Evidence-driven adaptation
A research program about what models should remember, what they should change and how to tell the difference.
Compare prediction and observation.Test proposed updates.
Sources of prediction error
A wrong prediction does not always mean a model needs new knowledge. Conditions may have changed, a familiar situation may have returned, or the measurement may be noisy.
EOS asks a simple question: what should a model check before changing what it remembers? Its proposed design separates making a prediction from deciding to keep a change.
- 01
Make the prediction first.
Use the current model and earlier learning to make a prediction. Keep it separate from the observation that will check it.
- 02
Check what you already know.
Try earlier learning and simpler explanations. Check whether noise or an incomplete view of the situation explains the difference.
- 03
Ask what would help you decide.
Choose a useful observation or experiment, and consider its cost. Permission to investigate is separate from permission to keep a lasting change.
- 04
Test a change before keeping it.
Keep a proposed change separate while it is tested. Use fresh evidence and state where it should apply. If the evidence is missing, leave the decision open.
Evaluating candidate updates
The proposed system keeps three things separate: a prediction, the result that checks it and the decision to keep a change. The full system has not been demonstrated.
Conceptual research illustration. The drawing represents the proposed architecture.
Research status
EOS is a proposed design. The papers describe the idea and early studies of specific parts. They do not show a complete working system.
- Architecture hypothesisA proposed system combining prediction, memory, reasoning and control over persistent change.
- Component studiesFormal results and constrained experiments. Each manuscript states its assumptions and limits.
- EOS-F1An ongoing small-model and systems research track. No public model release is available.
- Next stepsLearned nonlinear worlds, matched baselines, integrated evaluation and independent reproduction.
EOS publications
- 01
Evidence-gated causal world models
Architecture hypothesis · Sonne SystemsThe proposed architecture, interfaces and experiments needed to assess it.
- 02
Dual governance for continual agents
Governance study · Sonne SystemsWhy reversible information gathering and persistent model changes need different authority.
- 03
Consequence-certified state refinement
Related methods paper · Sonne SystemsWhen a partially observed model can merge histories, refine its state or consider new structure.
EOS is presented within 2πarc. The listed manuscripts were written by Aman Agarwal at Sonne Systems and retain that attribution. Titles and summaries are public. PDFs in the Rexzen directory use the shared password.