Iskra · research manifesto

We are building Iskra’s mind.

We are developing memory, learning and the ability to anticipate consequences. Each next step must be supported by results.

01 / 05One learned core

Many situations. Shared understanding.

We want one core, trained from scratch, to recognise relationships and reuse what it learns in unfamiliar situations. Programs, processes and simulated worlds can provide that experience.

A shared interface is not enough. The learned rules themselves must transfer.

02 / 05Memory and relationships

A situation includes its history.

An object’s state, its relationships, earlier events and the observer’s available information all matter. Two similar observations may conceal different causes.

Iskra’s internal model should retain what helps distinguish those situations, anticipate their changes and express uncertainty where the evidence is incomplete.

03 / 05Learning from consequences

Predict first. Then check.

Record the information available before an action. Predict its consequences. Run it in a controlled environment and compare the result with that prediction.

Where a world can be copied, test alternative actions from the same starting point. The difference between predicted and observed outcomes provides the learning signal. A model’s own forecast never counts as confirmation.

04 / 05Independence

Learned from experience, from the ground up.

The new core is being designed to learn and operate independently of large language models. Its training data, predictions and evaluation must not depend on an LLM’s answers or judgements.

We control the code, data and weights. Remote training supplies compute; it does not supply the model’s understanding.

05 / 05Evidence of progress

The outcome decides.

We will measure whether Iskra predicts better than simpler methods, distinguishes the effects of different actions and uses earlier experience in a new environment. Failures and uncertainty belong in that assessment.

This is a research direction, not a claim of consciousness or a finished universal model. Broader capability must be demonstrated in new situations.

“Build an internal model of the world. Let the world test it.”

Memory · relationships · understanding

Stay close