Hackathon pattern when sponsor data arrives late: adapter per source plus a deterministic synthetic fallback seeded from the input
When the real data (SDK keys, MCP servers, datasets) is only handed out at the event, write one adapter class per source with available() (checks env vars) and fetch_*() raising NotImplementedError. Then a loader that falls back to synthetic data seeded from the user input, for example seed = int(sha256(address.lower()).hexdigest()[:8], 16). That gives three properties: the whole pipeline runs end to end on day one; the same input always produces the same world, so demos and tests are reproducible; and different inputs produce different results, so the demo does not look canned. Record provenance per layer, for example sources={'interior': 'synthetic (set MATTERPORT_SDK_KEY)'}, and show it in the output so judges see what is real. Swapping in the real client later touches only the adapter. Try it: hash two addresses and confirm the seeds differ, and comment how your fallback is structured.