Doctrine
Why it works this way
Mintmark is a bet that the interesting question about AI in personal finance is not “what can the model do?” but “what should the model be structurally unable to do?” Every principle below is a constraint, and every constraint is enforced by code, not by good intentions.
Zero invention
No plugs, no estimates, no inferred dates. Nothing enters the book without a real origin — a document, a feed, or your explicit say-so. When the system meets a figure it cannot source, the correct output is a question — and a question that changes nothing is a fully successful result.
This is the load-bearing rule: a book is only useful if you can stop double-checking it, and one invented number taxes every other number forever.
Nothing settles without the owner
Everything that enters the book enters pending. Settling is the owner's action, not the system's. The structure makes overreach impossible, not merely discouraged.
Documents are the authority
Feeds are fast and occasionally wrong. Statements are slow and legally answerable. The system uses both and confuses neither: feeds keep the book current; documents give it truth. When the two describe the same event, the document wins.
Derive, don't accumulate
The book is a derivation from a handful of human-legible files. “Just fix the number by hand this once” is not merely discouraged — it is useless, because the next derivation erases the fix. Reproducibility is the shape of the thing, not a feature bolted on.
Two questions, kept apart
Is a change invalid, or is the book merely behind? Most systems collapse these into one bucket. Mintmark keeps them separate: invalid is caught at the gate and refused; behind rides beside the book as a verdict, never a veto.
Collapse them and the system rots: either everything blocks and the operator learns to override, or nothing blocks and the operator learns to ignore.
Code decides; the model narrates
The agent cannot approve its own plan. The model's judgment is useful — reading a statement, proposing a rule, answering a question — but never the thing that makes a change true.
It learns you
Most systems that learn are black boxes: the model improves — or drifts — and nobody can say exactly why. Mintmark inverts this. Every learned behavior has a named, versioned form — a rule whose provenance names the settled transactions that birthed it. Guesses decay into heuristics; heuristics crystallize into deterministic rules; a mature book runs mostly on rules, with the model left to handle only novelty. Every rule is born pending and approved by you. Learning that cannot be audited is just invention with a delay.
Over time the volume of pending work tracks how busy your money is, not how hard you fight it: batches arrive pre-tagged, citing the rules you settled into existence, and a clean batch clears in one tap.
Just say it
“Hey, btw — I took $60 out of an ATM to cover a $100 cash Facebook Marketplace furniture purchase.” The agent reads that, stages a plan — the withdrawal, the cash purchase, the tags — and summarizes it back to you. You confirm. Everything lands pending through the same action API the screens use. Nothing settles without you.
The agent translates casual language into structured actions, shows you exactly what it intends, and waits. A staged plan cannot confirm itself. What you get is natural-language access to every action the system has, with the same settlement gate standing between a proposal and the book.
Memory you can audit
The system's memory is data in a git repository, not a hidden vector store. Rules, rulings, resolved docket items — versioned, each with an author and a date. Unauditable memory is just another way to invent numbers.
Boring on purpose
Under the AI, Mintmark is aggressively conventional: text files, git, SQLite, one derivation path, invariants checked exhaustively. The novel part — an agent that reads your statements and drafts the code to parse them — earns trust only if everything beneath it is legible enough to check by hand. The exotic sits on the boring, never the other way around.