We build intelligence from advanced dynamics.
Phase-TITAN (PT) is our post-transformer architecture for representing complex systems as evolving state across machines, language, and interactive worlds.
Our post-transformer architecture
Phase-TITAN represents intelligence as evolving state.
PT puts this understanding into practice by continuously evolving compact state from observations, tokens, actions, and events.
State
What persists.
Transition
What changes.
Stability
What holds, deviates, and recovers.
Phase-TITAN applications
Phase-TITAN powers intelligence across machines, language, and interactive worlds.
PT Training-Free Dynamics Monitor
With no trained weights, it estimates normal dynamics and calibrates thresholds online for each machine.
PT Language Model
Persistent language intelligence without transformer layers or a token-growing KV cache.
PT World Model
A generative world whose stable dynamics were learned from independent random-seed gameplay traces and persist under live interaction.
PT Training-Free Dynamics Monitor
Normal dynamics, estimated live for each machine.
The PT Training-Free Dynamics Monitor has no trained weights. It estimates each machine's normal dynamics and calibrates its thresholds live, without pretraining or failure labels.
Watch every frame from one recorded fixed-wing flight. The external fault annotation is shown for auditability but never provided to the monitor.
Normal dynamics are estimated from the machine's own stream. No labeled failure set, training run, or fitted fault model is required.
The monitor confirmed a sustained change across several consecutive frames after the recorded engine power-loss annotation. No future sensor values were used.
The fixed-size, heap-free Rust decision path is built for constrained onboard targets without cloud connectivity.
This is a causal replay through the governed Python reference runtime, not a live flight or target-board timing claim. The source is the public CMU AirLab ALFA dataset.
Road vehicles and fixed-wing aircraft produce very different signals. We adapted how each sensor stream enters the monitor while leaving the monitoring logic unchanged.
How the monitor works
Estimate normal dynamics live. Flag departures.
Live dynamics estimate
Model weights are neither trained nor updated. The normal-dynamics estimate and alert threshold adapt continuously to the incoming stream.
When observed behavior departs from the estimated normal dynamics, the monitor surfaces the change.
Matched or exceeded the Matrix Profile baseline across four real ALFA fixed-wing fault types, measured by offline VUS-PR after detector warm-up with 2–23 flights per type.
Road vehicle: 4.8%. Fixed-wing UAV: 4.2%. Both were calibrated online from one shared risk setting.
This is reproducible offline experimental evidence, not a customer performance guarantee or hardware qualification.
PT Language Model
Long-context intelligence built for on-device operation.
PT Language Model v5 replaces token-growing context storage with compact evolving state.
No transformer layers. No KV cache. State size does not grow with context length.
Measured in the evaluated v5 1B runtime; excludes model weights and temporary runtime memory.
Eleven-task panel spanning retrieval, tracking, aggregation, and multi-document QA. Scores for every model exclude two multikey tasks because v5 was not trained on those formats.
All 4,441 BFCL-v3 cases, evaluated under strict word-10-gram decontamination.
PT Language Model v3.5 · 3B on Galaxy S26 Ultra
An earlier PT Language Model running locally on a consumer smartphone.
Watch on YouTubePT World Model
A world is not a longer video.
To our knowledge, the PT World Model is the first demonstrated persistent interactive video world model to sustain a learned world as a stable dynamical attractor in a fixed-size, attention-free temporal state.The attractor was extracted from gameplay traces recorded under independent random game seeds, not from one continuous run.
Starting from a short observed seed, the PT World Model generated an evolving kart world under live steering, throttle, and braking for approximately 33 minutes. The driver ended the recording; the model did not reach a temporal horizon.
The session ended when the driver chose to stop, not at a model-defined limit.
World-step count reached on the final branch during the continuous recorded session. Steps from abandoned branches are not added to this total.
Repeated traversal of the generated circuit under live user control.
Measured live in this research build. Excludes model weights, codec caches, and runtime workspaces.
The world persists because its dynamics are learned.
Training used separate gameplay traces initialized from independent random game seeds, not one continuous run. From these traces, the PT World Model learned recurring circuit structure and dynamics without reducing each interactive session to a replay.
Temporal state, not temporal context.
Token-level temporal attention must either retain a growing KV history or discard context beyond a finite window. The PT World Model instead carries its evolving world in a fixed-size PT state.
The world exists as operational state.
The world's complete neural state can be restored and evolved directly. The same state and commands reproduce the same continuation; different commands generate a different future.
Demonstrated at 512 × 512 in a kart domain. Fixed circuit geometry is rendered using a pose estimate derived from the model's own PT state. Generation receives neither future frames nor live ground-truth simulator state.
Work with TraceWeave
Build with us.
Discuss a pilot in machine intelligence, on-device language systems, or interactive worlds.