Skip to main content
The Spectral Geometric Manifold (SGM) is the sensor engine underneath every World. A single self-supervised instrument that reads the structure of an interaction (not what was said, but how it unfolded) across the same measurement space, in milliseconds, without retraining for each domain.

​
What it reads

Any World produces an interaction: text, transcript, audio, video, environment. All of them enter the same instrument and produce the same class of reading: a spectral projection, a fidelity index, and traceable evidence segments. The reading resolves along three axes:
Spectral Geometric Manifold

The SGM. One instrument. Same geometry.

The same instrument reads a pension advisor in London and a market trader in São Paulo, a human and an agent, a boardroom and a market stall.

​
Four properties

Deterministic. Same input plus same Protocol equals the same output, bit-identical, across any deployment. This is a mathematical property of the spectral transform, not an approximation. It is the precondition for regulatory admissibility. Universal. The same measurement space holds across domains, languages, modalities, and participants. Human and machine judgment become directly comparable inside a single reference. Traceable. Every reading reconstructs to the specific segments of the interaction that produced it. The evidence ships with the score. Sub-millisecond. Complexity scales linearly. Population-scale throughput without GPUs at read time.

​
Provenance

Four years of research proved the physics. Eighteen months of empirical validation across regulated institutions confirmed the instrument in production. Paper 0 (Self-Supervised Meta-Heuristic Mapping) is published and cited in the physical edition of The Oxford Handbook of the Foundations and Regulation of Generative AI (2025). Empirical properties: reported on the public surface at blankstate.ai/sgm.

​
Current version

SGM 0.2 is in production. API URL: https://ibf.blankstate.ai/api/v1/. The version is pinned on the Protocol, not on the URL.