Active development / October 2026
Building AI systems that can discover tools, test ideas, and carry useful knowledge forward.
Teledra is a research and development program exploring recursive self-improvement (RSI) through reusable software, causal reasoning, and controlled evaluation.
Current focus: AI systems development for Shipyard AS.
● A proposal becomes progress only after it is checked.
The research question
The program studies a practical route toward RSI: discover what already exists, identify a real limitation, make a bounded change, and evaluate whether it helped.
Today’s work consists of software tools, experimental reasoning components, and evaluation workflows. RSI is the research direction; a generally capable, autonomous self-improving system remains a goal.
Changes are developed in isolation, measured against defined tasks, and reviewed before adoption. Failed experiments and unresolved questions remain part of the record.
Current workstreams
Three connected areas make the research concrete. Each has its own implementation and evidence requirements.
A local toolkit shared by agents and developers through CLI and MCP interfaces. It supports discovering and reusing maintained operations for files, documents, images, data, and repository inspection.
Experimental grid perception, temporal comparison, action-effect learning, prediction, and bounded experiment selection. Observations, hypotheses, and uncertainty are kept distinct.
Research into actor identification, conditional action rules, and experiments that distinguish competing explanations. The next stage investigates committed, multi-action experiments.
Evidence over promises
A result is useful when its task, measurement, and limitations are visible. Strong performance on a narrow test is one piece of evidence.
One frozen evaluation measured action-rule discovery in 48 generated environments. These are internal research results, with a fixed action budget and defined scoring criteria.
Read the comparison and limitations →This evaluation does not establish a general ARC-AGI gain or a leaderboard result. The comparison also does not establish that information scoring outperforms its ablation.
How progress is judged
Reusable work should be easier to find, safer to inspect, and possible to reproduce.
Use existing capabilities where they fit. Record meaningful gaps instead of repeatedly generating the same helper.
Choose experiments that distinguish hypotheses. Charge setup actions, failed probes, and follow-ups to the same budget.
Separate observations from inferred rules. Keep ambiguous results and failure cases visible.
Freeze evaluation conditions before holdout testing. Permissions, deployment, and capability promotion remain explicit decisions.
The Teledra creative identityPeople & company
Sveinung Ramse
Co-founder of Shipyard AS
Current status: developing AI systems for Shipyard AS. The company is owned by Younes Louragli, who is also its registered chair.