The third-party work that establishes the requirement we build for, and our own research.
The founding study of the reference monitor concept concluded that software interpretation suffices only for restricted systems, and that general-purpose secure systems require hardware interpreters.
A joint study tested twelve published defences against jailbreaks and prompt injection. Most were bypassed with attack success rates above ninety per cent. The majority had originally reported near-zero attack success.
A peer-reviewed proof extends Gödel's incompleteness theorems to AI, establishing that no finite set of guardrails can be universally robust against adversarial prompts.
Google's security engineering organisation published that authorisation must move from the application to the individual action, and that beneath any AI reasoning layer there must be a floor of static policy that can be verified.
Agents escaped a purpose-built isolation environment, reached a third party's production systems, and tampered with their own transcripts to conceal it. OpenAI's post-mortem concludes safeguards must operate at the speed of the AI agents themselves.
Sources are linked. None of them endorses Evrmind.
Only one of these sources calls for hardware directly. The rest describe the gap our work sets out to fill.
The requirement has been the same for over fifty years.
The work that defines the limits and requirements the architecture is built to satisfy.
A category-theoretic framework defining the properties any physical enforcement layer must satisfy to guarantee safety constraints, and how those properties hold across electronic and photonic substrates. The architecture is built on this work.
Zenodo record 23581490
A frozen model can be steered at runtime towards a safety rule, but only so far. This work identifies two hard limits: a ceiling set by what the model's own read state already implies, and a horizon on how far a bounded correction can reach, measured on activations from real 7B and 32B models. The systems implication is that enforcement should not depend entirely on the mutable system it is meant to govern.
Zenodo record 23578032
Published on Zenodo, openly accessible.
A deterministic symbolic substrate that rewrites its own structure under a formal check applied before each change runs. Because the core holds no probabilistic model, the check binds to the computation that will actually execute, not to an assumption that a model behaves. Evaluated at controlled toy scale.
Zenodo record 23578696
A neuro-symbolic architecture exploring deterministic, checkable computation as an alternative to fixed-weight models. It grounds continuous sensing, retains earlier competence, and holds a reasoning core that revises its own program under proof rather than by adjusting learned weights. Validated at toy scale.
Zenodo record 23622574
Our early geometric approach to safety constraints, treating formal rules as geometric barriers on a continuous manifold, so they can shape learning without the information loss of discretisation.
Zenodo record 23569730
An exploratory physics-constrained synthesis framework treating energy cost as a first-order constraint rather than an externality. Revised following external peer review.
Zenodo record 23564348