A demo of an event-driven world model applied to critical infrastructure
What this demo shows
An AI model that was never trained on data from San Francisco’s December 20, 2025 blackout predicts its 911 emergencies from the outage, weather and calendar, then re-runs the day with power back at 3 pm.
Model trained on
San Francisco fire & EMS incident history (type, fire district, time) from before May 2025, alongside the power-grid state, weather and calendar at each moment.
What Dark Saturday proves about the approach
Supports exact, unlimited what-if analyses
Rewind to any moment, change one thing and run it again. Everything the change doesn’t reach stays identical byte for byte, so every difference comes from your change.
In demo
Restore the Mission substation at 3 pm and about 135 emergencies per run never happen. Battalion 10, outside that substation’s reach, came out identical in all 5 runs.
See it in 2 Rewind
Delivers precise causal tracing of events
Point at any single event, trace what caused it, and test whether it would still have happened without that cause, within the model.
In demo
A 3:03 pm electrical-hazard call in the Sunset exists only because of the outage. The outage raised the model’s rate for that call about 190-fold, and with power restored it never happens.
See it in 3 Why
Compute efficient training and inference
Trained and run on ordinary CPUs. Compute is spent only where something happens, so quiet parts of the world cost nothing.
In demo
Twelve days of the city simulated in about 78 seconds on one CPU processor. Built, trained and run entirely on CPUs.
Results describe the trained model, not a guarantee about the real world. Inferred details (routes, which buildings went dark) are labeled in the demo.
Dark Saturday
San Francisco, Saturday, December 20, 2025: the blackout replayed on the real city from 911 dispatch, grid-outage and building data, then rewound by an event-driven world model. Read the essay.
World model output
Customers out
ENGINES
TRUCKS
MEDICS
Reality
Model’s own rollout
real 911 dispatch and unit trips
trained only on data from before May 2025; given the real grid, weather and calendar; incident locations drawn from past incidents of the same battalion and class
Click any incident after 1:09 pm in a restored battalion to trace its chain, from the committed logs.
Restoration variants ranked by incidents avoided (mean of three runs) per 1,000 customer-hours restored early: a cost proxy, not a dollar cost.
What a general assistant can say without this world
Power outages are known to raise the risk of elevator entrapments, and restoring power sooner generally reduces emergency calls. No numbers, no specific incidents: general knowledge, not this event.
The world’s answer
This incident’s exact chain back to the failure, and a ranked, tested set of fixes, from the committed logs of a model that watched this city.
loading…
Pose your own alternative world
Pick a hypothesis and its settings. The world model runs it now, on this server, and the page draws each event as it is committed. A request someone has already run replays from the cache.
World model output
Beat 1: The model was never told
Saturday, December 20, 2025: a fire at PG&E’s Mission substation (1:09 pm) leaves about 130,000 customers, roughly 30% of the city, dark within two hours. The city is Google’s photorealistic 3D San Francisco, and each of 168,674 LiDAR-measured building footprints is tinted by its power state. Drag the divider: real 911 dispatch on the left, the event-driven world model’s own rollout on the right, from 12:55 pm onward. The model was trained only on data from before May 2025 and given only the real grid, weather and calendar. Its incident locations are drawn from past incidents of the same battalion and class, not measured. Its unit trails are its own dispatches, station to incident in a straight line, timed from assignment to release in its service log: inferred routing, real timing.
About this beat
Left of the divider is measured: SFFD calls and every engine, truck and medic trip, routed on the street graph. Right of it is the model’s own output from its committed logs, run 1 of 5.
Counters start at 12:55 pm, when the model takes over. The model’s “trucks busy” is its service log’s own count of open truck assignments.
Which buildings are dark. EAGLE-I is county-level: it says how many customers were out, not where. A deterministic fraction of the buildings in the reported-dark neighborhoods turns steel blue as the county curve rises toward its peak; buildings elsewhere stay lit. It is a visualization, not a claim about which buildings lost power. Both sides show the same buildings, because the model is given the real grid.
World model output
Beat 2: Rewind and change one thing
A rewind reconstructs what was knowable at the chosen moment from the record, changes one thing, and replays the rest as a possible world: everything outside the change’s reach stays identical, and what differs is the model’s claim, relative to what it has learned. Here, restore power at 3 pm, using the real per-battalion restoration data. Incidents identical to the unchanged world stay grey. Incidents the restoration prevents appear as violet rings, larger for elevator rescues, and that is the dominant effect. A few incidents appear only in the restored world, in magenta: the learned response isn’t the same in every class, so restoring power doesn’t reduce every incident type everywhere. Buildings in the served battalions relight as the model’s own outage input falls. The second alternative world adds one engine at Station 1 from 1 pm and follows the change as it spreads from station to station over three days. Or pose your own below the map.
About this beat
Matching. The two worlds are matched incident by incident, on (battalion, class, incident id); everything outside the intervention’s reach is identical by construction. The stats panel counts five runs, from the restoration to 6 am Sunday.
Which buildings are dark. Each battalion’s own outage input sets the share of its buildings without power, and each building belongs to its nearest fire station’s battalion; no per-building battalion boundary exists in the source data.
Customers out is real for both worlds: the restored curve is rebuilt from the per-battalion restoration series with the same formula that produced the real county curve, which it reproduces exactly on the real input.
The engine alternative world. Each battalion’s outage follows the loop world’s own grid, de-energization and repair-crew logs, and the county figure is the sum of those same battalion levels. Dispatch lines run straight from station to incident: white is the alternative world’s dispatch, grey dashed is what the unchanged world sent instead.
Flags are computed from the same sources; no number in a flag is invented. Neighborhood names for battalions B02, B03, B05 and B09 are approximations from their stations’ locations; no published battalion-boundary map exists.
Your alternative world runs the world model on this server with your settings and draws each event as it is committed, compared against the same run with no intervention.
World model output
Beat 3: Ask it why
Click any incident after the fire in a restored battalion (rings mark the class these logs show restoration preventing most) to trace its chain back through the committed logs: the outage input, the model’s own rate for that class there, the shared draw that accepted it, and a but-for test against the restored world. It is exact with respect to the learned model. “Cheapest fix?” ranks real restoration variants by incidents avoided per customer-hour restored early, both counted over the same window.
About this beat
The chain reads each value from the incident and rate logs’ own committed rows (run 1): the rate with and without the restoration, the normal rate, and the shared draw. Only model-generated incidents are eligible; an incident whose rows lack a value is left out, not filled in.
Suggested incidents are the class these logs show restoration preventing most, computed from the logs, not chosen by hand. The model produces almost no elevator rescues in the afternoon, under-predicting the blackout’s real elevator surge.
Cheapest fix. Each variant is matched against its own committed logs (three runs), with no scaling. Avoided incidents and customer-hours restored early are both counted from 1 pm Dec 20 to 6 am Dec 21.
Which buildings are dark follows the county curve, as in beat 1.
Example use cases for critical infrastructure
Restoration priority
Which substation or feeder to bring back first to avoid the most emergencies, tested before crews roll.
Utilities
Planned shutoffs
Run a wildfire safety shutoff on the model first and see which neighborhoods and services take the hit.
Utilities, regulators
Where to stage units
Test where one extra engine or ambulance does the most good during an outage, storm or heat wave.
Fire & EMS
After-action review
Show which calls the outage caused and which it didn’t, with a traceable cause for each.
Cities, regulators, insurers
Backup power rules
Estimate how many elevator rescues and hazards backup power in high-rises or signals would prevent.
City planning, building owners
Provenance
City: Google Photorealistic 3D Tiles, graded to night on this page. Buildings: DataSF’s 177,023 LiDAR-derived building footprints (ynuv-fyni), simplified (Douglas-Peucker, slivers under 15 m² dropped) to 168,674; each is a volume from its measured ground to its measured roof peak that tints the photoreal mesh inside it. Neighborhoods: DataSF’s 41 Analysis Neighborhoods (j2bu-swwd). SFFD Fire Department Calls for Service: DataSF, Dec 16–23, 2025 unit-level rows, for incidents, fleet gauges and every engine, truck and medic trip, routed on DataSF’s street centerlines (3psu-pn9h). EAGLE-I San Francisco County customers out: Oak Ridge National Laboratory, 15-minute resolution, CC BY 4.0, shifted from UTC to Pacific time. Fire stations (DataSF, wt27-4jim) and ED hospitals (OpenStreetMap, emergency=yes) locate trip origins and destinations. Mission Substation: 66 8th St (OSM way 35115839; PG&E), fire at about 1:09 pm. High-voltage lines: HIFLD Electric Power Transmission Lines, clipped to SF.
Inferred, and labeled as such:
Routes. Engine and truck origin is the home station by unit number (E07 is Fire Station #7); a medic starts from its own previous incident, or the nearest station on its first run. The path is the shortest path on the street graph, timed from dispatch to on-scene. A leg is skipped, not estimated, when its times or station are missing.
Transport destination. The nearest ED hospital by straight-line distance; the real destination is not in the data.
Reported-dark neighborhoods. The news and PG&E list is mapped onto the Analysis Neighborhoods; “Civic Center” falls inside “Tenderloin”, “Panhandle” inside “Haight Ashbury”, and “downtown/SoMa” maps only to “South of Market”.
Observed outage core (Tenderloin/Civic Center). Where SFFD calls and SF311 reports both ran at least four times their baseline density: not a utility-reported footprint.
The ~30% figure is PG&E’s own; no share of the city is computed from EAGLE-I’s modeled customer total.
World model output
Everything marked “World model output” is read from the event-driven San Francisco world model’s committed logs: five runs, each with and without the intervention, warm-started on real history through 12:55 pm Dec 20 and generated by the model alone from there. Only model-generated incidents are counted; replayed history is excluded everywhere. Each beat’s own notes sit with its text above.