Wakeline

Research

What we have measured, and what we argue from first principles.

Everything we publish carries one of three labels, so you can tell a measurement from an argument at a glance.

Start here

One architecture, the same principles, different environments

Most AI systems that work across several fields are several systems, each designed and trained for its problem. Wakeline's AI is one architecture.

  • Learning is part of operation. There is no pretraining and no retraining cycle; each instance starts with an empty memory.

  • One component per stream. Each input stream has its own component that learns to forecast it, and nothing is trained centrally.

  • Experience is kept as memory. Forecasts and decisions are read from recorded situations and what followed them.

  • Consequences are the teacher. The next observation grades the previous prediction. To act, the system needs only a preference for larger outcomes.

  • Change is handled in memory. Old experience loses weight as new experience arrives, with no reset and no retraining.

  • No neural networks, no backpropagation, no GPU. Hundreds of live streams run on one CPU server.

Production

Evidence from continual learning systems running in production.

SystemEnvironmentLive sinceScaleStatus
Market Edge Wholesale power market April 2026 83 live streams Running

Empirical

Experiments with numbers, reported with their method and limitations.

Learning in Worlds

Showing 7 of 7 studies

DateTitleEnvironmentCompared againstHeadline resultTags
The CLS Forecasting Benchmark on the 15-Minute Market Day-ahead power prices, 15-minute resolution TabPFN 2.5 91.4% of the perfect-foresight battery revenue vs 89.8% for TabPFN 2.5 over the first year of 15-minute trading
  • Energy
Forecasting Traffic Without an Offline Training Phase Real road sensors (PEMS-BAY, METR-LA) DCRNN, Graph WaveNet PEMS-BAY: lower mean absolute error than published DCRNN and Graph WaveNet at every horizon; METR-LA within 1.9–3.6% of the DCRNN checkpoint
  • Traffic
Learning the Shape of a Rule It Was Never Shown Simulated stock market, shaped rules Tabular Q-learning Recovered every committed rule shape (exemption, kink, cap, surcharge band) on three unseen seeds; tuned Q-learning failed the entry test
  • Finance sim
Learning while the world changes Simulated road network Its own control arms With forecast input, 29 seconds saved per journey on average against a matched learner without it (8 runs, 3,200 scored journeys)
  • Traffic
Keeping a Learned Rule Current When Nobody Announces the Change Simulated stock market, changing rules Q-learning Registered every unannounced rule change and re-acquired each new rule (medians 70 and 294 trading days); Q-learning variants missed 3 and 14
  • Finance sim
Learning a Hidden Cost Structure from Outcomes Alone Simulated stock market Tabular Q-learning Recovered a hidden 25% capital-gains tax schedule from outcomes alone, within about 6% at every bounded level of gain
  • Finance sim
The CLS Forecasting Benchmark: Continual Learning Measured in Euros Day-ahead power prices, hourly resolution TabPFN, Prophet, NeuralProphet ~14% higher cumulative trading P&L than TabPFN over 2,097 trading days
  • Energy

Theoretical

First-principles arguments and definitional work, with no new data.

Foundations

Continual learning defined

Drift and change

Uncertainty

Architecture and safety