ModelScope

Inspiration

Learning an equation is different from knowing when to trust it. Students often memorize equations without learning their assumptions or valid regimes. A high fit score can hide systematic residual disagreement. ModelScope makes model adequacy an evidence-based learning task: when should we stop trusting this approximation for these measurements? ModelScope responds to the AI + Education prompt by connecting concepts to observations and student-owned experiments.

What it does

ModelScope compares measurements with a configured scientific baseline, searches continuous segmented alternatives, and exposes the supporting evidence. Students explore four 3D scientific systems, inspect linked plots and measurements, or bring CSV, pasted tables, and manual data. Outcomes distinguish supported, ambiguous, no clear transition, and insufficient evidence.

How we built it

Pure TypeScript performs baseline fitting or fixed-theory prediction, residual analysis, penalized segmented comparison, sustained same-direction disagreement checks, and an influential-observation safeguard. The same complete observations remain in the reported fits. The transition sensitivity range describes sampling and candidate sensitivity; it is not a confidence interval. Reproducible JSON preserves provenance and numerical evidence. Gemini proposes only a supported configuration from an experiment description and exact column headers. The student reviews and confirms it, supplies a justified response scale, and separately runs analysis. Gemini can also explain an existing deterministic finding using minimal summary context. It cannot fit the model, select the breakpoint, invent uncertainty, or change support. AI explains. The analysis engine decides.

Challenges we ran into

Distinguishing sustained model disagreement from an attractive fit improvement or one influential point. Keeping AI proposals reviewable, explanations bound to the current finding, and scientific results reproducible. Making four scientific scenes and dense evidence views usable on desktop and mobile.

Accomplishments that we're proud of

A working public product with four 3D systems, custom data, linked evidence, JSON exports, and real bounded Gemini. The verified release has 306 passing tests with typecheck, lint and build passing. The demo uses actual production recordings, including explicit AI setup confirmation.

What we learned

Model adequacy depends on assumptions, sampling and residual structure. Better fit alone does not establish a transition, and a transition does not establish a physical cause. Keeping AI outside the calculation path makes results reproducible and explanations easier to review.

What's next

Evaluate the educational experience with teachers and students, then test independently collected classroom datasets and the suitability of the current assumptions. These are proposed future steps.

Responsible AI and privacy

Gemini proposes a supported setup requiring confirmation and explains deterministic evidence. It cannot select a transition, fit the model, invent uncertainty, or change support. Raw measurements stay in the browser for normal analysis; optional AI sends description and headers or a selected finding with limited summary context. Free-tier inputs may improve Google products; avoid sensitive context.

Limitations and disclosure

Built-in datasets are synthetic educational demonstrations, not laboratory validation. The prototype supports restricted baselines, one possible hinge, ordinary least squares and assumed response scales. It has no calibrated false-positive rate, confidence interval, inferred mechanism or classroom outcome study. Luis Contreras, solo, started ModelScope from scratch during the overlapping eligible period; retained Git history begins October 4, 2026.

Built With

• gemini
• next.js
• papaparse
• react
• react-three-fiber
• recharts
• tailwindcss
• three.js
• typescript
• vercel
• vitest
• zod