feat: AI transparency page — model + LLM + data-API inventory (#108)

[William-Hill]

Summary

Adds /ai-transparency listing every AI or AI-adjacent surface running in the dashboard, with: algorithm, inputs, training data (if any), where the inference runs, what data flows on invocation, and the retention policy. Closes #108.

What's inventoried

• 6 homegrown ML models (XGBoost / Random Forest, all on-prem, no third-party data flow): retention, time-to-credential, credential type, gateway math, gateway English, low GPA.
• 3 OpenAI gpt-4o-mini routes with explicit data-flow disclosures: prompt→SQL analyzer (app/api/analyze), query-result summarizer (app/api/query-summary), course-pairing explainer (app/api/courses/explain-pairing).
• 1 rule-based NLQ fallback at lib/prompt-analyzer.ts (no LLM, no outbound flow).
• 1 external data API at schools.syntex-ai.com — disclosed honestly, even though it's project-controlled, because it isn't on the institution's infrastructure.
• 1 in-development entry for the SHAP narrator.

Discoveries that shaped the disclosures

• CLAUDE.md and README.md claimed "5 ML models" and "OpenAI for natural language query analysis" — both inaccurate. The pipeline trains 6 models, and there are 3 distinct OpenAI routes (not just one). Filed fix: docs drift — CLAUDE.md and README claim 5 ML models / 'OpenAI for NLQ' but reality is 6 models + 3 distinct OpenAI routes #125 to fix the docs.
• The default query path can hit schools.syntex-ai.com even when a local DB is configured. Filed feat: deployment-hardening flag to force useDirectDB = true end-to-end (eliminate schools.syntex-ai.com data flow) #126 to add a FORCE_DIRECT_DB flag for institutions that need a fully-local data path.
• FERPA exclusion (Student_GUID) is enforced by prompt only, not by code. Filed test: enforce FERPA exclusions in NLQ-generated SQL (Student_GUID never appears in SELECT) #127 to add a runtime guard + test suite.

Implementation notes

• Content lives in codebenders-dashboard/content/ai-transparency.ts as a typed AISurface[]. Chose typed TSX over markdown because (a) the codebase convention (/methodology is the precedent) is structured TSX with content arrays, (b) typed entries enforce required disclosure fields at compile time — the failure mode of a transparency page is omission, and TS catches that, markdown can't.
• Page is linked from nav-header.tsx (visible to all roles) and from a new "See also" section at the bottom of /methodology.
• No new dependencies.

Test plan

• npm run dev and navigate to /ai-transparency; verify all entries render
• Click "AI Transparency" in the nav header from each route
• Visit /methodology and follow the new cross-link
• Confirm Deployed vs In development badges render correctly
• Confirm Homegrown vs Third-party: <provider> badges render correctly
• npm run lint passes
• npx tsc --noEmit passes

Related

• Part of epic: Spring 2026 AASCU convening follow-ups #124 (spring-convening-followup epic)
• Filed during this PR: fix: docs drift — CLAUDE.md and README claim 5 ML models / 'OpenAI for NLQ' but reality is 6 models + 3 distinct OpenAI routes #125 (docs drift), feat: deployment-hardening flag to force useDirectDB = true end-to-end (eliminate schools.syntex-ai.com data flow) #126 (FORCE_DIRECT_DB hardening flag), test: enforce FERPA exclusions in NLQ-generated SQL (Student_GUID never appears in SELECT) #127 (FERPA exclusion runtime guard)

Deviations from acceptance criteria in #108

• "Markdown-sourced for easy maintenance" → typed TS module instead. Reasoning above. Same maintainability win, no new dep, matches codebase, gives compile-time guarantees that markdown can't.
• "Linked from the dashboard footer" → there is no global footer in the app. Linked from the nav header instead, plus from /methodology. Equivalent reachability.