Work you can interrogate.
Interactive maps, replication reports, reference tools — each built the same way: every claim sourced, every finding adversarially reviewed, delivered as a single self-contained page.
Athlone to Nottingham
A live, in-browser illustration of Tero-type (Physarum) conductance dynamics on a stylised network of OpenStreetMap roads and car ferries between Athlone and Nottingham. Watch the flow settle across the routes and see where it concentrates. It illustrates the mathematics behind Many Nets, One Field. It is not an experiment: the costs are stylised, it uses no traffic data, and it is not a navigation tool.
Open the illustration →Many Nets, One Field
Nearly every mechanism behind a Physarum-style shared-field router has a precedent, in ant-colony optimisation, floor-field cellular automata, GPU agent modelling or placement practice. What we could not find, within the reach of our searches, is their combination as a label-per-net, capacity-limited router for interposer and RDL assembly, evaluated at real sizes; none of the Physarum routers we could read reports contest-suite results or overflow. A projection against current capability, not a result (revised 2026-10-02).
Read the thesis →Scaling Out Lane Parallelism
Thousands of random programs run at once on TEMPO's own gate-level netlist, each compared with an independent reference model, at up to 18,000 programs a second on a single Mac mini. Four real gaps between the reference model and the RTL found and confirmed, since resolved in the model along with eight more; the GDS built for submission, read back by an independent extractor, identical on every program run on it; a GPU-compiler bug traced to one wrong gate; and every speed-up proven identical to what it replaced before it counted (revised 2026-09-27).
Read the writeup →Retracing Two Stars
Jane Street's ASIC puzzle shipped a chip layout with every name stripped. RETRACE rebuilt its netlist from geometry with two independent extractors, recovered readable RTL proven equal to it, found an 11 by 11 Star Battle checker, and solved it two independent ways to a unique answer: (* TWO STARS *). Then the same extractor became an independent LVS for our own tapeout.
Read the writeup →The Calculemus Audit
A published Trail of Bits compiler bug in Lean 4 forced an agentic theorem daemon to answer a question it could not settle from its own logs: when the system stamps Q.E.D., did a kernel decide, or did something merely say so? Nine adversarial rounds, six decision records, and the finding that the most dangerous failures were the ones that returned success.
Read the audit →NaLa Under Replication
An independent replication of NaLa and secant — two open-source tools for editing and explaining transformer attention. Every test passes and the core identity holds to machine precision, but three of the project's headline claims do not survive contact with its own code and data.
Read the replication →Rack-Scale Compute Ecosystem Map
Rack-scale AI platforms — NVIDIA NVL72-class racks, AMD Helios, hyperscaler ASIC racks, China supernodes — mapped against the consortia, standards and vendor programs that convene them, with every tie sourced and adversarially reviewed.
Open the map →AI Trust Ecosystem Map
The Open Secure AI Alliance mapped against 33 consortia, standards bodies and policy frameworks across AI risk, reliability, safety, security, governance and policy.
Open the map →