Evidence between attempts.
The internal prototype connects hypotheses, code changes, experiments and scientific memory. Retained run reports describe earlier findings shaping later research.
RECURION / INSIDE RESEARCHOS
From idea.
To evidence.
To the next idea.
Research software that connects what to investigate, how to test it, and what the result changes.
Explore the approach01 / THE THEORY ATLAS
Give new ideas
a scientific basis.
The atlas maps mechanisms, assumptions, predictions and tests. Research agents use it to develop competing explanations—not just retrieve related text.
For example, our point-cloud atlas describes how noise can make local geometric matches unreliable. That suggests different matching methods, uncertainty-aware designs and targeted tests before a costly training run.
This is an example of a theory-derived research direction, not a claim that the proposed intervention already improves the pilot model.
Mechanism→Prediction→Test
02 / SCIENTIFIC CONTINUITY
Keep the reasoning.
Not only the result.
Each research attempt connects the hypothesis, implemented change, evaluation and interpretation.
The structured experimental record feeds the next research request. Findings can redirect the investigation, narrow a claim, or leave an explanation unresolved. An unsuccessful experiment is useful only when its evidence informs a better decision.
The broader ambition is continuity across people, teams and projects. Useful transfer at those boundaries still needs to be demonstrated.
Evidence→Revision→Next attempt
03 / WHAT EXISTS TODAY
Working loop.
Continuing validation.
A 3D reconstruction pilot.
The next application investigates model improvements for reconstructing detailed shapes from sparse point-cloud data.
Value beyond an internal run.
Customer productivity, cross-team continuity and better research methods require their own comparative evidence. No blanket speedup or performance guarantee is claimed.
04 / RECURSIVE IMPROVEMENT
Make the method
an experiment.
First improve a model. Then test better ways to generate hypotheses, choose experiments and interpret evidence.
ResearchOS’s method-improvement infrastructure is distinct from proving a better researcher. Proposed methods must earn that claim on independent research problems.
05 / BEFORE A PILOT
A practical
starting point.
What problem is a good fit?+
A valuable model or algorithm, a measurable improvement goal, a repeatable evaluation and a team able to review the results. AI and robotics workflows are the initial focus.
What would the customer receive?+
A proposed paid pilot covers scoped experiments, tested changes, evidence and a justified next decision. We also want to test whether another researcher can continue from the scientific record. Scope, pricing and delivery are agreed individually.
How is this different from a coding agent?+
Coding tools are part of the workflow. ResearchOS also addresses what to investigate, the scientific reasoning behind the change, how to evaluate it, and what the result should change in the next attempt. Other AI research systems overlap with this work; comparative value must be measured.
Who controls execution and adoption?+
The team sets the goal, resource limits and approvals. Test results do not automatically authorize deployment. Customer deployment, data access and ownership requirements must be agreed before a pilot; no blanket security certification is implied.
Does every result become accepted knowledge?+
No. Observations, interpretations and accepted claims remain distinct. Uncertainty, failed assumptions, contradictory evidence and limits on transfer matter as much as positive findings.
START WITH YOUR SYSTEM
A question
worth answering.
Bring the model.
Let’s define the experiment.