A fast coupling model, graded by outside solvers
6.1×
Median: less error than lab-defined sum; one of several
The sum: each pair solved alone in FastCap, then added up; the lab’s model and the sum were both graded against FastCap’s full solve.
Autonomous Proof-Carrying Engineering.
The lab behind five companies, for teams that build AI’s chips, networks and agents: license or buy a checked result with the files behind it, or commission an evaluation.
- 6.1×Smaller coupling error than pair-by-pair sums; FastCap-graded median of 24 simulated layouts
- 29/29Crash points the lab chose; each recovered whole (OrbitalProof)
What the lab has shown
Results from the lab’s record, in chip packaging, chip printing, AI-cluster networks, AI assistants and AI inference; the full list is on the results page.
Fixed-size memory for reordering AI network data
1.965x
Reorder memory the lab’s STrack model needed, relative to the lab’s design: geometric mean over 64 simulated settings
In the lab’s network simulator, a network-card design that tracks out-of-order data in small, fixed memory, against the lab’s model of STrack, a published design.
Checked brightness ranges for chip prints
232 test masks
On the lab’s test masks, its detailed simulation fell inside the brightness range its check predicted.
Crash recovery for AI assistants’ multi-step changes
29
Crash points the lab chose; each recovered whole
At each of those points a restarted program finished or undid the interrupted change, for plans the layer admits in advance.
Capacitance at the ends of one simplified via, in simulation
about 40%
Share of one simple wire’s capacitance (its stored charge) that sits at its two ends, which flat models miss, in the lab’s 3D simulation.
An Ultra Ethernet recovery design, proved in a model
A machine-checked proof that, in the lab’s small model, its recovery design never hands data on out of order or twice.
Pair-by-pair coupling estimates overstate the worst case, in a model
at least 1.10×
Minimum overstatement of the worst case, proved
Adding up the coupling between wires one pair at a time overstates the worst case for every layout in one family.
Tool combinations that leak a secret
2,379
Runs on real files and a real database
In the lab’s test world, three combinations of an AI assistant’s tools leak a stored secret.
Where deletion receipts can be fooled
A signed deletion receipt can pass while the data is still recoverable, in simulation.
Signing off some tiles near a chip-mask edit, in simulation
0 violations
In the lab’s single-edit tests, no tile its check approved came out worse than a full re-simulation.
An automated attack search against quantum-safe Wi-Fi sign-in
11,200
Combined Wi-Fi attacks tried; none got through, against the lab’s own models
A search with no starting attack combined known Wi-Fi attacks against ten of the lab’s own models; a null result, within a small fixed budget.
A memory ceiling for post-quantum Wi-Fi, proved in a model
64 KiB
Ceiling on memory held for half-received sign-ins, proved in the lab’s model
In the lab’s hand-written Lean model, an access point never holds more than this for half-received sign-ins from unverified devices, however many connect.
Least bookkeeping a shared AI cache needs, in a model
T*n bits
Least bookkeeping a leak-free shared AI cache must keep, in the dense rights model
Proved in the lab’s model: the least bookkeeping a shared AI cache needs, but only if any item can be shared with any group.
All fourteen resultsThe formal statements, for a technical readerEach company’s own result pages, with their headline figures
Why it matters, why now, and who it is for
Why it matters
The systems AI runs on can fail in ways ordinary tests miss, and the lab’s results study four of them, in its own models and runs: an assistant that dies halfway through a change, a network card that runs short of reorder memory, a coupling estimate graded only by its own vendor, a mask approved by an approximate check.
Why now
AI assistants have begun changing real files and databases, and research groups have begun publishing ways to make those changes recoverable (Cordon, SagaLLM). The Ultra Ethernet Consortium has published a new standard for AI-cluster networks. Chip packages are stacking chiplets vertically (UCIe), and simulating how masks print already costs tens of billions of processor hours a year (NVIDIA).
Who we expect would buy
- Chip-package design teams and extraction-tool vendors
- Network-card and switch-chip makers for AI clusters
- Computational-lithography and mask sign-off teams
- AI agent-platform and runtime-security teams
- AI inference and cloud providers, and their auditors
What each result offers each team is on the for buyers page.
VerifyCore Labs: the engine behind five companies
The lab’s AI agents do the research and engineering; each company takes the results into its own market: ChipletOS, OrbitalProof, AxiomLimit, FluxZero and Lattice Graph. A buyer deals with the lab for any of them.
What it has produced
Fourteen results
in chip packaging, chip printing, AI-cluster networks, AI assistants and AI inference, each with the lab’s files behind it. All results
2
of the fourteen graded by, or cross-checked against, outside physics solvers the lab did not write (FastCap, Palace). OrbitalProof’s harness also reproduces NIST’s own test vectors.
Checked against references the lab did not write
- ChipletOSA fast coupling model, graded by outside solvers6.1× smaller coupling error than a pair-by-pair sum the lab defined (each pair solved alone in FastCap, then added up), both graded against FastCap’s full solve; median over 24 simulated layouts; one of several figures in the lab’s recordPalace, a second outside solver, graded one slice of layouts. Both are physics solvers the lab did not write.
- OrbitalProofPost-quantum key exchange and signature keys on liboqs255 NIST test vectors reproduced byte for byte by liboqs, an outside library, in the lab’s test harness (ML-KEM; ML-DSA key generation only)This checks the harness and liboqs, not a lab result; published on orbitalproof.com.
What a buyer can get
- License a result. The result, the lab’s files behind it, its proof where it has one, and the right to use it.
- Buy a result outright. The same, by purchase instead of licence. How to start
- Commission an evaluation. A scoped technical evaluation, with its evidence and pass mark agreed before work begins. Consulting
One lab, five companies
The companies are one lab’s work, each carried into its own market; each result sits under the company whose market it serves. FluxZero’s coolant work and Lattice Graph’s materials data are published on their own sites.
ChipletOS
chipletos.comChip packaging and chip printing: coupling, capacitance and mask checks that carry their evidence, proofs where a result says so and measurements otherwise.
Its results ( opens the same result on chipletos.com)
More on chipletos.com: How our proofs cover a whole family of layouts at onceWiring models built to stay passive in exact arithmeticA design-check certificate that records its inputs, model and boundsChoosing which rigorous printing simulations to run, against a proved floor
OrbitalProof
orbitalproof.comNetworks and AI assistants: post-quantum Wi-Fi, Ultra Ethernet, and transactions and tool checks for AI assistants.
All 255 NIST test vectors the lab ran on liboqs reproduce byte for byte (ML-DSA: key generation only). The check on orbitalproof.com
Its results ( opens the same result, or its results page, on orbitalproof.com)
More on orbitalproof.com: The lab’s admission-engine chip block, checked against its own source designOne protocol rule turned into checks for chips, firmware, kernel and logsA permission grant cut to the minimum on one workload, partly on the otherCandidate code fixes scored by proof, some left undecidedSix weakened safety checks that random fuzzing missed and coverage-guided fuzzers found
AxiomLimit
axiomlimit.comAI inference infrastructure: a network-simulation result on reorder memory, a simulated deletion-receipt study and an abstract bound on cache bookkeeping.
Its results ( opens the same result on axiomlimit.com)
More on axiomlimit.com: Where to place a shared AI cache, on one test instanceThe most an empty AI cache can hit, on a slice of a public trace
FluxZero
fluxzerofluid.comCooling: two-component coolant blends, enumerated and scored on five axes: heat flux, permittivity, safety, climate impact and boiling window.
Lattice Graph
latticegraph.comMaterials data, served through a public production API and a Python client (pip install latticegraph).
On latticegraph.com
Why this team
Nick Harris, founder. CTO of VivaMed BioPharma; co-founder of MedSim.ai, FastRead.io and Formulai.
The lab’s AI agents did the research and engineering behind every result here. In the lab’s current method, each piece of work names its check before it starts, a separate verifier re-runs that check, and a failure gets one fix before it stops for a person. The counts are on How we work.
Where we are: No customer, no revenue, no pilot, no third-party audit.
License or acquire a result
You get the result, the lab’s files behind it (and its proof, where it has one) and the right to use it, by licence or by outright purchase. Terms are to be discussed; no price is stated here.
To start, write to nick@latticegraph.com and name the result.