∴ Verifiable AI for high-stakes work
AI for the domains where a wrong answer actually matters.
Most AI is built to sound convincing. In tax, law, medicine, and hiring, sounding right isn't the job — being right is. We build systems that show their reasoning, cite their evidence, and admit when they don't know.
01/The Problem
Fluency arrived
before truth.
Language models were trained to produce text that reads as correct. That is not the same as being correct. A model will state a wrong threshold with the same confidence as a right one, cite a clause that was never in the contract, and never signal the difference.
In low-stakes work, that is a quirk you edit around. In tax, law, medicine, and hiring, it is the whole risk — and the person relying on the answer is rarely the person equipped to catch it.
SOUNDS RIGHT
“Yes — any sales into the state create nexus.”
✕ confident · flat · uncheckable
IS RIGHT
“Conditional. Above the economic-nexus threshold, yes; below it, no.”
✓ rule checked · assumption surfaced
02/What we hold to be true
Four things we refuse to pretend aren’t true.
Plausible is not provable.
A fluent explanation can still violate the rule it is quoting. The confidence in the tone tells you nothing about the correctness of the fact.
Rules are not prose.
Real domains run on dependencies, thresholds, exceptions, and definitions. Flatten them into a paragraph and the edge cases are exactly what gets lost.
A score is not a guarantee.
“92% confident” is not an audit trail. High-stakes decisions need reasons that hold up to scrutiny, not probabilities that merely sound reassuring.
Review that doesn’t scale isn’t automation.
If an expert has to re-check every answer by hand, the bottleneck didn’t disappear. It just moved — and the cost moved with it.
03/The Thesis
For high-stakes AI, the output isn’t the answer. The proof is.
A system worth trusting can show, for any answer it gives, four things. If it can't produce them, it hasn't earned the decision it's being asked to make.
RULE
the rule it applied
ASSUMPTIONS
the assumptions it made
CONSTRAINTS
the constraints it checked
FAILURE
the point it would break
04/The Approach
We turn domain knowledge into something a machine can check.
The point isn’t a smarter chatbot. It’s a pipeline that treats an answer as a claim to be verified, and refuses to ship one it can’t stand behind.
Capture
We encode the statutes, guidelines, and expert process of a domain as structured rules — not vibes, not a prompt that hopes for the best.
Formalize
Those rules become checkable logic, with the dependencies, thresholds, and exceptions made explicit instead of left implied.
Reason
The system generates candidate answers and tests each one against the constraints, the way an expert pressure-tests their own conclusion.
Evidence
You get the answer with its trace: sources cited, assumptions surfaced, and the gaps it couldn’t close flagged in plain sight.
05/Domains
Where being wrong is expensive.
We don’t chase general intelligence. We go domain by domain, into the places where the cost of a wrong answer is paid by a real person.
Tax & statutory reasoning
Taxability, nexus, exemptions, and credits — answers that have to survive an audit, not just sound authoritative.
ACTIVELegal & regulatory compliance
Obligations, thresholds, and exceptions read the way a regulator would read them, with the clause attached.
ACTIVEHealthcare & clinical safety
Guideline-grounded reasoning where a confident wrong answer carries a cost no probability can excuse.
RESEARCHSoftware, protocols & autonomous systems
Specifications and invariants that should be proven to hold — not assumed to, until they don’t.
RESEARCHHiring & talent
Where we started, and from the candidate’s side. The first home for this is Dossier.
SHIPPED∴ Work with us
The hardest problems aren’t unsolvable. They’re unformalized.
If your domain runs on rules that have to be right, we should talk. Bring the hard, high-stakes corner everyone else hand-waves — that is the part we want.