In development · honest pre-launch
Your words. Your relationships. Receipts for everything.
Cartographer maps your working relationships from your own correspondence — with consent, per source — and hands your AI assistant the context and the constraints before it drafts a word in your name. Every statement in its map keeps receipts — links to the real messages behind it — and a statement that can't show its receipts refuses to display. Nothing sends itself: that part stays yours.
The problem
AI drafting has a quiet tax: the audit.
Every generated draft is a small interrogation. Is that how I talk to her? Did he actually agree to that? Would I ever open with this? The assistant is confident; you are the one doing the checking, message after message, because it doesn't actually know these people — and it doesn't know what it doesn't know.
So the tool that was supposed to save your attention spends it. You end up proofreading a stranger's impression of your own relationships.
The inversion
Review should feel like recognition, not inspection.
When a draft is built only from things that verifiably happened — your phrasing, that thread from March, the commitment you actually made — checking it stops being detective work. You read it and think: yes, that's right. Not because you trust the machine, but because every part of the draft can show you where it came from.
That is the whole design goal. Not a model that sounds plausibly like you. A map of what is actually true between you and the people you write to, and a drafting process confined to it.
How it works
Four steps, one owner.
-
Map
With your consent, granted source by source, it reads your own correspondence — read-only — and builds a map of how you and your people actually correspond: who answers, what stalls, what was promised. Every statement it keeps carries receipts: links back to the evidence it came from.
-
Retrieve
When you're about to write to someone, it assembles that moment's context. Constraints come first — what not to touch — then what's known, receipts attached, then what it doesn't know: sensible defaults labeled as exactly that, and plain questions rather than guessed answers.
-
Draft inside constraints
Your assistant writes with that packet as the boundary: the facts and limits of your relationship with this person, not a general model's hunch about people like them.
-
You own the send
Nothing sends itself. There is no code path in the system that sends anything — drafts end in your hands, every time, by construction rather than by setting.
The trust doctrine
Rules, not vibes.
The trust rules aren't a policy page; they are enforced in code, with tests that fail when a rule is broken. Filters before weights. Behavioral language only — every stored sentence has to survive being read aloud to the person it describes. A binding list of things that are never scored. Provisional until validated. Texture, never on facts. The operator owns every send.
Where things stand
It isn't launched yet.
Cartographer is in development, and this site will never pretend otherwise — no invented user counts, no borrowed testimonials, no screenshots of data that doesn't exist. If the doctrine reads like something you've been waiting for, leave an email and you'll hear from us once, when there is something real to show.