Results
Measured results from committed artifacts: accuracy against baselines, simulation-to-real gaps, and acceleration benchmarks.
PPO dispatch against SPTF and LP
Not yet run
What will be reported: Paired difference in tonnes per hour over paired seeds, with its 95 % interval
Pre-registered criterion: At least 30 paired seeds; "beats SPTF / LP" only if the paired 95 % interval of the t/h difference excludes 0, otherwise "no significant difference" (SC-007-01)
Attention policy against SPTF, LP and PPO
Not yet run
What will be reported: Paired difference in tonnes per hour, overall and per fleet size
Pre-registered criterion: At least 30 paired seeds; the same decision rule as for PPO; the per-fleet-size comparison is reported (SC-007-02)
Slope forecasters against inverse velocity
Not yet run
What will be reported: Time-of-failure error as a share of lead time on seeded synthetic events, per learned arm
Pre-registered criterion: Median error at most 0.10 of the lead time at both noise levels; "better than inverse velocity" only when the paired interval lies above 0 (SC-010-01, SC-010-02)
Graph-network surrogate against the DEM solver
Not yet run
What will be reported: Repose-angle and run-out error on held-out geometries
Pre-registered criterion: Repose error at most 1.5 degrees and run-out error at most 5 % on every one of the 20 held-out geometries (SC-011-01)
U-Net against watershed with a Swebrec fit
Not yet run
What will be reported: Paired difference of the x50 error with its 95 % interval, real and synthetic images stated separately
Pre-registered criterion: "U-Net beats watershed" only when the interval of the paired difference excludes 0 in its favour, otherwise "no significant difference" (SC-012-03)
Pre-registered criteria of the 13 models
| Model | Pre-registered criterion | Status |
|---|---|---|
| D-FINE-N and D-FINE-S | D-FINE-S mAP50 >= 0.80 on the pooled synthetic held-out test set (SC-009-01) | Not yet run |
| RF-DETR-Seg-N | RF-DETR-Seg-N (fp32) on pit-family held-out images: mask AP50 of at least 0.70 (SC-009-05) | Not yet run |
| U-Net fragmentation | MRE of x50 <= 0.15 against the exact pile PSD (SC-012-01) | Not yet run |
| GNS granular surrogate | repose within 1.5 degrees and run-out within 5 % on held-out geometries (SC-011-01) | Not yet run |
| FNO surrogates | relative L2 <= 5 % on held-out terrains (SC-011-03) | Not yet run |
| PPO and attention dispatch | beats SPTF or LP only if the paired 95 % CI of the t/h difference excludes 0 (>= 30 paired seeds) (SC-007-01); same decision rule against SPTF, LP and M4 (>= 30 paired seeds) (SC-007-02) | Not yet run |
| IL-1 haul-truck policy | >= 95 % route success without collision over 100 seeded episodes (SC-014-01) | Not yet run |
| IL-2 excavator policy | IL-2 fill factor of at least 0.8 in at least 80 % of FEE episodes (SC-014-06) | Not yet run |
| Slope forecasters | median TTF error <= 10 % of lead time over 500 test events (SC-010-01) | Not yet run |
| Mine-to-mill meta-model | R² >= 0.95 against the chain on held-out sweeps, for kWh/t and t/h (SC-017-01) | Not yet run |
| DEM calibration | The 95 % interval of the recovered parameter covers the synthetic truth in at least 90 % of trials, at least 45 of 50, for each arm (SC-011-02) | Not yet run |
| Cosmos Reason 2 | Cosmos beats the Qwen base model only if McNemar p < 0.05 (SC-015-03) | Not yet run |
| TensorRT engines | Per-engine parity on every TensorRT version and execution provider: fp32 rtol 1e-3 and atol 1e-5; reduced precision within 1 percentage point; failing variants reported as rejected, none dropped (SC-016-01) | Not yet run |