0.875
PhysioNet AUC
Models · Research preview
Time is an input,
not an assumption.
Continuous-time model for irregular signals — small, fast, and time-aware.
≈5×
Smaller
8.2×
Faster kernel
Category
Continuous-time · generation 1
Reads
Irregular, time-stamped signals
Best for
Clinical, sensor, streaming & edge
Stage
Research preview · gate-validated
What RL-L1 is
A model that treats time as a first-class input
Most sequence models assume one step equals one tick. Real-world signals — vitals, sensors, market events, telemetry — do not arrive on a clock. RL-L1 is a continuous-time model: timing is part of the input, not an afterthought, so the model behaves correctly across irregular gaps, missing samples, and live streams. The how is our own — what we show here is the behaviour and the numbers.
Why it is different
Three structural choices
RL-L1 is not a transformer and not a point-ODE network. It makes three deliberate bets that pay off on real, irregular data.
L1
Time is in the math
The model evolves in continuous time, so the interval between two observations changes the computation. Zero out the timing and accuracy collapses — proof that the time signal is doing the work.
L2
Small and stable
A compact, well-behaved state means strong results at tiny parameter counts, with bounded, predictable dynamics. The internals are ours; the stability is measured.
L3
Built for the edge
Designed to run fast and tiny on real hardware, and co-designed with our dissipative silicon line so the model and the substrate converge on the same physics.
Measured, not asserted
The numbers, including where we lose
RL-L1 wins on irregular, small-to-medium-scale signals and on the edge. It does not win at frontier scale or at language — the table says so on purpose.
Benchmark RL-L1 Baseline Note
PhysioNet 2012 mortality 0.875 AUC 0.874 (GRU-D) 18-seed ensemble · CI [0.868, 0.883]
Timed associative recall 0.004 MSE 0.68 MSE vs point-ODE; 9.61 if time-blind
Mackey-Glass (≤50k params) 0.052 MSE 0.094 MSE 1.2–1.8× better, small-model regime
Mackey-Glass (≥214k params) 0.087 MSE 0.046 MSE transformer wins at scale
Edge footprint (int8) 24 KB · 0.40 ms 65 KB · 0.44 ms faster and 2.7× smaller
PhysioNet 2012 · set-a → held-out set-b · leakage-safe · multi-seed
Where it wins
- Irregular & streaming time series — timing carries the signal.
- Energy- and memory-constrained edge inference.
- Online adaptation without a full retrain.
Where it does not
- Frontier-scale language — transformers win, and we say so.
- Large-parameter regimes where the efficiency edge fades.
Measured
Smaller, and time-aware
0.875
AUC on PhysioNet 2012
In-hospital mortality from irregular ICU vitals — 18-seed ensemble, 95% CI [0.868, 0.883].
≈ 5×
Fewer parameters than the baseline
The clinical result is reached at a fraction of the baseline Transformer's parameter count.
8.2×
Faster runtime kernel
Parallel-field evaluation vs the reference loop (0.23 ms vs 1.90 ms), at exact numerical parity.
Status
A gate-validated reference
- Gates
Ten falsifiable gates passed
Continuous-time advantage, stability, binding, lifelong retention, and substrate co-design each have a pass/fail test. They pass — and where the model should lose, it loses honestly.
- Clinical
Validated on real ICU data
Benchmarked on PhysioNet 2012 in-hospital mortality — real, irregular vitals — with a leakage-safe train/test split.
- Today
Research preview
Reference implementation is production-tested on Apple Silicon. We are opening it to selected research partners.
Where RL-L1 fits
Built for signals with a clock
Irregular vitals & monitoring
ICU streams, wearables, and clinical time series where samples are sparse, uneven, and timing is the signal.
On-device & streaming
Tiny footprint and fast inference make RL-L1 a fit for sensors and embedded systems that cannot afford a large model.
Honest continuous-time evals
A reference for studying where continuous-time models genuinely beat transformers — and where they do not.
“It wins where timing matters. We say so where it does not.”
— RL-L1 design note
