The published lead.
17.21min error, M2G4RTP
A number needs its context.
The future picks the inputs.
2 hoursof future data decided which parcels were scored
Right time. Wrong parcel.
+9.7 minmore error once scored by parcel, not route position
A fair rerun.
24%lower error on the same 149,263 predictions
Published results
Reported performance
The Yantai results come from LaDe, Table 4. M2G4RTP has the lowest mean absolute error (MAE) among the four methods shown. These are published results, not our reruns.
The audit describes the inspected code, not every historical paper configuration.
Comparison limits
Published results, checkpoint rescores and the matched rerun use different inputs and populations. Do not compare the paper’s 17.21 minutes directly with Hermes’s 58.64 minutes.
Input and scoring audit
Future information
The inspected pipeline keeps pending parcels only if they finish within the next two hours. It also uses future delivery order and full-period statistics. This information would not be available at prediction time.
The parcel examples are fictional.
Input filter ↗Parcel identity
The inspected M2G4RTP evaluator matches times by route position, not parcel identity. Rescoring local checkpoints by parcel increases MAE:
| Pipeline / predictions | Position MAE | Parcel MAE |
|---|---|---|
| Original / 131,454 | 12.47 | 22.19 |
| Partly cleaned / 200,534 | 12.09 | 23.27 |
MAE is in minutes. Compare scoring rules within each row, not performance across populations. Neither row is the matched rerun.
Partial cleanup restricts statistics to training data, removes true-order area features and retains busy days. Both versions still use future filtering and task windows.
Evaluator ↗Matched rerun and limitations
Matched comparison
On 149,263 matched Yantai parcel-checkpoint predictions, M2G4RTP records 77.11 minutes MAE and Hermes 58.64: a 24% reduction in error for pending parcels.
Training precedes September 1; model selection uses September 1–15; evaluation uses September 16–30. M2G4RTP epoch 12 was selected on stopping performance.
This is a local adaptation, not an exact paper reproduction. Saved metrics and training history are available; the full matched-run script is not archived.
Limits
September informed development. Queries repeat across parcels and courier-days; 149,263 is not a unique-parcel count. No repeat-run uncertainty or untouched October evaluation is available.
Recorded completion is not verified receipt. Data availability is approximate, and pending records may be incomplete.
The result is an average improvement, not a guarantee for every parcel or evidence of deployment readiness. Stock photographs are illustrative.
Dataset card: Apache-2.0. Paper: CC BY-NC 4.0. Confirm applicable terms before redistribution. No raw records or trained models are included. AI assisted coding and presentation; predictive models ran locally.