The numbers, and their limits.
Tournament t001, 1500 held-out simulated field cells on five carriers (450 MHz 25 RB, private LTE, private 5G, satellite-backhauled; 15 RB held out). The bar and the result are on the proof page.
Hidden-load error per class
See the tables below.
Interval coverage
See the tables below.
Breaking-load error against arithmetic
See the tables below.
First breaking class right
See the tables below.
Same cases, same rules
Scored against, untouched
- Resource-block arithmetic on the planner's mix
- The counter gauge: 100 % over today's resource-block use
- Per-group rules of thumb
- Today's deterministic scheduler (the Rust link engine)
- External check: the M/D/1 closed form on a single class
Thresholds set on validation data and frozen for every test, including the cases the engine never saw. The truth model differs from the engine's own.
The bar, set before the first run
The stochastic scheduler must move the breaking class or load beyond its interval, against resource-block arithmetic.
On surge and storm cells, arithmetic puts the break -33 % off for scada polls, -37 % off for alarm bursts, -29 % off for push-to-talk, -19 % off for event video, -37 % off for pump cards (median against the scheduler), and names the wrong first class in 54 % of breaking cells.
First proof
Two to three weeks: bursty arrivals and an uplink poll queue with an oil service mix, 2,000 seeded runs on composite 450 MHz cells; the break by arithmetic against the scheduler, per class; then hidden load from synthetic counters against known truth.
The break: which class, at what load
| First class right | Break error, median | …on cells that break | Says “fits” when it breaks | |
|---|---|---|---|---|
| Headroom: counters → load → scheduler | 75 % | 0.0× | 1.0× | 35 % |
| Resource-block arithmetic (the plan) | 17 % | 0.0× | 3.0× | 88 % |
| Counter gauge: 100 % ÷ PRB use | 17 % | 0.3× | 0.7× | 11 % |
| Rules of thumb | 17 % | 1.2× | 1.3× | 19 % |
| Today's deterministic scheduler (Rust engine) | 46 % | 1.0× | 1.0× | 8 % |
Truth: the true mix ramped on the stochastic scheduler (9 scales × 4 seeds); a class breaks when its packet-delay budget is missed by more than 5 % (push-to-talk, video) or 10 % (data, polls), or its admission blocks above 1 %. Beyond 6× is scored as 7×.
Hidden load per class
| Load error | Inside 90 % interval | |
|---|---|---|
| Headroom: counters → load → scheduler | 40.7 % | 71 % |
Alarms, push-to-talk, video and crew data on cells whose export has per-class counters (polls and pump cards are configuration). Without them the estimate refuses per class: refusals correct on 100 % of the held-out cells without per-class counters.
The other answers
| Link edge: precision | recall | Reconnection error, classical | post-quantum | |
|---|---|---|---|---|
| Headroom: counters → load → scheduler | 56 % | 51 % | 11.2 s | 10.2 s |
| Resource-block arithmetic (the plan) | 0 % | 0 % | 14.2 s | 10.3 s |
Endpoints flagged at the link edge against the true SNR. Reconnection: seconds until 95 % of the fleet is re-authenticated after an outage; median true times 32.9 s classical, 35.0 s hybrid post-quantum (unfinished within the 60 s sample counted as 60 s).
Against the bar: arithmetic against the scheduler
| Median offset | 90 % interval | Cells | |
|---|---|---|---|
| SCADA polls | -33 % | -36 % to -29 % | 153 |
| Alarm bursts | -37 % | -43 % to -32 % | 88 |
| Push-to-talk | -29 % | -33 % to -25 % | 162 |
| Event video | -19 % | -23 % to -13 % | 131 |
| Pump cards | -37 % | -42 % to -32 % | 107 |
On surge and storm cells, arithmetic puts the break -33 % off for scada polls, -37 % off for alarm bursts, -29 % off for push-to-talk, -19 % off for event video, -37 % off for pump cards (median against the scheduler), and names the wrong first class in 54 % of breaking cells.
What is proven. What is not.
Proven, in simulation
- Names the first class to break in 75 % of held-out cells; resource-block arithmetic 17 %, today's deterministic scheduler 46 % (simulation)
- Per-class load from counters: 41 % RMS error where per-class counters exist; refuses per class on every cell without them
- On surge and storm cells the stochastic scheduler puts the break 19–37 % below arithmetic for every class, beyond the interval
- Reconnection after an outage: median 33 s classical, 35 s post-quantum on the simulated cells
Not proven
- The counter gauge (100 % ÷ today's PRB use) reads the breaking load better: 0.7× against 1.0× median error on breaking cells, and says “fits” wrongly less often (11 % against 35 %)
- Load intervals cover 71 % of true loads, not 90 %
- No real counter export used
- The uplink scheduler is new and class-level, not per flow
- KPIs on satellite-backhauled cells fall outside the standard's measurement condition
- Interference localisation not built
The first real number comes from your data.
A month of your exports, replayed offline, gives the figure that matters on real operations. We will show it to you whether it is good or bad.