web/sveltekit/tests/e2e/demo-script.md

Riprap demo script

Generated via npm run test:demo. Each query lands a full-page screenshot in test-results/demo-screenshots/<name>.png. Use the curatorial top-5 ordering for the 5-min pitch slot.

Curatorial top-5 (5-min pitch)

Order	Screenshot	What lands	Audience
1	red-hook.png	Visual language: claim glyphs in the gutter, 4 sections, citations 1-4 in the drawer, MapLibre with Sandy + DEP layers, trace UI ticking through 9 specialists. The "this is what Riprap is" frame.	Everyone
2	red-hook-houses-nycha.png	Specific damning number: ~85% of the 2,878-unit Red Hook Houses footprint inside the 2012 Sandy Inundation Zone. NYCHA register card with per-row provenance grid.	Journalists, agency analysts
3	hollis.png	The Mellea reroll demo — Riprap catches itself getting an order of magnitude wrong (0.19% → 19%) and corrects in front of the audience. The system showing its work.	Researchers, methodology audience
4	battery-live-now.png	Shifts gears from archival to real-time: NOAA Battery gauge + NWS active alerts + TTM r2 surge nowcast. Streams in seconds, not minutes.	Agency analysts (OEM), planners
5	downing-street-london.png	Silence-over-confabulation. The briefing politely refuses to invent NYC content for an out-of-scope address. Closes the philosophical loop.	Researchers, anyone evaluating epistemic discipline

Full screenshot index

Pre-vetted (passing on local Ollama with RIPRAP_HEAVY_SPECIALISTS=0)

Screenshot	Query	Intent	Stakeholder fit
red-hook.png	80 Pioneer Street, Red Hook, Brooklyn	single_address	Everyone — canonical opening
far-rockaway.png	Far Rockaway flood exposure briefing	neighborhood	CB-14, Queens-borough planners, journalists
coney-island.png	Coney Island Brooklyn	neighborhood	CB-13
hollis.png	Hollis	neighborhood	Researchers — Mellea reroll path

Added (need RIPRAP_HEAVY_SPECIALISTS=1 for register-card cells)

Screenshot	Query	Intent	Stakeholder fit
red-hook-houses-nycha.png	Red Hook Houses NYCHA	single_address	Equity / housing journalists, NYCHA-policy watchers
nyu-langone.png	NYU Langone Hospital, Manhattan	single_address	Healthcare-resilience analysts; Sandy memory
battery-live-now.png	current conditions at the Battery, Manhattan	live_now	NYC OEM, NWS forecasters, anyone in surge mode
gowanus-superfund.png	Gowanus Canal Superfund flood exposure briefing	neighborhood	EPA Region 2 + DEC + DEP; methodology
sheepshead-bay.png	Sheepshead Bay flood exposure briefing	neighborhood	CB-15; canonical register-card showcase per spec §15

Honest scope

Screenshot	Query	Outcome	Stakeholder fit
downing-street-london.png	10 Downing Street, London	Geocodes; every NYC specialist falls silent; ErrorCard all-silent OR honest "no grounded data" briefing	Researchers, evaluators of epistemic discipline

Reading a screenshot

Each image captures the full page after the SSE done event. Look for:

1. Header: wordmark, query echo, live status pill
2. Briefing prose (left column): four sections numbered 01–04, claim glyphs in the gutter (filled square = empirical, open square = modeled, filled circle = proxy, striped square = synthetic prior), citations numbered [1]–[N]
3. Map (right column, sticky): four tier-coloured polygon/dot layers. Layers with zero features are omitted from the legend (silence over confabulation, applied to the map)
4. Citation drawer (right column, below map): every cite line shows tier glyph + source + vintage + doc_id
5. Register cards (when present): per-row asset table for subway / NYCHA / schools / hospitals
6. Trace UI (bottom): one row per FSM step with fired · silent · errors summary; tier badge per row when applicable

Re-running the deck

# Local Ollama (granite4.1:8b on Mac, ~3 min/query average)
cd web/sveltekit
RIPRAP_HEAVY_SPECIALISTS=1 npm run test:demo

# AMD MI300X (vLLM, ~15 s/query)
RIPRAP_LLM_PRIMARY=vllm RIPRAP_LLM_BASE_URL=http://129.212.182.52:8000/v1 \
RIPRAP_LLM_API_KEY=<vllm-token> .venv/bin/uvicorn web.main:app --port 7860 &
cd web/sveltekit && npm run test:demo

The screenshots overwrite on each run; back them up under pitch/screenshots/<date>/ before iterating on the briefing prose.