S
Decompose and project
Spectral
Read every dimension at once.
A spectral decomposition gives the model access to many dimensions of the signal in one pass. Fast enough to be live, structured enough to remain auditable.
The Spectral Geometric Manifold
Any signal.
One model.
A self-supervised spectral model. It reads interaction across text, audio, video, and agent worlds, at sub‑25 ms per reading.
Self-supervised. Multimodal. Privacy-grade.
SGM 0.2 · in deployment
Spectral Geometric Manifold
blankstate research
0.2 · in deployment
Reads
text · audio · video · world
Speed
sub-25 ms per reading on CPU
Returns
deterministic, fidelity-qualified
The acronym is the architecture
Spectral. Geometric. Manifolds.
Three functions, one construction. Energy is the substrate that ties them.
G
A continuous geometric space
Geometric
Meaning lives in a learned geometry.
The model is a geometry. Directions carry concepts, distances carry similarity, regions carry whole families of meaning, and trajectories carry change over time. A reading is a projection through this geometry, so the structure of the answer is always visible alongside the answer itself.
M
Position and trajectory
Manifolds
Actants live on a surface.
Each actant occupies a position on a learned manifold and moves through it as the interaction unfolds. The model reads the curve. A single construction, zero-shot across domains and actants.
E
One unit, end to end
Energy
Energy is the unit.
Energy is the unit at both ends. At training, the model minimises predictive energy in latent space. At output, fidelity reports the captured energy of each reading. The same number, the same meaning, throughout.
any signal · one unit of measure · one fidelity index
Two readings
Across the interaction. Per actant.
The same model produces both. One reading follows the interaction as it unfolds, the other follows each actant through it.
01 / Across the interaction
The interaction as it unfolds.
One reading follows the interaction as a whole. Each segment is read in context of what came before, so the score reflects the arc of the exchange across every moment. Same for a dialogue, a call, a video, or an agent trajectory.
02 / Per actant
Each actant, read distinctly.
The other reading is per actant. Each one carries a separate trajectory through the exchange, so the same signal from different actants produces different scores, and the direction of influence between them is itself measurable.
any signal · two readings · one unit of measure
Privacy by construction
Energy reads the interaction.
A fidelity score is what comes out.
SGM measures the fidelity of an interaction and returns a score. That score is all that comes out. The model runs where the data already lives, and what leaves is a number. Privacy is structural, built into the measurement itself.
Signal in · interaction session
Raw signal
conversation_session_4f29.m4a
audio · 4 actants · 00:08:42
Content sealed
Energy out
0.87 fidelity
No PII crosses
Scalar only · non-reconstructible SGM reads the energy of the signal · what crosses the boundary is a score · fidelity on a named axis
01
Scalar output
Every reading is a number on a named axis. That number is all that leaves the signal.
02
On-device viable
Sub-25 ms per reading on CPU. The model reads where the data already lives.
03
Federated by design
Trained on signal, serves only numbers. The model runs where the data lives.
Across modalities
Text, audio, video, world. One unit of measure across all of them.
Schematic
Same model. Same frame. Four signals.
01 · dialogue, prose, code
Text
Multilingual exchange measured directly. The same six-axis frame across English, Portuguese, Spanish, Arabic, Japanese.
02 · speech, dialogue, calls
Audio
Audio enters the same model unchanged. Per-actant readings carry across paired exchanges, with no per-modality fine-tuning.
03 · demonstration, behaviour
Video
Video enters the same model unchanged. Zero-shot readings on held-out task pairs, with no per-domain fine-tuning.
04 · agents in a generated environment
World
Trajectories through procedural worlds become readings. The same instrument scores reasoning, choice, and recovery in real time.
one unit of measure · one six-axis frame · four signals
See for yourself
One protocol. Four modalities. Same instrument.
The spectral field below is what the model sees. The three traces are the named axes of the protocol, projecting through it. The numbers on the right are what comes out.
live walkthrough · protocol = client-pressure / 3 axes
text · dialogue
at playhead 0.74
captured energy ∈ [0, 1]
0.78
Area under the three axis ridges. Everything else is energy outside the protocol.
fidelity ∈ [0, 1]
0.82
Share of the field the protocol explains. Below the coverage floor: no score reported.
resonance · per axis A1 · A2 · A3
actant flow F(A → B)
A
B
0.62
urgency · A drives
input
note A reading on text. Discrete turn-bursts shape the field. Energy concentrates in urgency framing, fidelity above the coverage floor.
illustrative · procedural readings · no raw content shown
A new raw primitive
What comes out is a new primitive of interaction.
Every reading carries five surfaces, derived from the same projection. Together they form one composable measurement object, ready to read, audit, and replay.
sgm reading · session #4f29
protocol
client-pressure / 6 nuances
fidelity
0.78
01 · resonance
segments × nuances
04 · fidelity
∈ [0, 1]
0.78
captured energy
coverage floor · τ = 0.20
02 · actant profile
C × B × I · per actant · per turn
05 · causation
directed · per axis · time-lagged
Signed flow per axis, on the geometry, with a measurable lag. A1 leads in one turn, A2 trails by three. The graph is read from the structure.
illustrative · mock data · no raw content shown, ever
Resonance
[segments × nuances]
How strongly each segment of the exchange activates each named direction in the protocol. Eigenvalue-weighted, sums to 1 per segment.
Actant profile
C × B × I over time
Cognitive, Behavioural, and Intentional position for each actant, evolving turn by turn. The shape of how one actant occupies the interaction.
Spectral flow
F(A → T)
Directional asymmetry between actants. Who drives which dimension, and how much. A signed scalar per pair, per axis.
Fidelity
∈ [0, 1]
The share of the signal that landed inside the protocol. Below a coverage floor, the instrument reports no score. Honesty by construction.
Causation graph
directed · per axis · time-lagged
Directed edges between actants, signed and asymmetric, decomposed per protocol axis and resolved with a temporal lag. Causation is read on the geometry over time, from structure.
The path
Toward a unified geometry across all interaction.
Each version makes the geometry broader. The horizon is one instrument that reads every interaction worth measuring, in the same units.
0.1
0.2 · now
horizon
0.1 Principle proven
One modality, deterministic.
The first version established the construction on text-based interaction. Deterministic readings, glass-box traceability, protocols added at measurement time rather than retrained in.
- · deterministic
- · glass-box
- · protocols at measurement time
0.2 In deployment
Multilingual, multimodal.
The same model now reads audio and video alongside text, and runs zero-shot across multiple languages. The geometry is shared across modalities, so a protocol authored once applies anywhere the model reads.
- · text · audio · video
- · multiple languages
- · shared geometry
→ The horizon
The unified geometry across all interaction.
The construction is built toward a single geometry that reads every interaction worth measuring, in the same units. Modality and language become projections through the same learned space.
- · unified spectral model
- · all signals
- · one unit of measure
Latest research
What the readings look like.
Four readings from the latest research, with full methodology and falsifying controls in the public manuscript.
-
Result · 01
Six-axis spectral emergence
The top six eigendirections each top-1 align to a distinct named primitive at orthogonality residual 5.9 × 10⁻¹⁶, mean alignment 0.472. Latent structure concentrates into a measurement-grade subspace.
-
Result · 02
One unit of measure, three modalities
The spectral model extends, unchanged, from text to audio to video. Per-actant readings stay stable across paired audio at 0.997; zero-shot reading of 0.86 on held-out video task pairs (n = 68).
-
Result · 03
Cross-lingual, zero-shot
Multilingual readings hold across nine languages with a zero-shot mean of 0.904, and across five language families with centroid agreement of 0.975.
-
Result · 04
Reproducibility by construction
Eigenbases are bootstrap-stable at 0.998 / 0.988 / 0.96 across 200 resamples (text / audio / video). The eigendirections themselves are reproducible measurement objects.
By construction
Deterministic. Zero-shot. Glass-box. Real-time.
Four properties held jointly, in the same instrument, on every reading.
01
Deterministic
Same input, same protocol, bit-identical output.
02
Zero-shot
A new protocol is a closed-form rank-1 update on the model.
03
Glass-box
Every reading traces back to specific segments and specific evidence.
04
Real-time
Sub-25 ms per reading on CPU. Hundreds of thousands per second per node.
Build on it. Measure with it. Ship under it.
Get in touchAt civilisational scale
Measurement, in service.
The SGM is the instrument. PKnet is the civic federation it operates inside. The same model put where it is needed, under consent, with custody held by the public — so the measurement reaches the population it is meant to protect, and the benefit of the technology stays with the people who fund it.
The public deployment
PKnet — the consensus mesh
consenting protective gateways · published protocol registry · per-citizen legibility · ε-bounded aggregation