VerifyCore Labs One-page brief
An AI research lab whose AI agents work on the systems AI runs on.
We lead with one problem: AI assistants that use tools can change real files and databases. One that dies halfway through a multi-step change can leave them half-updated, with nothing to finish or undo it. The lab also works on the networks that link AI chips, and on chip packaging and chip printing. Each result below links the file it rests on, and states its limit.
The problem
- AI agent platforms. An assistant that dies halfway through a multi-step change can leave files and database rows half-updated, with nothing to finish or undo it.
- Chip-package design. Flat, two-dimensional package models treat each vertical wire as a cross-section that goes on forever, so they miss the electric charge that collects at its two ends. A fast model graded only by its maker’s own software gives a buyer no outside check.
- Chip printing. After a small change to a mask design, the affected small pieces of it (tiles) are normally fully re-simulated, which is slow at full-chip scale.
- AI-cluster networks. Traffic spread over many paths reaches the network card out of order, and holding it until it can be put back in order costs memory that grows with the traffic.
What we showed
A fast coupling model, graded by outside solvers
On every sample set FastCap graded, and on the one slice of layouts Palace graded, the lab’s fast model of how electrical signals couple between the vertical wires of a chip package came closer to those two outside physics programs than the lab’s own pair-by-pair sum did, for one property, the charge they can store (capacitance). That sum is a baseline the lab defined, and the lab’s own records say full-wave field solvers do not use it.
two outside physics programsOn every sample set FastCap graded, the lab’s fast model came closer to it than the lab’s own pair-by-pair sum did; Palace, the second outside program, graded one slice of layouts only.
- Who we expect would buy.
- Extraction-tool vendors and package signal-integrity teams (no customer or pilot yet).
- Limit.
- Capacitance only, against a pair-by-pair baseline the lab defined; the two outside programs disagree with each other about how accurate the lab’s own solver is.
Fixed-size memory for reordering AI network data
In the lab’s own network simulator, a network-card design that keeps track of out-of-order data in a small, fixed amount of memory, compared on the same simulated grid with the lab’s model of STrack, a published fixed-size design, and with a second bounded-state design in the lab’s simulator.
1.965xThe lab ran its own network simulator over a grid of 64 simulated cases. In it, the lab’s model of STrack, a published fixed-size design by Le, Pan and Newman, needs this many times the reorder memory that the lab’s design does, on average (a geometric mean). Read it as about twice.
- Who we expect would buy.
- Network-card and switch-chip makers for AI clusters (no customer or pilot yet).
- Limit.
- Against a second bounded-state design in the lab’s simulator, which the lab calls CTS, the lab’s record gives 1.443x, CTS better in 31 of 64 simulated cases, a ratio of CTS’s memory to the lab design’s whose averaging it does not state; gaps are often tens of bytes, so the result is mixed; STrack needs less in some cells. A simulation, not silicon.
Checked brightness ranges for chip prints
A check of how a chip pattern will print that gives, for every point of the image, a range for its brightness. On every test mask the lab’s detailed simulation fell inside the range, apart from a rounding error in one internal step that the range leaves out.
232 test masksFor every pixel of each one, the brightness the lab’s detailed simulation gives fell inside the range the check computed. An approval check built on the ranges made no wrong approvals once a fault the lab found was fixed, on some of the same masks where the fault was found (a smaller approval test, not a held-out one).
- Who we expect would buy.
- Computational-lithography vendors and foundry mask sign-off teams (no customer or pilot yet).
- Limit.
- Relative to the lab’s own imaging simulation on a finite set of test masks, not silicon.
Crash recovery for AI assistants’ multi-step changes
A transaction layer meant to finish or undo an AI assistant’s multi-step change after a crash, for plans it accepts before they run. At every crash point the lab chose, it recovered the change whole or not at all; killed at random moments, some recoveries were not, and the lab has not fixed that yet.
atomic recovery recorded at every one of those 29 pointsThe lab killed the program mid-change at each of them, and each time a fresh program recovered the change whole or not at all from the log on disk.
- Who we expect would buy.
- AI agent-platform teams (no customer or pilot yet).
- Limit.
- 2,893 further kills at seeded random moments found 31 recoveries that were not atomic (recorded in the lab’s claim; the run record is not published); the lab has not fixed them yet. Power cuts are untested.
The capacitance flat models leave out
The lab’s own three-dimensional solver measured how much of one simplified chip-package connection’s capacitance (its charge-storing capacity) flat, two-dimensional models leave out.
about 40%of one simplified connection’s capacitance (its charge-storing capacity), measured with the connection driven as one of a pair, sits at its two ends, which flat, two-dimensional models leave out. The lab’s own solver was checked first against shapes with exact answers, then against FastCap, an outside solver run on the lab’s own mesh of the shape.
- Who we expect would buy.
- Chip-package design teams and extraction-tool vendors (no customer or pilot yet).
- Limit.
- One simplified connection, a bare cylinder; whether the missing share changes any figure the lab’s own models sign off on has not been tested.
Every result, with all its limits, and the lab’s exact wording for a technical reader: verifycorelabs.com/theorems
Why now
Research groups have begun publishing transaction layers for AI assistants’ tool calls (Cordon, submitted 2026-06-16, and SagaLLM, first submitted 2025-03-15). A platform that lets an assistant change customer data has to answer for the change it leaves half-done when it dies. The Ultra Ethernet Consortium has published its specification for AI-cluster networks, which sets the ordering rule for network cards that use its ordered-delivery mode.
The companies
- ChipletOS chipletos.com: Chip packaging and chip printing: coupling, capacitance and mask checks that carry their evidence, proofs where a result says so and measurements otherwise.
- OrbitalProof orbitalproof.com: Networks and AI assistants: post-quantum Wi-Fi, Ultra Ethernet, and transactions and tool checks for AI assistants.
- AxiomLimit axiomlimit.com: AI inference infrastructure: a network-simulation result on reorder memory, a simulated deletion-receipt study and an abstract bound on cache bookkeeping.
- FluxZero fluxzerofluid.com: Cooling: two-component coolant blends, enumerated and scored on five axes: heat flux, permittivity, safety, climate impact and boiling window. Its work is not among the results published here.
- Lattice Graph latticegraph.com: Materials data, served through a public production API and a Python client (pip install latticegraph). Its work is not among the results published here.
Why us
Nick Harris, founder. CTO of VivaMed BioPharma; co-founder of MedSim.ai, FastRead.io and Formulai.
AI agents do our research and engineering, and no outside firm has audited the work. Each result links the lab’s file that states it; the command the lab ran is named in the lab’s file; its code is not public.
- For the transaction layer and the mask-edit sign-off, we tried to break the check on purpose: a separate program that did not build the result changed a number the check reads, and the check caught it.
- For the capacitance measurement, no automated check recomputes the headline share on each run, so it is a measurement reported once.
- That separate program works inside the lab’s programme, so this shows that the checks can fail, not that anyone outside the lab has graded the results.
Stage: No customer, no revenue, no pilot, no third-party audit.
Contact
Acquisition, licensing and partnership: nick@latticegraph.com · verifycorelabs.com