Research
Four areas, one shared requirement
The research advances across four distinct areas, which exercise the foundation described in The method on real substrates and measure how far it holds. All of them serve the same aim: deterministic, certifiable embedded AI, currently at the prototype stage. Each area has its own detailed page, with protocols and results.
Generalisation, level by level
The question fits in one sentence: what exactly does a system learn when it is said to have learned? I answer it with a three-level scale. The first is learning by direct consequence, where a favourable experience is to alter a later preference. The second is generalisation to situations neighbouring those encountered. The third is the recomposition of known elements into a combination never seen before.
The result verified to date is this: on a learning circuit modelled from a part of the fruit-fly connectome, my prototype learns a task, then a second one, and partly retains the first, above chance; the learning protocol is supervised. The levels of the scale structure the roadmap, and the detailed page explains why this scale is the programme’s most useful measuring instrument: it indicates where the design must evolve, not merely how much more training would be needed.
The roadmap of the three levels →
The connectome testbed
The neural wiring of the fruit fly, published by the FlyWire consortium, is one of the few complete connectomes available at this scale — 139,255 neurons and their connections. I use it as a large-scale test structure: in paper D, the topology provides a realistic and demanding graph on which execution determinism holds at full scale, with reproducible fingerprints and exact replay (two runs give the same trace). To test my hypotheses on learning, I use a much smaller circuit, modelled from a part of this connectome.
The topology serves as a wiring substrate and nothing more: the dynamics, the weights and the learning rules belong to the project. This area shows that bit-for-bit replay of the engine holds when moving to scale, checked by double execution up to the complete graph (a test, not a proof).
The simulated organism
A simulated quadruped gives the system a body, and therefore consequences. This is the area where the properties of the verifiability layer become concrete demonstrations: replaying an execution identically, resuming it after an interruption at the exact point where it stopped, establishing, by disconnecting a part, that it is necessary to a given behaviour — over hours of continuous execution.
The demonstration concerns the continuity and auditability of a simulated organism that lasts; what it learns belongs to the roadmap, and no form of cognition is claimed.
The mathematical work
Two pieces of work consolidate the programme at its base. The arithmetic specification describes integer computation completely enough that any two machines produce the same result bit for bit, with controlled rounding and explicit saturation. It is deposited, public, and can be implemented independently by a third party; three independent implementations, in C, Python and Rust, already produce the same conformance hash.
If one of these areas meets your own
The specifications, the public kernels and the conformance formats are available, under declared licences.