The project asks whether repairing stale or inaccurate travel-time inputs can improve route efficiency and robustness. It does not try to imitate the historical driver sequence.
Source data and simulated conditions
The starting route, stop and package data come from the 2021 Amazon Last Mile Routing Research Challenge Dataset. The dataset contains historical deliveries from 2018. Its location data are obfuscated; the map shows approximate schematic locations.
Amazon’s directed travel-time matrices are historical averages, not measurements of current traffic. Current traffic conditions, report ages, recent observations and input faults in this project are synthetic. An input’s age alone cannot reveal an arbitrary traffic shock.
The source dataset is attributed to Amazon and the authors listed in the AWS dataset registry, which describes its Creative Commons Attribution-NonCommercial 4.0 license. The downloads below retain this project’s export provenance.
How a recorded comparison is produced
Separate complete routes chronologically. Training, validation and evaluation receive different routes before synthetic variants are created. Variants of a route stay together.
Create a simulated decision snapshot. Historical inputs seed synthetic current conditions and the reports available at the modeled decision time. The hidden reference stays fixed across paired method comparisons.
Prepare the planner’s travel-time input. Compare raw reports, a simple repair rule and CatBoost repair. Repair receives observable evidence, not hidden current values, defect flags, future measurements or evaluation outcomes.
Run the same route planner with the same budget. The equal-budget comparison isolates changes in the travel-time input. A recorded guarded policy may choose a fallback; its decision and reasons belong to that saved policy.
Evaluate the saved routes. Measure directed travel, service, waiting, lateness and validity against the fixed simulated reference. Report unfavorable cases and estimate uncertainty at the route level.
Review the recorded decision. The demo supports accepting a plan, keeping the recorded raw plan or deferring. A saved review is a decision record; it does not dispatch a vehicle.
What the replay means
Each time point is a separate counterfactual restart of a complete route from its depot under that decision snapshot. Moving the time control does not show a vehicle progressing through a continuous trip. This is recorded simulation, not live tracking.
The website displays frozen plans and outcomes. It does not train or run a model in the browser. Map paths connect stops through an illustrative city; they do not represent real street geometry, actual building heights or turn-by-turn navigation.
Estimated metrics describe the information used by the planner. Revealed outcomes use the retrospective simulated reference. Review effort and computational runtime are different quantities and should be reported separately. Viewing results or downloading evaluations can inform a reviewer’s judgment.
Limits of the current evidence
Round 8 contains 16 previously inspected routes and 1,024 correlated planning snapshots. It is not a fresh final holdout.
Synthetic traffic and reporting faults cannot cover all operational conditions. A repair may help one condition and harm a clean or unfamiliar case.
Historical raw runs with an older planner budget are separate baselines. Differences against them do not measure the effect of repair alone.
A heuristic search does not establish an optimal route. Duration improvements alone do not establish acceptable lateness or route validity.
Missing evidence remains missing. No recorded aggregate effect guarantees operational savings, safety or deployment readiness.
Human review and deferral must be counted, including the effort and outcomes of decisions left unresolved.
Amazon Last Mile Routing Research Challenge dataset (2021), Amazon. Distributed under CC BY-NC 4.0. Dataset description: 2021 Amazon Last Mile Routing Research Challenge: Data Set, Transportation Science, doi:10.1287/trsc.2022.1173.
retrievalDate
Null — not available in this export
retrievalDateNote
Exact original retrieval timestamp was not recorded. Saved source metadata records checksum verification on 2026-10-17; no download occurred for this export.
Recorded scientific compute is historical; exporter only reads, verifies and serializes saved artifacts.
2
Human effort and monetary costs have not been measured.
aiUseDisclosure
AI coding/research assistants developed the simulation and export tooling and helped interpret saved results. The displayed values are copied from hash-verified saved artifacts; narrative labels and example selection are illustrative.
limitations
Provenance manifest: limitations / 13 records
Record
Value
1
Four examples were selected retrospectively from inspected development results; they are illustrative, not representative or fresh validation.
2
Amazon travel matrices are historical averages. Current traffic, ages, report faults and supporting feeds are simulated.
3
All snapshots restart a complete route at clock zero. timeMinutes selects the saved report snapshot, not an already driven route.
4
Lateness is measured at service start against sparse package deadlines. Stop windows summarize cached bounds; they do not replace the original package deadline arrays.
5
Coordinates are rounded to three decimal places. Orders are schematic directed stop sequences, not street geometry.
6
Per-stop package counts are unavailable in the cached route object and remain null; no per-package data is exported.
7
Estimated metrics only copy saved observable schedules. Missing estimated validity remains null; retrospective evaluation never fills estimated fields.
8
At most 24 selected edges per plan are shown, prioritized by observable repair/report disagreement. The full order is always retained.
9
Shared supporting observations are not independent per-edge reports; supportCount at snapshot level copies the saved unique_feed_reports counter.
10
Original raw is an older-budget historical plan and is the exact guard fallback; matched raw uses common starts and the new budget.
11
No new model loads, training, prediction, planning, simulation, reference-matrix decoding or outcome evaluation occurs in this export.
12
No policy qualified in round eight. No real Amazon savings, optimality, calibrated safety or unseen-test claim for S79.
13
Actual human review/deferral count is zero; human effort, cost and inference uncertainty are unmeasured/null.