Assignment, sequence, and promise under uncertain demand and travel time.
VIA is not a standalone router. It is a domain-specific instantiation of the Principia Intelligence OS, pointed at last-mile and long-haul fleet logistics (including the real challenges of EV charging time and queueing on extended routes). The OS supplies the calibration, uncertainty, and provenance discipline; VIA supplies the routing decision (CVRPTW with robust buffers and priced disjunctions). The intelligence it consumes comes from p1-chaos — the same forecasting, calibration, regime, and conformal engine that serves every other application on the stack. VIA is what happens when that intelligence is handed to an optimiser that knows how to spend it.
across 3 banded route-days where chaos beats baseline on P&L (64-cell regional sweep; most days chaos costs more on $/parcel — see objective profiles). Long-haul is the flagship savings story.
Held-out scenario replay in ViaDB; marketing slices are fixed seed-42 route-days with real geometry.
Empirical coverage in replay is profile-dependent: efficient (min_cost) trades narrower buffers for lower booked cost; reliable (balanced) maximises promise hit-rate.
At a glance
Assignment, sequence, departure, and promise under uncertain demand, travel time, service time, and delivery success.
P1-VIA is the last-mile and long-haul member of the P1 family. It solves real-street CVRPTW (dense urban as well as multi-city corridors such as Lille–Paris–Lyon–Marseille) against calibrated probabilistic inputs from p1-chaos rather than point estimates. It explicitly models EV charging time and queueing on long routes so the A/B comparison honestly reflects electric fleet realities. Given today's stops, vehicles, and windows it decides which vehicle serves which stops, in what order, when it departs, and what coverage-guaranteed promise window to publish.
The intelligence layer (p1-chaos) supplies travel-time quantiles, dwell distributions, success probabilities, demand, and regime. The optimiser reasons over those distributions with robust buffers and priced disjunctions. Value is isolated the P1 way: identical solver, mean inputs vs chaos inputs. The delta is Chaos in action.
--
xx
--
xx
--
xx
The Bottom Line
VIA leads on the thing a dispatcher actually publishes: the promise. Same solver, calibrated inputs instead of averages, and last-mile promise coverage goes from 48% to 88% against an 80% target. On long-haul it's starker — the baseline books all 112 stops and keeps 6% of its word; VIA serves 87, declines the 25 it can't profitably promise, keeps 80%, and does it at −54% cost per parcel. Across dense-metro, suburban, rural and long-haul, a window quoted at 80% delivers ~83% out-of-sample; the naive baseline keeps 21%.
Every figure here is SIMULATED — model-vs-model, same solver both arms. Current results are a baseline, not an upper bound — what the model shows before your routes, your telemetry, and your constraints enter the loop. Coverage is a validity claim, verified against real deliveries only after a pilot.