INFRASTRUCTURE
AgentKit
A local toolkit for discovering and reusing maintained operations through CLI and MCP interfaces. It supports work with files, documents, images, data, and repositories.
Implemented components · continuing developmentIndependent AI research & development
AI systems.
Measured
progress.
Exploring how agents can use tools, learn from actions, and improve through experiments.
Led by Sveinung Ramse · Developing AI systems for Shipyard AS
THE RESEARCH APPROACHObserve → test → learn
From an observation
to a tested explanation.
to a tested explanation.
Explicit uncertainty Repeatable evaluation
Featured research / recurrent memory
RNNs deserve
a closer look.
Recurrent neural networks carry a changing internal state through a sequence. They give our experiments an important conventional baseline for questions about memory, timing and learning.
In a recent middle-pulse task, a better-tuned RNN outperformed the tested Lizard variants. That result corrected an earlier comparison and shows why strong baselines matter.
Explore the RNN research and its limits →A baseline changed the picture
Middle-pulse accuracy · 20 fresh seeds · higher is better
Lizard + ratchet74.4%
Lizard + order trace75.9%
RNN · 64 units78.5%
RNN · 256 units79.2%
01 Reusable agent tools
02 Causal learning
03 Controlled evaluation
01 / Research direction
How can an agent improve
the way it works?
Teledra studies a practical route toward recursive self-improvement: discover what exists, identify a limitation, make a bounded change, and measure whether it helped.
The work brings together reusable software, causal reasoning, and experimental evaluation. Today’s outputs are tools and research components; a generally capable, autonomous self-improving system remains a goal.
Progress means a result that can be inspected, reproduced, and understood within its limits.
02 / Current work
Connected lines
of inquiry.
From the tools an agent can find to the experiments it chooses to run.
PERCEPTION & LEARNING
ARC-SENSE
Experimental grid perception, temporal comparison, action-effect learning, and prediction. The system keeps observations, hypotheses, and uncertainty distinct.
Implemented components · bounded evaluationsSCIENTIFIC EXPLORATION
Learning by testing
Research into actor identification, conditional action rules, and experiments that distinguish competing explanations. Follow-up prototypes investigate multi-action experiments and the validation of competing explanations.
Experimental · active researchCODEBEE / CONTROLLED ADAPTATION
Learning that lasts
Checkpoint geometry, update interactions, and candidate tool-use habits. Current evidence spans early model studies and CPU controls; learned-adapter benefits remain open research questions.
Experimental studies · design review · proposed follow-ups03 / Selected research record
Evidence you can
put in context.
Results are reported with the task, measurement, and limitations attached.
ARC-004 / INTERNAL GENERATED-WORLD EVALUATION
Read the research record ↗Scientific explorer
One frozen evaluation of action-rule discovery in 48 generated environments, with a fixed action budget and defined scoring criteria.
48generated holdout worlds
189correct supported action mappings
5 / 48tasks completed by the scientific policy
0false supported mappings under the stated criterion
Preserved 23 October internal evaluation, not an official ARC-AGI score. The comparison does not establish that information scoring outperforms its ablation. Read the boundaries →
04 / Working principles
A clear question.
A checkable result.
Useful research leaves a trail from the proposed change to the evidence behind it.
Discover before rebuilding 01
Use existing capabilities where they fit. Record meaningful gaps instead of repeatedly generating the same helper.
Test competing explanations 02
Choose experiments that distinguish hypotheses. Charge setup actions, failed probes, and follow-ups to the same budget.
Preserve uncertainty 03
Separate observations from inferred rules. Keep ambiguous results and failure cases visible.
Review before adoption 04
Freeze evaluation conditions before holdout testing. Permissions, deployment, and capability promotion remain explicit decisions.
05 / People & affiliation
Research with
a person behind it.
Sveinung Ramse
Co-founder of Shipyard AS
Leading the Teledra research and development program. Currently developing AI systems for Shipyard AS.
sveinung@teledra.dev ↗Get in touch ↗