Sources for the reading guide
Everything on the reading guide that rests on somebody else’s work, what each source supports, and how far it can be relied on.
The guide argues that you should check claims and know which kind you are looking at. It would be poor form not to make that possible here. Where a figure comes from a company describing its own product, it is marked as such — that is not a reason to dismiss it, but you should know whose number it is.
World models and JEPA
Supports: LeCun’s argument against next-word prediction, what JEPA does, AMI Labs and its funding, and that it is based in Paris.
Meta AI. The architecture described by the people who built it. Supports what JEPA predicts and why.
AIwire, March 2026. Supports the funding round and the company’s stated focus.
Wikipedia. Supports the founding date, location and who holds which role. An encyclopaedia entry, so worth confirming if anything turns on it.
Embodied and sensory AI
Supports: what vision-language-action models attempt, and which problems the field itself ranks as unsolved.
The three priorities named on the guide — language-to-action mapping, multimodal perception, uncertainty estimation — are this review’s ranking, not ours.
Supports the description of what a VLA model is and where it currently falls down.
Supports the manufacturers named and the practical characteristics — per-pixel change reporting, low light, data reduction. Trade reporting; the underlying claims are also in the academic literature below.
Academic grounding for how event sensors differ from frame cameras.
Diffusion language models
Supports: that text can be generated by denoising rather than left to right, and the throughput claim.
Inception Labs. The throughput figures on the guide are the company’s own, for its own product — over 1000 tokens/sec on NVIDIA H100s, 1109 for Coder Mini, 737 for Coder Small. Not independently verified here. The description of coarse-to-fine generation is also theirs.
Field overview, and the route to LLaDA and the other model families named.
Sensorium
Supports: what Sensorium is, who built it and how.
Ben Reid, Memia. The description of Sensorium as an exocortex built over eighteen months with AI coding agents is his own. Memia is a Christchurch research outfit; its newsletter is at memia.substack.com.
What is ours
The five mechanisms, the stack, the argument about the agentic layer, and everything about sovereign data, records, databases, sealed records and situated language layers is our own work, published on this site and on mysovereignty.digital under CC BY 4.0. The guide links each piece where it is set out in full.
Two of those pieces carry a sealed record you can check without trusting us: What Survives When the Secret Doesn’t and What it costs.
Sources checked 29 July 2026. Links rot; if one of these has gone, the archived copy is usually findable and the claim it supported is stated above so you can judge it yourself.