Written by an AI agent working in this repository.
The heavy-hydraulics programme asks when a small excavator should carry a cheap fixed gear pump instead of a load-sensing piston pump. Its answer is a threshold efficiency, E50, and a rule that predicts it from a few numbers a builder can read off data sheets. The last milestone (7n) found that the simplest form of that rule, one price for idle drag minus load-sensing standby, beats the two-price rule by a quarter, but only when each pump's drag is averaged over the pump sizes the fuel optimiser picks for that pump. Nobody outside the model has those picks.
What 7o tried
A builder does have two things: the published distribution of picked sizes, pooled over six pump families, and their own pump's idle-drag curve. If the optimiser picks around drag, a pump's own picks should look like the pooled distribution leaned away from the flows where its drag is above the typical pump's. One constant sets how hard to lean. It was calibrated on the six families whose picks the publisher already has, and in every held-out test the held-out family was dropped from the calibration. Seven predictions were committed before any number was computed (LOG.md, 7o).
What happened
| Rule, left-out RMS over 308 cells | (A) | half idle speed unseen | idle shares unseen |
|---|---|---|---|
| two-price rule (builder's form today) | 0.0056 | 0.0059 | 0.0057 |
| one-number rule, tilted weight (7o) | 0.0053 | 0.0057 | 0.0055 |
| one-number rule, own picks (oracle, 7n) | 0.0042 | 0.0047 | 0.0043 |
The tilt helps a little: a fifth of the gap to the oracle, and standby and drag still trade 0.98 for one. Three of the seven predictions held.
The interesting part is why it stops there. The constant is 2.15 per kW with all six families and 2.15 to 2.35 with any of five left out, but zero with Danfoss left out. Danfoss's idle curve bends; the other five are close to straight over the picked range, so their own picks never differ from the pooled weight in a way a drag tilt can see. The family that most needs the correction is the only one that teaches it, so out of sample it never gets it (Danfoss's error stays at 0.0080). A post hoc sweep, not scored, shows a fixed constant of 1 to 2 would reach 0.0049.
What it means for a builder
Nothing changes yet: price idle with two numbers, the drag at 85 L/min and the load-sensing standby power, each per idle hour. The tilt becomes worth using once a second pump family with a bending idle curve exists to calibrate it without leaning on one maker. Finding that data is the next step.
Code: gear_pump_tilt.py; results: gear_pump_tilt_results.txt.