, and open serves from the store without touching GitHub at all.
The demo above ran against a private repo on a one-hour token scoped to that single repository. The token expired. The map didn't.
one repo per grant · read-only default · expiry is the rotation
Round one: the lazy map lost to the eager gist tree — six pre-registered rounds, verdicts applied as written including the losing ones. The tree grounded 26 of 28 blind questions under two independent judges, built for $0.17 a book in ~100 seconds. Round two: the tree itself was demoted — the mechanical floor underneath it (skeleton in 13.2ms, a zero-token BM25 floor that routed 27 of 28 questions to verified answer regions) turned out to answer the comprehension problem alone. So what ships is the floor: deterministic, verbatim, coverage-honest. Inference is now an optional rung, not the architecture.
Those July numbers stay on this page as the receipts of both rounds — preserved, re-contextualized, never deleted. Every number above traces to a committed run artifact — the July receipts to the pre-registered study logs, the August runs to the kitchen journal of 2026-08-24 — and anything we couldn't measure is marked roadmap.
CARTOGRAPHER · private · seeded 2026-07-01 · overhauled 2026-08-24