Give TypeSafe’s model, Jev, a set of permitted answers and it chooses between them directly. The result is faster, cheaper and far easier for software to act on. And, because the model cannot stray outside those permitted answers, TypeSafe makes the striking claim that its AI cannot hallucinate.
For insurance, that proposition deserves closer scrutiny.
Insurance decisions have to reconcile commercial interests, contractual obligations and regulatory expectations about customer outcomes. As AI moves from assisting underwriters and claims adjusters to participating in decisions themselves, the standard has to be higher than producing an answer a computer can more easily consume.
And under closer scrutiny, three problems emerge.
1. Zero hallucination isn’t quite what it sounds like
Start with TypeSafe’s most provocative claim.
Give the model: Approve / Reject / Refer
and it cannot invent a fourth answer. That’s the guarantee.
TypeSafe’s 0% hallucination claim is about ensuring that the output conforms to the permitted choices, not establishing that the choice itself is correct.
So:
TypeSafe can guarantee that the AI won’t invent an answer you didn’t give it. But that tells us nothing about whether the answer it selects is the correct one or whether the correct answer was among the choices in the first place.
An answer can be perfectly valid within the choices TypeSafe was given, and completely wrong in the real world.
That’s the difference between controlling the form of an answer and establishing that the answer is correct.
For insurance, the latter is the problem that matters.
2. Selection isn’t determination
This points to a more fundamental problem with the TypeSafe proposition: what it means to make a decision.
Consider two questions that look almost identical:
“Does this policy cover cancellation following emergency surgery?”
and:
“My travelling companion has had emergency surgery. Is this cancellation claim covered?”
Computationally, they can be very different questions.
The first may be answered by finding and interpreting the relevant provisions in the policy wording.
The second asks whether a particular claim is covered. That requires more than recognising that emergency surgery can fall within the policy; the relevant facts of the claim must also satisfy whatever the policy requires.
That distinction becomes especially important when AI is used directly with customers.
A customer asking whether their policy covers cancellation following emergency surgery is unlikely to be interested only in an abstract interpretation of the wording. What they usually want to know is whether their particular circumstances are covered.
An AI system may still answer by finding the relevant wording and selecting ‘covered’ as the answer most strongly supported by the information available. But that does not establish that the claim itself has been determined.
TypeSafe can attach confidence to that prediction, allowing uncertain cases to be referred. But confidence in a prediction does not turn it into a determination. A confidence score tells you how strongly the model supports an answer; it does not establish that every condition capable of changing this particular claim has been resolved.
The consequential decision problem is therefore not simply:
Which permitted answer is most probable?
It is:
What must be resolved before one outcome is actually determined by the policy and the facts?
Those are not the same computational problems.
3. TypeSafe assumes the decision model
That leads to the third problem.
Jev isn’t intended to be the entire decision system. It operates inside one.
But that means the correctness of the overall decision still depends on the model around it: what questions are asked, which facts matter, how the answers interact and when those answers are sufficient to determine an outcome.
A policy may still exist as a 100-page PDF written for humans. AI can find relevant provisions, interpret the language, identify definitions and exclusions, and turn previously inert documents into computationally accessible information.
But understanding the content and structure of a policy is not the same as being able to compute the determination that the policy and facts are meant to justify.
That requires a model of the decision itself.
Rules can encode decision logic. Learned models can infer relationships from data. Constraint systems can specify conditions that an answer must satisfy.
But they all encounter a fundamental difficulty: the decision model has to come from somewhere.
Rules require relationships to be authored. Learned models infer them probabilistically. Constraint systems require the relevant constraints to be specified. None of those approaches, by itself, establishes that the model contains everything necessary for the determination.
TypeSafe doesn’t solve that problem. Its own evaluation methodology makes the dependency explicit:
“Assume the harness is correct.”
That’s entirely reasonable when testing a component. But it also defines the boundary of TypeSafe’s contribution.
TypeSafe provides a probabilistic component inside a decision system whose correctness it assumes.
That might be useful. But it isn’t a breakthrough in computational decision-making.
The question of where the structure required for the determination comes from remains.
From prediction to determination
That is where Geometric starts.
Geometric AI is a frontier lab based in the Knowledge Quarter around King’s Cross, London, developing its research in partnership with Claim Technology, with insurance as its initial domain.
It starts with a fundamentally different computational objective.
Rather than beginning with a decision model and asking which answer it produces, Geometric asks what must be resolved before an answer can be determined at all.
The computation then determines which outcomes remain possible under the facts and applicable rules. If one admissible outcome remains, it is determined. If more than one remains, no determination can yet be made, and what remains unresolved can be identified.
A determination exists because the alternatives have been eliminated by the facts and rules, not because one answer scored higher than the others.
Crucially, Geometric does not simply assume that someone has already built a complete decision model. That modelling problem is part of the research itself. Geometric asks whether the structure required for a determination can itself be established computationally rather than supplied in advance.
That places it beneath, rather than in competition with, rules, learned models and constraints. Each may still have a role in the computation. The deeper question Geometric addresses is what structure must be present for their outputs to contribute to a valid determination.
Determination means eliminating what cannot be true until only one possible outcome remains.
A stronger meaning of hallucination-free
And that brings us back to where we started.
TypeSafe describes Jev as “hallucination-free” because its possible outputs are defined in advance.
But preventing a model from producing an answer outside a permitted set says nothing about whether the answer it produces within that set is warranted.
A valid output is not the same as a valid determination.
Geometric’s claim is different: the computation cannot manufacture a determination where the structure, facts and rules do not warrant one.
Which gives us perhaps the simplest distinction of all:
TypeSafe constrains what answers the model can produce.
Geometric constrains when an answer can be determined.
The TypeSafe red herring
TypeSafe has identified a genuine inefficiency in today’s AI.
Using a machine designed to generate language when a computer ultimately needs approve, reject or refer is inefficient and unnecessary. TypeSafe makes probabilistic AI easier for software to act on.
The red herring is mistaking that advance for a breakthrough in decision AI.
It improves how probabilistic AI returns an answer while leaving the harder question untouched: What makes that answer the determination?
That is the problem that matters in insurance.
Insurers operate in a world of enormous complexity: different products, policy wordings, jurisdictions, facts, evidence, commercial considerations and regulatory obligations—all interacting in decisions that need to be made quickly, consistently and explainably.
Geometric offers a different proposition:
Make the determination itself computable.
Not by predicting through the complexity, but by resolving the facts of each case against the rules that apply, determining an outcome when one can be determined, and identifying what remains unresolved when it cannot.
Generative AI made language computable.
TypeSafe has shown that when software needs an answer, generating an essay first is unnecessary.
But making probabilistic AI easier to consume doesn’t solve the decision problem. It just makes the prediction easier to automate.
For insurance, determination is the breakthrough that matters.
That’s the frontier Geometric AI is opening, and why mistaking TypeSafe’s advance for the future of decision AI is ultimately a red herring.
