Crucible

Inspiration

The persuasive part of a research result is often easier to find than the experiment behind it. Which metric was chosen before the result? Which examples were already inspected? Does the comparison support the claim? Crucible keeps those questions attached to the answer.

What it does

Crucible turns a bounded research claim into an inspectable contract, separates discovery from confirmation, and exports the computation with its evidence. The demo follows two completed tabular folios.

An authored nonlinear sensor experiment compares a random forest with logistic regression and requires a two-percentage-point balanced-accuracy advantage. One condition was registered before confirmation predictions. Across 24 independently generated groups, with training seeds averaged within each group, the candidate's mean group-level advantage was 44.94 percentage points. That conclusion belongs to the authored XOR generator, conditional on fixed training data and seeds. It is not a finding about public research.

The office occupancy folio measured -1.57 percentage points across six usable calendar days in one room. Those adjacent days are not established as independent, so the verdict remains inconclusive. Crucible does not turn uncertainty into a claim of equivalence.

How I built it

React and TypeScript present the workbench; a Node service and SQLite preserve transitions and exposure history. Trusted Python code executes authored numerical methods. Separate verifier code runs in a separate process under the same operating-system identity. All eight mechanical checks passed for the synthetic result, without implying external scientific review.

Nemotron Ultra on Fireworks drafted the historical tabular contract and ranked discovery challenges. Separately, Nebius Token Factory served NVIDIA Nemotron Ultra for qualified primary-source extraction. That receipt did not produce the completed tabular results. Tavily made a functional runtime search that returned a primary SAHI source for the newer wheat-image workflow.

Challenges and what I learned

The hard part was preventing an attractive result from rewriting its question. Registration, exposure tracking, independent calculations and retained failures became first-class parts of the interface. I am proud that an inconclusive result is useful and legible rather than treated as an error.

What comes next

The repository also contains a source-informed wheat-image experiment with an acquired, audited corpus and qualified source extraction. Detector fitting and scientific measurements remain unrun; no wheat-detection result is claimed. Next steps are completing that declared experiment and obtaining independent research review. Completed tabular capsules already reproduce their frozen computations without model calls after dependency installation.