A tax answer that cannot be proved should not be given.
That sentence is the product. Everything else here — the deterministic calculators, the citations, the coverage matrix with its gaps on it — is what you have to build once you decide to take it seriously.
The failure worth designing against is confidence, not ignorance.
Generic AI tools confidently invent tax rates, cite laws that do not exist, and cannot tell you which tax year they used. The problem is not that it sometimes does not know. It is that it does not know that it does not know, and the output looks identical either way.
For most questions that is an inconvenience. For a tax position it is somebody filing on a number nobody checked, finding out two years later, and having no way to reconstruct which law the answer was based on. Everything in this product follows from treating that as the failure to engineer against.
What that costs
A system that refuses is harder to demonstrate, harder to sell, and scores worse on any benchmark that rewards answering. It also cannot be built by wrapping a model in a prompt — it needs calculators, a corpus with provenance, and an evaluation harness that is a second project the first one depends on.
That is the trade, stated plainly. If a system that always answers is what you want, this is not it — and the coverage page will tell you where this one stops before you spend an hour finding out.
Exact inside declared coverage. Outside it, the honest answer is that there is not one, and the product says so rather than producing something that looks like one.
Five positions, each one checkable on another page.
None of these is a value. Every one is a property of the system that somebody can go and verify, which is the only kind worth writing down — a principle nobody can check is a slogan.
- Check it
Exact, or nothing at all
Most systems in this category are built to always produce an answer, because an answer looks like value. That is the wrong objective for tax. A figure that is confidently wrong costs more than no figure, because somebody files on it — so the certainty gate can refuse, and refusing is a successful outcome rather than a failure to be engineered away.
- Check it
Numbers come from calculators, never from a model
A language model that produces a tax figure is guessing at arithmetic, and it will be right often enough to be trusted and wrong often enough to matter. Every number here is computed by a deterministic engine from a versioned parameter dataset, and the model's job is to decide which engine and explain what it did.
- Check it
Every legal claim points at a real document
Not a plausible-sounding section number: a specific locator in a real official source, checked back against the evidence one claim at a time. An unsupported material claim blocks the answer rather than shipping with a caveat nobody reads.
- Check it
A person approves anything that matters
The product does not file returns, does not move money, and does not post to a ledger on its own. Those are not missing features — they are the line between decision support and a system that can be wrong on your behalf without anybody noticing until it has been.
- Check it
The gaps are published beside the coverage
A coverage map that lists only what works is marketing. Ours lists 84 jurisdictions with what each one can and cannot do, including the ones where the honest answer is that there is a dataset and no calculator behind it.
Two sections you would expect, and why they are missing
There is no team grid, because there is no team to introduce yet. A row of grey circles under a “meet the team” heading is a claim about the size of a company, and it is the first one a buyer checks.
There is no founding timeline either. Every dated event we could put on one carries the same date — the day the documentation became public — and turning one event into four by inventing quarters would be manufacturing a history on a page arguing for provable claims.
Both appear here when there is something real to put in them.
Two routes, if you want a person rather than a page.
A question costs nothing and gets a real answer.
Five routes, each with the address it reaches and how long a reply takes, published before you write anything.