AIYES Track 1 · AI application development

EvidenceLab

An integrity-safe AI evidence auditor for student experiments.

The software organizes evidence. The student owns the scientific judgment.

01 · Problem

Data is not yet evidence.

Students can collect measurements and still miss blank values, possible anomalies, weak repeats, uncertainty, and confounds. Generic chatbots may skip those checks and write an authoritative-sounding verdict.

Need: a guided path from raw data to inspectable reasoning without replacing the student.

02 · Product flow

Safety before coaching.

Describe the projectScreen safetyPreview untrusted CSV dataRun deterministic auditBuild a student-owned evidence brief

03 · AI strategy

Deterministic first. AI constrained.

  • TypeScript calculates data summaries and audit flags locally.
  • Optional OpenAI synthesis sees privacy-safe aggregates, not raw datasets.
  • Structured output is schema-validated and screened for forbidden verdict language.
  • Disabled features, missing credentials, provider errors, or invalid output fail back to deterministic text.
  • No open-web browsing and no AI-calculated statistics.

04 · Experience

Promise → Observe → Audit → Decide

The landing world uses generated depth plates, a bounded CSS perspective camera, and scroll-linked transforms without scroll hijacking.

EvidenceLab immersive audit chapter

05 · Evidence view

Every flag stays reviewable.

Possible outliers are labels, not automatic deletions. Group summaries report descriptive statistics without deciding which explanation is correct.

EvidenceLab deterministic audit workspace

06 · Trust boundaries

Built to fail closed.

  • Red-tier projects receive no procedure or next-trial guidance.
  • Guest projects and raw rows remain browser-local by default.
  • Raw-data save and one-time raw export are separate explicit choices.
  • Direct student identifiers are neither requested nor logged.
  • No scientific-verdict field exists in the interface or exports.

07 · Verification

Proof, with the gap visible.

579unit and integration tests passed
92cross-browser E2E tests passed
87.13%overall branch coverage
0moderate dependency vulnerabilities

Live AI, RLS, billing, legal/privacy, and qualified science-safety review remain separate proof gates.

08 · Next proof

A small, honest pilot.

Run a consented teacher-reviewed pilot using non-identifying aggregate outcomes, verify deployed service boundaries, and expand experiment support only after real student and educator feedback.

Live: ouija-olive.vercel.app
Source: github.com/rushtanu14/ouija/tree/rushex/evidencelab-submission