How the ambulance run was built

The recording

Same pickup and hospital; 30,000 on road at the matched starting state. The wet case prescribes a 2,000 ft diameter, 0.20 m deep flood, closing 83 road links.

One shared clock. After dry arrival, its entire image and traffic are held while the wet recording continues.

Blue circle: prescribed flood footprint. Yellow → orange → red: increasing modeled road occupancy, not measured speed. The ambulance icon follows the recorded trip.

Network and demand

Motion uses a prepared NYC DCP LION 26C network: 10,111 nodes and 16,870 directed road links. Preparation uses complete-polyline lengths, contracts links shorter than 10 metres, selects parallel links and trims to the borough boundary. Missing lane and speed values use declared exporter defaults. No turn-restriction table is supplied; graph-permitted routes do not certify real-world turn or access legality. The importer includes existing private-street source features as service roads, without independently certifying emergency access.

Synthetic weekday-morning demand draws on Census LODES employment journeys, NYMTC household travel and modeled origin–destination sources, Hub Bound controls and TLC trip sources. Boundary routing uses a separate full-city OSM graph. Boundary mapping retains 18,519 of 19,595 weighted journeys; expansion gives 148,943 unit-weight vehicles, then native reachability excludes 32. A priority vehicle and 1,206 authored road-entry vehicles bring allocation to 150,118. At the shared starting tick, 30,000 are on road and 90,333 trips are pending. Allocation is not simultaneous road occupancy. Each actor uses one passenger-car unit (PCU); this is not a calibrated mixed fleet. Queues outside the selected network are not simulated.

Motion, queues and capacity

The native Rust engine advances in one-second ticks. Directed routes, finite storage, movement-specific first-in-first-out queues, discharge credits, downstream receiving checks, one-lane ordering and priority rules constrain motion. This is a point-motion and queue model, without collision-free vehicle bodies, lane changes or gaps.

The inherited setup declares triangular kinematic-wave defaults on primary, secondary, tertiary, residential, unclassified and service links: jam density 1/7.5 PCU per metre per lane, capacity 0.475 PCU per second per lane, and backward-wave speed derived from free-flow speed. Class-selected links too short to store the largest supported vehicle remain spatial queues. Released downstream space takes time to reach the entrance; priority cannot bypass receiving supply.

The retained model uses a legacy storage floor. Ordinary spatial queues use length × lanes / 7.5 metres, with a minimum of ten PCU. This floor can overstate short-link geometric storage and delay spillback. Wave links use their diagram storage without that floor. Global flow and storage factors are 1.0.

Spatial-queue headways are 1.565 seconds for free→free or free→jammed, 1.73 for jammed→free and 1.65 for jammed→jammed; occupancy above 40% of nominal storage selects the jammed regime. Sending uses lanes/headway × green-time fraction × flow scale × road speed factors; wave links use diagram capacity × green-time fraction, subject to receiving supply. The native motorway rule has a 12.5% capacity drop, 22% breakdown and 15% recovery occupancy; these defaults are not measurements or a claim of activation in this pair.

After ten seconds of consecutive refusal, a gridlock escape can move an ordinary queue head beyond destination storage. It must still pay sending credit and respect closures, legal graph movements and wave receiving supply. It does not force the priority vehicle. No run-wide activation count is established; this rule limits claims of strict geometric capacity.

Signals, priority and route decisions

No timed signal plans or preemption transitions were supplied or recorded. A declared 0.5 green-time fraction reduces throughput where backed timing is unavailable; it does not simulate measured red/green cycles. Emergency signal policy is obey. Priority unobstructed speed ratios are 1.15 on links at or above 40 km/h and 0.8 on slower links. A 5 m/s yielding corridor is declared for eligible multi-lane spatial queues, but neither journey records a corridor entry. No continuous vehicle-body corridor or footprint mode is configured.

Ordinary drivers make discretionary attempts every 60 seconds, known-invalid retries five seconds apart and predeparture attempts only when due. These are attempt rules, not guaranteed route changes; congestion feedback and staggered reviews are inherited. Priority actors are exempt from the ordinary budget: reviews occur every 30 ticks and in staggered slots (the wet diagnostic records a 100-tick stagger interval). Revisions preserve an already owned traversal and replace its continuation.

This pair uses current-state queue-price routing, not physical future replay or measured-flow BPR. The wet journey travels west and returns through the pickup before heading south. A diagnostic reproduction matched original capture bytes through tick 2800 and identified cheaper optional priority replans at ticks 2670 and 2760. That explains the modeled decision sequence, not global optimality or a real-world need to reverse. Subtracting its 94-second prefix cannot establish a no-loop arrival time.

Prescribed flood

The wet input prescribes uniform 0.20 m street depth within radius 304.8 m (2,000 ft diameter), centered at latitude 40.749788, longitude −73.974121, from tick 1800. Complete road polylines are intersected with the extent: the at-grade closure policy derives 83 directed-link closures with zero speed and capacity. The treatment supplies no manual road-ID list. The receipt records zero excluded or assumed-at-grade intersecting links.

Closures already exist at journey departure, tick 2700. Closed links refuse entry and discharge; actors already on them are held. Priority and gridlock escape do not override them. Partial passability and priority-only flood access are absent. This prescribed footprint does not run terrain-derived bathtub calculations, rainfall/runoff, rupture hydraulics or a hydrodynamic solver. Water does not spread or recede through a hydraulic simulation.

Recorded time and reproduction

Both saved runs resume at tick 1800 and advance 3,600 one-second ticks to 5400. Departure is 2700; arrivals are 2912 dry and 3646 wet. Full actor/queue capture is every two ticks (1,801 source frames per run); selected-actor measurements are a separate 60-tick series. Rendering interpolation smooths display, without another physics calculation.

Each corrected capture selects 450 endpoint-inclusive images: sample i maps to 2700 + 212 × i/449 dry, or 2700 + 946 × i/449 wet. The comparison selects the latest genuine image at or before the common elapsed-time clock, repeats images on a 60 fps encoding grid and holds the actual dry arrival image, including its frozen traffic. Each map can lag that clock by up to its sample spacing: about 0.4722 model seconds dry or 2.1069 wet. This does not create 60 independent observations per second. A 45-second active recording ends with a 0.5-second hold; dry arrival ideally falls at 10.0845666 seconds and first appears at encoded frame 606 (10.100 seconds). Neither source includes pre-journey or flood-onset footage. Exact arrival images have no remaining ambulance actor.

The shared input is a reconstructed synthetic warm state retaining demand, routes, queues and RNG state, with recorded point reassignment, reset wait history and a retargeted priority trip. It is not exact historical continuation. Final inputs differ only in case name and prescribed flood; initial actor-byte hashes match. Source construction uses seed 17; the seed alone cannot recreate the retained state.

Reproduction references: backend source 980ebf2f0bf8ca05f6f12d089e3b88c5feb4ac0e; native binary SHA-256 c53e458d91229c8eb4036c3a483fbd05169dac0897f0faf1445f63a46a3bd2b1; prepared road binary 531c3e52b7bcfe1075b0d2dc58d542c1c99775944005f2ebb66be6cd34aa625a; common initial execution input 6d4a72d9658610892a5576f115523245add4e03853daa1aba58774305205af1d. Matching them needs the exact authorized inputs and engine; no downloadable reproduction bundle is supplied here.

The wet journey lasts 946/212 = 4.4623 times the dry journey. Completion, byte verification, diagnostic equivalence and contract tests establish the specified software behavior and provenance. They do not calibrate demand, validate local ambulance forecasts or prove bitwise reproduction over the full horizon on every machine. The arrival difference is the total model response, including endogenous queues and routing, without attributing it to one isolated mechanism.