Research notebook / 18 October 2026

The observer
is part of the event.

What the next step costs

Our small simulated world contains six coordinates: three for the surroundings and three for the observer’s body, resources and access. Iskra learns their numerical changes from observed outcomes. The environment supplies the coordinate meanings, connections, action vectors and availability thresholds; these were not discovered by the model.

We continued the existing core in three controlled trials: 504 learning steps and 324 evaluation transitions. All three met the criteria set before execution, compared with a frozen observer and the inherited knowledge before the new experiences. These are three parameter settings of one family, not evidence of universal understanding.

The saved model predicted that two consecutive probes would make a third unavailable. A probe, resource renewal and another probe remained possible. We recorded both forecasts first, then executed copies of the same world. The availability and lengths agreed with the dominant predicted branch.

What memory still does not do

A separate control used twelve new contexts and 108 actual transitions. We froze only the hypothetical branch memory. Every physical path remained exactly the same, despite changing weights in the full version. The criterion of a five percent improvement failed in all three trials.

Each path keeps one mechanism and follows actions supplied in advance. Updating its beliefs does not yet change those actions. The next capability to build and test is learned action selection from acquired information. The earlier return-to-a-familiar-world criterion also remains unmet.

What is running here

The public live view uses an earlier checkpoint. Each new biography starts from the same saved parent; learning continues within that biography. The newer observer trials are separate, explicitly identified research continuations. The interface is not evidence that the public model already has their capabilities.

We preserve forecasts, outcomes and failed criteria. Source protocols and reproducible evidence are stored with the project. The goal remains a core that learns which experience to seek and how to use what it discovers.

Previous research entry: returning to a familiar world →