This is the public update surface for research releases, architecture milestones, demonstrations, company announcements, and selected field notes.
Revalidation before execution, revocation, evidence-change handling, and causal decision chains are now first-class runtime behaviors — authority granted at decision time cannot silently persist through changes in evidence, policy, or operational state.
Read the RegOS overview →The public writing trail now spans 12 Hugging Face articles covering governed AI, low-precision training, quantization, and the IDA research program, plus 2 SSRN papers on collaborative governance and anti-discrimination. Source →
RegOS now treats revalidation before execution, revocation, evidence-change handling, and causal decision chains as first-class runtime behaviors. Source →
A controlled evaluation harness exercises governed systems over long horizons with incomplete information, where ground truth is held by the evaluator. Source →
Signed execution lineage and persistent device identity connect physical state, device continuity, and operator authority to the governed runtime. Source →
Validation work across 4-bit floating-point formats and selective-compute scheduling supports training and evaluating specialized model bodies on native GPU infrastructure. Source →
The company is advancing mission-assurance research that connects governed authority, long-horizon evaluation, and physical trust for mission-critical and regulated operational systems.
A Hugging Face article documenting the port of the IDA native training engine from Hopper-class infrastructure to consumer Blackwell hardware. Source →
A root-cause investigation showing how a fixed activation-row clipping ceiling became the primary cause of divergence in an FP8 mixture-of-experts training run. Source →
