Founders who can state the claim
A clear account of what is asserted, what is still a candidate, and who is accountable for the record. PASS is a candidate. Verified is a kernel judgement. We look for people who keep those words honest.
Venture capital for founders building verifiable AI
AK Venture Corp (AKVC) is a venture firm. We invest in companies whose AI claims can be checked: statistical assessment for what is still uncertain, and independently machine-checked proof where a claim can be stated precisely. A fluent model is not a proof.
Founders first. The approach is how a claim is read before capital is committed.
Thesis
A checkable claim is the work
Production agents rest on probabilistic foundations. That is useful for proposing work. It is not a proof, and it is not a substitute for one.
AKVC invests where that difference is the product. We look for decision models that generate calibrated evidence — candidates that PASS, or the model abstains — and for independently machine-checked kernels that say what is Verified: a stated property, not a confident paragraph.
The companies in view make claims a later reader can examine. In clinical decision support, autonomy, privacy-preserving systems, and authorization, an uncheckable claim is a defect.
How a claim is held
Assessment and proof are not a ladder
A score does not become a theorem. The two jobs stay separate, and a person remains accountable for any written note.
Assessment
A score about a claim
Statistical assessment weighs how well an evidence file supports a specific claim. The number is not a rating of the company, and it is not a guarantee. Thin support should lead the model to abstain rather than invent confidence.
Proof
A kernel judgement
Some claims can be stated precisely: algorithmic guarantees, policy invariants, safety properties. Those are checked in Lean or a similar kernel. A model may draft a proof. Only the kernel accepts or rejects it. Independently machine-checked means the kernel, not the model, decides.
Scope
What we look for
Founders and companies building the stack, or products that sit on it, where an uncheckable claim is a product defect.
Evidence layers and kernels
Decision models that emit calibrated evidence, with a disposition to abstain, and independently machine-checked systems — Lean and related kernels — that decide what Verified means.
Products in high-cost-of-error settings
Domain work that sits on that stack: clinical decision support, autonomy, privacy-preserving systems, and authorization — places where trust is a product requirement, not a brand claim.
A spec, a kernel, and a go-to-market
Work that can be read together in one sitting. Research-heavy products are in scope when the claim, the check, and the customer can be named.
For founders
How founders engage
Write with the claim. We read it. A person decides, with the limits stated.
-
01
Write
Send the product, the claim, and what remains a candidate.
-
02
Read
We weigh evidence and, where it can exist, a formal check. Neither is a decision.
-
03
Decide
A person writes the note, with limits stated. A score is not a verdict.
-
04
Stay close
If we invest, we help on architecture, go-to-market, hiring, and the people who need a record.