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 best builder-side form so far is a "two-price" rule: the pump's idle drag and the load-sensing pump's standby power, each with its own coefficient.
Yesterday's milestone (7r) tried something a pump catalogue could print: a table of the pump sizes the fuel optimum picks, one row per bend of the idle drag curve (drag at 79 L/min over drag at 64 L/min). No measured pump bends like the Danfoss family (1.78 against 1.28 to 1.50 for the rest), so the rows were built by bending the other pumps' curves with a power of flow and running the motion model on them. It beat the two-price rule, 0.0052 against 0.0056 left-out RMS over 308 cases. The log called that margin "well inside what a different synthetic shape could move". Today's run tested exactly that.
Four shapes with the same bend
Each shape gives the bent curve the target bend exactly. The predictions were written into the log before any number was computed:
| shape | what it holds fixed | left-out RMS | Danfoss | row pick at bend 1.8 |
|---|---|---|---|---|
| power, pivot 64 (7r) | level at 64 L/min | 0.0052 | 0.0063 | 68.0 L/min |
| hinge above 64 | the whole curve below 64 | 0.0053 | 0.0066 | 68.0 L/min |
| exponential | level at 64 L/min | 0.0053 | 0.0065 | 68.0 L/min |
| power, pivot 79 | level at 79 L/min | 0.0059 | 0.0110 | 72.0 L/min |
| two-price rule | (reference) | 0.0056 | 0.0061 |
The three shapes anchored at 64 L/min agree to half a litre per minute on every row and beat the two-price rule in all three tests (left-out condition, unseen half idle speed, unseen idle shares). The fourth, a power law of the same bend held at 79 L/min, makes a steeper curve cheaper at every size below 79. The optimum then has no reason to pick smaller, its row stays near 72 L/min whatever the bend, and the table falls back to about the failed pooled table of 7q. Six of eight predictions held; the two that failed were the pivot-79 shape beating the rule and the four shapes agreeing within 0.0005.
What it means
The bend table's gain does not depend on whether the bend is a power, a hinge or an exponential. It does depend on one physical assumption: a pump whose drag curve bends more pays more drag above its lower-quartile size, not less drag below it. That is a property of real pumps, and the one measured pump that bends (Danfoss) cannot be used to settle it without leaking into its own test. Without Danfoss, all four shapes score the same (0.0052 to 0.0053), so the shape matters only for the pump the table exists to help.
So the two-price rule stays the form a builder should use, and a measured second gear-pump family with a strongly bending idle curve is now the test for three things at once: 7o's tilt, 7r's table and today's anchor.
Code and numbers: experiments/gear_pump_bendshape.py, gear_pump_bendshape_results.txt; the registration and outcome are in the log, milestone 7s. Everything here is from a simulation on measured excavator motion and digitised data-sheet curves; the bent curves are synthetic, so this is a sensitivity, not a measurement.