For investors
The research is done. The engineering is next.
We have built a verified execution substrate for model-written code. It runs on the target processor and is published for anyone to check. What remains is hardware and conformity assessment. Both are expensive and slow, and both sit on a path thousands of machinery vendors have already walked. A round here buys execution on a route that is already mapped.
What is already built
We reached this without outside capital, and you can check every line of it without asking us for anything.
| Trusted core | 880 measured lines of no_std Rust, an interpreter and an envelope monitor |
|---|---|
| Machine-checked | thirteen properties discharged, one exhaustive over all 232 inputs |
| On the target | 1,204 bytes of RV32 machine code, five checks passing on that instruction set |
| Adversarially tested | a simulated cell, four attack classes, every attempt and its outcome published |
| Open to inspection | source-available, with the proof harnesses runnable by anyone |
We are not asking anyone to fund a hypothesis. Whether a machine-checked interpreter can carry model-written code inside a bound small enough to audit was the open question, the half of this problem people expect to be a research project, and we answered it in public, with the artefacts to check the answer.
Why the timing is not ours to choose
Every vendor putting a model near an actuator faces the same three requirements on the same date. Regulation (EU) 2023/1230 applies from 20 January 2027: machinery with self-evolving behaviour must not act beyond its defined task and movement space, must record its safety-related decisions, and must remain correctable at all times. Annex I Part A puts machine learning that ensures a safety function into mandatory third-party assessment with no self-certification route.
The architecture that answers them, an untrusted planner behind a verified supervisor, is the one industrial robotics has shipped under type certification for years. The substitution is that the planner is now a language model.
The queue is the part people miss. The rules governing notified bodies applied from 20 January 2024, so bodies have been designating against the new requirements for two years and assessment capacity is finite. Being early in that queue is worth more than being early to the idea.
What the first round does
Nine months, three hires, and four results that turn this page into numbers.
- A board, measured
Every projected latency on this site replaced with a measurement: bring-up on the selected FPGA, timing closure, and an end-to-end figure.
- The chain completed
Audit chain and effect construction written and checked, and the memory-protection configuration validated, which is the last link below the software that is not yet proven.
- Partners in the loop
Two or three platforms running their command path through the gate over PWM passthrough, each producing an evidence record they keep.
- Assessment, quoted
Written quotes from notified bodies, which converts the largest line item in any later plan from an industry rumour into a contract.
Three technical hires alongside the existing team: FPGA and firmware for the board, embedded verification for the audit chain and the protection configuration, and a functional-safety engineer to own the conformity path from the first week rather than the last.
The bottom-up cost basis, the salary and hardware assumptions, and the staging rationale are in a memo. Ask and you will get it.
Where the value sits
The value is not in the code, and we say so before diligence does. We publish the interpreter; the architecture has been public for months; nothing is filed and the EPO applies absolute novelty. Anyone can read what we built.
What is hard to copy is the position at the end of this round: a certified device with a notified-body relationship, an evidence format that customers have written into their own safety cases, and a public record of every claim this company has made and revised. Certification is a moat because it is slow, and the clock started for everyone on the same day. The trust record is a moat because it cannot be assembled retroactively.
Nobody else is building this shape. Adjacent rounds price AI design tools and testing platforms. Chipmind at $2.5M pre-seed, Alpha Design AI at $3.1M, ChipAgents at $21M Series A, Momentic at $15M Series A. None of them carries hardware in a safety chain or a conformity path. There is no valuation anchor, and there is no crowd.
What has to be true
Four things. Three are measurements we will have inside nine months, and we publish them either way.
- The budget closes
Roughly 57 to 144 interpreter instructions fit a 2 ms tick at 25 MHz. If the fast multiplier is not enough, the answer is 250 Hz, which most of the target platforms run at anyway. Measured on the board in month three.
- Assessment is affordable
The industry figure is twelve to twenty-four months and mid six figures, and it is folklore until an assessor writes it down. The regulation directs notified bodies to set fees with SMEs in mind, in a sector that is 98% SME. Quoted in month two.
- Somebody wants it now
The requirement lands in January 2027 and the platforms it lands on mostly have no safety-rated envelope to fall back on. Whether that urgency is felt today is what the design-partner conversations settle.
- No one is further ahead
Synapticon markets "AI behavioural safety" with an A3 award and a TÜV Süd relationship, and has published no definition of what it covers. It is the one competitive question we cannot yet answer, and the first thing we would want answered too.
Two boards, three written quotes and one design partner settle almost everything on this page. That is what the next nine months are for, and the list is short enough to hold us to. We publish these because the alternative is finding them in diligence, and because three of the four are measurements rather than opinions.