Neuro-symbolic runtimeruntime neuro-simbólico

Probabilistic models guess.
This verifies the guess.

Symbolic Engine turns image, audio and video signal into symbols, then checks decisions by hyperdimensional similarity in microseconds — no training, no GPU, no black box.

Los modelos probabilísticos adivinan. Esto verifica la respuesta — sin entrenamiento, sin GPU, sin caja negra.

Run the live arena Contact

What it is

A resolver, not a model

Signals are encoded as 10,000-bit hypervectors. Decisions come from symbolic similarity between those vectors — not from weights adjusted by gradient descent. Las señales se codifican como hypervectores de 10.000 bits. Las decisiones salen de similitud simbólica, no de pesos ajustados por descenso de gradiente.

// example trace, abbreviated
encode(frame_1882) → sym:0x7fa2…c91 [2.1ms]
resolve(sym, policy_set) → action:HOLD [0.08ms, hamming=412]
verify(action, executed) → match:true [auditable]

Bit-level similarity means every output is traceable back to the input that produced it — nothing is interpolated. La similitud a nivel de bit significa que cada resultado es trazable hasta el input — nada se interpola.

One-shot, three senses

Vision, hearing, video — no training set

Same resolver, three input types. Each demo below is a real recorded run, not a mockup. Mismo resolutor, tres tipos de entrada. Cada demo es una corrida real, no un mockup.

Vision

One-shot shape classification by encoding position and intensity as a hypervector.

Clasificación de formas one-shot, codificando posición e intensidad.

100% accuracy, translation + noise
Watch the video →

Hearing

Encodes band energy over time; recognizes tone, noise and sweeps.

Codifica energía de banda en el tiempo; reconoce tono, ruido y barridos.

2 examples per class
Watch the video →

Video delta stream

Reads symbols emitted by a sparse video codec — no raw pixels.

Lee símbolos de un codec de video disperso — sin píxeles crudos.

one-shot: calm vs. explosion
Watch the video →

Live proof

Symbolic Arena

Twenty autonomous bots share a single symbolic arbiter, running on a free-tier CPU. No GPU involved. Watch it decide in real time, then read the trace behind any decision. Veinte bots autónomos comparten un árbitro simbólico, en CPU de capa gratuita. Sin GPU.

AgentsAgentes

~60

Ticks / secondTicks / segundo

GPU utilizationUso de GPU

100%

Auditable traceTraza auditable

Open the arena

Measured, not claimed

What the lab actually shows

Internal benchmarks (HDC/VSA substrates, D=10,000, FHRR & BSC), reproducible, methodology shared on request. Benchmarks internos, reproducibles, metodología disponible a pedido.

Zero catastrophic forgetting Learn 200 associations (A), then 200 new ones (B): accuracy on A stays at 1.000 after learning B. Cero olvido catastrófico — aprende 200 asociaciones, luego 200 nuevas, y la precisión en las primeras se mantiene en 1.000.

~2.7–20µs per operation Update and inference cost at D=10,000, on CPU — no GPU, no batching required. ~2,7–20µs por operación — costo en CPU, sin GPU ni batching.

800 associations before first drop At D=10,000, a single hypervector holds up to 800 associations at ≥0.99 accuracy before degrading. 800 asociaciones antes de la primera caída — un hipervector sostiene hasta 800 con ≥0,99 de precisión.

Adapts in ~3 interactions, keeps 100% of old skill When an opponent switches strategy mid-game, the delta rule re-stabilizes in a median of 3 interactions while retaining prior behavior in 100% of episodes. Se adapta en ~3 interacciones y conserva el 100% del skill anterior.

Reward reversal: 3–4 trials vs. 53+ When the reward for a learned context flips sign, the delta rule re-adapts in 3–4 trials — plain Hebbian with no decay takes 53+. Reversión de recompensa: 3–4 intentos vs. 53+ con Hebbiano puro.

Perfect collision detection, sub-pixel decode Continuous 2D space in unit phasors: collision detection at r=4px hits F1=1.000 in 6–14µs, position decodes with ~0.17px median error. Detección de colisión perfecta (F1=1.000) y decodificación con error mediano de ~0,17px.

Where it fits

Robotics, edge, anything audited

Robótica, dispositivos edge, cualquier cosa que deba ser auditada.

Edge & robotics

Real-time decisions on hardware that can't carry GPU cost or cloud latency — the resolver runs where the signal is.

Decisiones en tiempo real en hardware sin GPU ni latencia de nube — el resolutor corre donde está la señal.

Verifiable systems

An LLM can stay the natural-language front end. The symbolic core checks that what gets executed matches what was verified.

Un LLM sigue siendo la interfaz en lenguaje natural. El núcleo simbólico verifica lo ejecutado.

Performance figures above (vector width, tick rate, GPU usage) are our own benchmarks, published as-is — happy to walk through methodology on request. Cifras propias, publicadas tal cual — con gusto explicamos la metodología si lo piden.

Probabilistic models guess.This verifies the guess.