Retrieval isn’t the sin. Chunk soup without receipts is.
Retrieval pipelines hand back disembodied fragments — no structure, no way to check them against the source, and a re-embed bill every time the corpus moves. Cartographer serves a map instead: deterministic lookups from the first millisecond, a gisted tree of record seconds later, and a trace to source lines under every row. We know which architecture wins because we bet against it, ran six pre-registered rounds, and published the receipts.
Index like a database
- Chunks lose the document's structure and hierarchy
- Answers arrive citation-free; you trust the pipeline
- Authors rewrite docs — frontmatter, summaries, retrieval contracts
- Embeddings to build, host, and rebuild when the corpus moves
- No account of what the index does and doesn't cover
Navigate like a map
- Zero-token floor in milliseconds: catalog + lexical + exact match
- Tree of record lands in ~100s at cutter prices — then it's resident forever
- Structure is the interface — TOC, tiers, trace ranges
- Every row is a pointer to source lines — trace and verify
- Documents stay untouched; the map is a derived sidecar
Glance. Zoom. Trace.
Three rungs, like a map at three altitudes. Try it — this demo navigates Cartographer's own repository, the same way the product navigates any resource.
The skeleton is the glance: structure, tiers, and trace ranges, extracted mechanically in milliseconds. Click any row to zoom — the gist is the map speaking. Then trace to read the source lines themselves.
Navigable at zero tokens. Grounded at seventeen cents.
The production posture, measured on a full 90K-token book (4,904 lines). A deterministic front answers from the first millisecond while the tree of record builds behind it; when the tree lands it takes over as the map, and the deterministic floor stays underneath it permanently.
Instant for the findable. About a hundred seconds for everything. One-time per resource — the tree is cached by source-sha, so every future session reads it free.
42°47′N · plate IIIPay as you navigate. Never pay twice.
A zero-token floor, from the first millisecond
Catalog, BM25, and exact match serve the gap while the tree builds — deterministic, free, and they routed 96% of the study's questions to verified answer regions. A model front cost fifty cents, took ten seconds, and grounded less.
buys: time-to-navigable, hours → milliseconds, at $0.00Build the tree of record at cutter prices
The full eager tree — gist wave, recursive tiers, book gist — costs seventeen cents on a cheap cutter and grounds 26 of 28 blind questions with two independent judges in exact agreement. Built once, resident for everyone after.
buys: the grounding championship, for pocket changeGround truth one trace away
Every row on the map is a pointer to source lines. The map routes; it never testifies. Verification is a fetch, not an act of faith.
buys: answers you can checkHonest coverage, non-negotiable
The coverage ledger says exactly what has been folded and what hasn't. The map never pretends to cover territory it hasn't surveyed.
buys: trust that survives an auditThe surface is the product. The verbs are the protocol.
Cartographer is the first product on kirigami — a frozen twelve-column record format and eight verbs, pinned at v1.0.0 as spec-of-record. Every Cartographer operation decomposes into those verbs, and every map it produces is validated in CI against the independent live kirigami oracle: drift caught by a judge that cannot itself drift, because it's frozen. Founding evidence: two isolated model instances, given only the verbs, derived this same architecture independently.
Dark tiles: the Cartographer MCP surface (working vocabulary). Light tiles: the kirigami verbs underneath every one of them.
40°01′N · plate VWe don't estimate where we can measure.
| Claim | Number | Provenance |
|---|---|---|
| Skeleton extraction, full book (4,904 lines, ~90K tokens) → 276 rows + 40 bins | 13.2 ms | Measured, skeleton pilot 2026-07-01 |
| Tree of record: grounded answers on 28 blind held-out questions, two independent judges in exact agreement | 26 / 28 | Measured, Round 2, 2026-07-02 |
| Full tree build per book, cheap cutter — gist wave + recursive tiers + book gist | $0.17 | Measured, Round 2 run-r |
| Time to full tree, parallelized (139.4s baseline → clean repeat runs at 95.7s and 103.8s) | ~100 s | Measured, perf pass 2026-07-03 |
| Zero-token BM25 floor: top-5 routes to a verified answer region — including two questions no model configuration found in six rounds | 27 / 28 | Measured probe vs. all graded runs (routing, retrospective) |
| Entire six-round study — every verdict pre-registered and applied as written | $12–13 | Measured, spend within caps throughout |
Honesty note, because it's the brand: Cartographer's founding thesis was the lazy map — and the study falsified it. Six pre-registered rounds, verdicts applied as written including the losing ones, a model grader with its defects registered, and a routing metric labeled as routing, not answering. The architecture on this page is the one that won on receipts, not the one we started with. Every number above traces to a committed validation artifact.
38°54′N · plate VIEverything is a map of the resource you point it at.
One globe, many projections. Books, transcripts, codebases, corpora — folded once, navigable forever, verifiable always.