Smok Przewodnik

The per-person constraint model (allowed, prefer_avoid, must_avoid, five presets, guest mode) genuinely changes the verdict per user, but the demo chain is broken: no PDF, no video, form unfilled, no live demo, and routes initially drawn from a simulated Bezier fallback. B's 9 is an outlier its own reasoning undercuts, and the evidence supports the more critical view of A and C.

The build is real and large, about 30k LOC across a .NET Clean Architecture backend (PostGIS, Supabase auth, Valhalla, SignalR) and an Expo app, with 79 commits by 5 authors during the event. It loses points because the required conflicting, incomplete and source-down case is nowhere demonstrated, 44 test files measure zero test cases, the only described screenshot is the stock Expo starter, and the Leaflet WebView map has no text alternative.

The model plans for provenance (confidenceState including disputed, fact statuses, one sample venue labelled unverified), but delivered data is largely invented: wheelchairAccess derived by hashing entity ids, street-surface notes hardcoded as facts, and synthetic results setting hasStepFreeAccess true while stating 'Brak zweryfikowanych danych'. The evidence supports A's and B's 3 over C's lenient 5, and no Kraków open-data ingestion is visible.

Clean Architecture with separated infrastructure, PostGIS and a Valhalla engine configured via tile URLs makes a multi-city setup plausible at the architecture level. The frontend is Kraków-hardcoded, no add-a-city path is documented, and no ingestion pipeline feeds the API, which caps the score near the middle.

No business model, pricing or numbers appear anywhere in the README, docs or code, and the submission form and deck that would normally carry it are missing. All three members independently scored it 0.