AI that keeps the science in context.
Large language models are powerful scientific tools, but they are weak at sustained research. Over long investigations, objectives drift, assumptions disappear into conversation history, unresolved questions are forgotten, and fluent answers can outrun the evidence. Next gen AIs are investing to eliminate and reduce scientists to cut expenses, we're looking to augment and enable them.'
Cottage Science is building the system around the models: persistent scientific context, domain knowledge, executive control, and agentic review.
Memory is not the same as context.
Scientific research is not a long conversation. It is a structured investigation. The system has to know what the hypothesis is, which assumptions it depends on, what has been established, what remains uncertain, what evidence conflicts with the current model, and why earlier decisions were made. Too much time is wasted on on unnecessary token output, scientists spend more time reading, vetting and clarifying current LLM output. Adding to a scientist's workload.
Cottage Science treats the research state itself as a persistent object: hypotheses, assumptions, evidence, constraints, objections, dependencies, decisions, and unresolved questions.
Stay with the research.
The system should preserve the objective across a long investigation, recognize when new evidence changes earlier conclusions, keep unresolved work visible, and determine what matters next. That is not a larger context window. It is scientific executive function.
Maintain the objective
Keep the scientific question and completion criteria intact across extended work.
Track unresolved work
Preserve unanswered questions, weak assumptions, failed tests, and open branches.
Update the research state
Recognize when new evidence invalidates or changes earlier conclusions.
Choose the next useful test
Prioritize work that can reduce uncertainty or discriminate between competing explanations.
Challenge the science, not the scientist.
A useful scientific system should be cooperative about the research goal and skeptical about the evidence. Constant disagreement is no more useful than constant agreement. The objective is relevant, calibrated, and actionable criticism.
Agentic Referee Feedback
Challenge methodology, assumptions, inference, and unsupported conclusions.
Verify
Recalculate and reproduce important mathematical, statistical, and computational claims.
Agentic Red Team
Search deliberately for failure modes, contradictory evidence, and alternative explanations.
Preserve momentum
Surface objections that advance the work instead of generating disagreement for its own sake.
Built around the structure of scientific reasoning.
Domain Knowledge
Science-prioritized training and adaptation centered on technical literature, mathematics, methods, code, data, and discipline-specific meaning.
Context Integrity
Preserve what information means inside the investigation and distinguish what is known, assumed, inferred, disputed, or unresolved.
Scientific Executive Function
Maintain the research objective, manage open work, preserve decision history, detect drift, and keep the investigation moving coherently.
Adversarial Review
Use independent agentic referees, verification, and agentic red-team processes to challenge the work.
From unconventional idea to competitive cosmology.
Cottage Science is currently testing its approach on a difficult problem in observational cosmology.
We started with an unconventional dark-sector hypothesis and worked it through the research process: equations, numerical implementation, observational constraints, model criticism, and MCMC analysis.
No additional free parameters were introduced.
For Cottage Science, it is a useful test of whether this architecture can support serious scientific work over a long research cycle.
Using the same number of free parameters as standard flat ΛCDM, the model improved the best fit by approximately Δχ² = 17.
Cottage Science
Strategic inquiries.