The heavy-hydraulics programme asks when a small excavator should carry a cheap fixed gear pump instead of a load-sensing pump. Its selection rule predicts the gear-pump efficiency at which the two tie on fuel. The builder's best form so far is the "two-price" rule, with an error of 0.0056 (left-out RMS over 308 cases).
Two milestones ago (7r) a pick table did better, at 0.0052. Each row lists the pump sizes the fuel optimum picks for a given bend of the idle drag curve (drag at 79 L/min over drag at 64 L/min). The rows were built by bending measured pumps' curves synthetically. This morning's milestone (7s) found the gain depends on one choice. When a curve bends more, does it keep its level at 64 L/min and get dearer at large sizes (the table wins)? Or does it keep its level at 79 and get cheaper at small sizes (the table loses, 0.0059)? The log called that "a property of real pumps".
Asking the measured pumps
There are 13 measured pump variants besides Danfoss, the one pump that bends hard (1.78). Under a power-law bend that pivots at some flow q*, the slope of log drag at 64 L/min on log bend across pumps tells you where q* is. So the run read q* off the other pumps for each held-out family, never touching the pump being predicted. It used two estimates: a pooled slope over all donor variants, and a slope within the two families that come in several variants. The predictions went into the log before any number was computed.
| table anchor | pivot (Danfoss fold) | left-out RMS | Danfoss error |
|---|---|---|---|
| fixed at 79 L/min (7s) | 79.0 | 0.0059 | 0.0110 |
| fixed at 64 L/min (7r) | 64.0 | 0.0052 | 0.0063 |
| pooled slope | 65.4 | 0.0052 | 0.0068 |
| within-family slope | 61.1 | 0.0050 | 0.0054 |
| two-price rule | 0.0056 | 0.0061 |
Both measured reads put the pivot near 64 L/min, not 79. With either one, the table beats the two-price rule in all three tests: the left-out condition, unseen half idle speed and unseen idle shares. With the within-family anchor, the table's Danfoss error comes in below the two-price rule's for the first time. Seven of eight predictions held.
The prediction that failed, and why it matters
I expected Danfoss's drag at 64 L/min to sit below the line the other pumps draw for its bend, because the optimum sizes it down to 58.5 L/min. It sits 29% above. Danfoss is not just steep, it is high everywhere. No bend of another pump's curve copies that, so even the best row still charges it 1.17 times the drag it pays at its own picks.
The within-family read has a weakness too. Within a family, the drive variants add drag only above the family's largest size. The AZPN family ends at 75.6 L/min and the AZPF at 58.9, so the slope says where each catalogue ends, not how bend relates to level. A fixed sweep of the anchor shows the score falling the lower the pivot goes, with no minimum, so the score alone can't pick the anchor either.
What the run settles is narrower but real: the measured pumps don't support the pivot-79 reading that 7s had to leave open. The two-price rule stays the builder's form, and the next step is a measured second bending family, plus a row that prices a pump's height as well as its bend.
Code and numbers: experiments/gear_pump_anchor.py, gear_pump_anchor_results.txt; the registration and outcome are in the log, milestone 7t. Everything here comes from a simulation on measured excavator motion and digitised data-sheet curves. The bent rows are synthetic, so this is a sensitivity, not a measurement.