DeadZone's demo deadline is 10 September 2026. This is a six-day sprint, not a product build. The bias is: fewer features, working reliably, demonstrable offline. A judge who watches one SOS cross three phones in airplane mode and then sync will remember it — five half-working features leave nothing behind.
Every cut below is deliberate. If a task seems to need one of these,
stop and ask rather than implementing it. Where a cut leaves a
TODO(post-sih) in code, that marks the correct-but-slower path for
after the Grand Finale (December), not a bug.
Cut. Typed SOS only. Voice capture and local transcription add a model, a permissions surface, and a failure mode (background noise, dialects) with no offline fallback story we can build in six days. Typed input is reliable and testable end to end.
Cut. Hand-translated fixed Hindi strings; the LLM answers in English. A full translation model is another multi-GB dependency on the edge node competing with Phi-4-mini for VRAM, for a feature (fluent multilingual LLM output) that isn't the core pitch. UI copy that needs Hindi is translated once, by hand, and shipped as a fixed string table.
Cut. Straight-line haversine distance and bearing. Turn-by-turn road routing needs offline road-network data prepared and hosted for the demo area, plus a routing engine to stand up and keep running. Haversine distance/bearing is enough to rank "which responder is closest" for the demo and has zero infrastructure cost.
Cut. Greedy nearest-available only. Optimal assignment matters at scale; for a six-day build with a handful of demo responders, greedy nearest-available produces the same visible outcome without an optimisation library or its edge cases.
Cut. Unsigned envelopes for now. Signing matters for a real deployment resisting spoofed SOS traffic. It does not change what the demo shows, and it adds key management (device identity, key exchange over an ad hoc mesh) that isn't solved yet.
Cut. The ttl field exists in the envelope, but only A → B → C
is demoed.
True multi-hop mesh routing needs a routing/forwarding strategy and
loop prevention beyond what a hop-count field alone provides. The field
is reserved in the contract so the demo path (one relay) is a real
subset of the eventual behaviour, not a special case that gets thrown
away.
Cut. Plain PouchDB sync, last-write-wins.
The same SOS document arriving via multiple relay paths already
collapses automatically because the message UUID is the PouchDB
document _id — see the architecture note in CLAUDE.md. Beyond that,
custom merge logic for concurrent edits is a real distributed-systems
problem that doesn't need solving to demonstrate offline sync working.
Cut. Unstyled but legible; visual design arrives from outside the
build team.
Screen time this week goes to the offline SOS → sync path actually
working, not to visual polish. Styling is centralised behind tokens
(packages/ui, still a placeholder) so a design pass can drop in later
without touching component logic.
The kept scope is exactly what's needed to demonstrate the narrative spine — COMMUNICATE → UNDERSTAND → PRIORITISE → COORDINATE → DELIVER — end to end and offline. Everything on this list is real work that improves correctness, scale, or security, but none of it changes whether the judges see a working offline rescue flow on 10 September.