For investors

A proof-based AI runtime.
In a market that bet on wrappers.

Three patents filed at the USPTO. The wrapper market consolidated under the legacy security incumbents in 2024–25 and left the gap we sit in — the gap none of them built proof-based AI to fill.

The stakes

The guess used to cost money.
Now it can cost lives.

Every failure that kills a pilot traces to one fact: the model guesses. For five years that guess cost money — a refund, a bad disclosure, a deleted database. The database was recoverable. As the same models move from chat windows onto factory floors, roads, and hospital equipment, the guess starts driving a physical actuator. The person is not recoverable.

Jul 2025 · in production

An AI agent deleted a production database.

At Replit, an autonomous coding agent ran a destructive command against a live production system during a code freeze it had been told to honor — four safeguards in place, all four bypassed, 1,200 records gone. No wrapper, no dashboard, no policy caught it in time. The guess is already destroying things people cannot watch fast enough to stop.

Jun 2026 · the warning shot

The actuators are multiplying faster than the proof.

At a public Children’s Day demo in June 2026, a Unitree G1 humanoid roundhouse-kicked a child; a year earlier another flailed beside a factory worker. No cause has been published — that is the point. These machines are entering production while the guess is still inside them.

why us

We don’t ship robotics — and we don’t need to.

The gate admits or refuses an action before it fires, and the autopsy proves a control law before it runs — whether the action is a loan disclosure or the instant before a machine moves. Reaching a physical workload is a deployment decision, not new R&D. That is the by-construction thesis paying off.

The market

The wrapper bucket consolidated in nine months.
None of them built proof-based AI internally.

Cisco, Palo Alto, Check Point, and F5 each acquired one of the four loudest wrapper-style vendors. Roughly a billion dollars across the four deals. Every one of them strapped a classifier onto an existing security stack. None of them built a proof-based runtime. That’s the gap.

Cisco → Robust Intelligence
Aug 2024 · rumored ~$400M. Now “Cisco AI Defense” inside Cisco Security. Distribution at Cisco-security scale.
Wrapper / classifier
Palo Alto → Protect AI
Apr 2025 announced, Jul 2025 closed. Now part of Prisma AIRS. Model scanning + guardrails inside the Palo Alto stack.
Wrapper / scanner
Check Point → Lakera
Sep 2025 · $300M. Prompt-injection defense rolled into Check Point’s Global Center of Excellence for AI Security.
Wrapper / classifier
F5 → Calypso AI
Sep 2025 · $180M. Now “F5 AI Guardrails” + “F5 AI Red Team” inside F5’s application delivery stack.
Wrapper / orchestrator

The category words — “AI security,” “AI governance,” “guardrails” — are commoditized. The construction method underneath is not. That’s the moat.

The moat

Why a hyperscaler doesn’t close this in 18 months.

The wrapper paradigm became the default because it was easy to build on the mainstream construction method. The proof-based paradigm is the opposite: it requires a discipline the field does not scale by adding engineers.

01 · incentive

Hyperscalers sell inference, not gates that refuse it.

Deterministic admission gates that block actions cut against the revenue motion every cloud provider is built on. The bundled-safety primitives ship free because they have to, not because they want to. That’s structural.

02 · discipline

Lean 4 + applied formal methods is a twenty-year bench.

Galois has been doing applied formal methods for twenty-five years; the field grows by training, not hiring. The IP and methodology behind the proof discipline cannot be acquired off the shelf. SMARTHAUS is productizing what consultancies have historically delivered as bespoke engineering.

03 · architecture

The proof object lives at the gate, not the model.

The common formal-methods objection — “you can’t verify a seventy-billion-parameter network” — is correct and irrelevant. We don’t verify the LLM. We verify the contract the LLM’s output must satisfy before it commits as an action. Verifying an admission contract is decidable.

04 · timing

Regulation is moving from policy to evidence.

SR 11-7 demanded demonstrable independent validation in banking for fifteen years — and its April 2026 successor, SR 26-2, deliberately left generative and agentic AI outside its scope — handing each bank the job of governing AI on its own. The EU AI Act, Colorado, the FDA, and NYDFS apply the same posture across sectors. Customers who buy proof-based architecture now pay no retrofit bill. Customers treating regulation as a future problem are queuing up for the retrofit at the worst possible moment.

Where the category’s capital already sits

The rooms where this thesis is already understood.

SMARTHAUS is bootstrapped — we have taken no outside capital to date. The names below are not current investors. They are the strategic and venture firms that funded the wrapper wave: the rooms that already know the category, already paid for the output-watching version once, and are where the construction-method bet gets made next.

Strategic / corporate VC · category rooms, not backers

M12 / MicrosoftSalesforce VenturesSamsung NextCiti VenturesLockheed Martin VenturesIn-Q-TelKPMGBooz Allen HamiltonMozilla VenturesBloomberg BetaGitHub Fund

Pure-VC repeats in this set

Acrew CapitalEvolution Equity PartnersPaladin Capitalboldstart venturesSequoia CapitalTiger GlobalIndex VenturesGreycroft

The shape only counts if you know who’s building it.

A proof base can’t be cloned in a weekend.