Ask a clear question.
Follow the evidence.
Explore how SlothLab turns hypotheses into measured experiments—and how LazyLoop decides what is worth running next.
Prior correction under prevalence shift
3 confirmation seeds.
Real measurements. Preserved uncertainty.
Every finding has a history.
Oracle correction under a fixed prevalence shift
Δ Brier = corrected − baseline. Negative values favor correction. Each bar is an original measurement; reuse is counted separately.
| Seed | Phase | Baseline Brier | Corrected Brier | Δ Brier |
|---|
A scoped result, not a universal law.
Three reserved confirmation seeds supported consistent improvement for one synthetic condition. True evaluation prevalence was supplied as an oracle input.
Completion means the predeclared operational criteria passed. It does not establish that correction helps every dataset, or that autonomous research generally saves compute. Reusing a result does not create a new replication.
How to design a careful investigation →Make the next question testable.
Draft your research plan, check the essentials, and take your notes into the local app.
Define the comparison.
Say what changes and what stays fixed.
Reserve confirmation.
Keep at least three independent seeds out of exploration.
Declare a stopping point.
A budget stop is useful information, not confirmation.
Your answers stay in this page until you download them. These are planning notes, not an executable app configuration.
Ready to run your own investigation?
The full app supports local history, portable guests, pause/resume, and local or remote Ollama-compatible inference.