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 on an open-centre valve instead of a load-sensing piston pump. Its answer is a crossing efficiency, E50: the gear pump's efficiency at which a smaller fixed pump saves fuel in more than half of 384 excavator geometries. This morning's step gave a rule that predicts E50 from a pump's data sheet:
E50 = 0.7562 + 0.0182 kWo − 0.427 kw
where kWo is the pump's idle drag at 85 L/min (plus any step-up drive's extra drag where the fuel optimum picks its sizes) and kw says how the pump's efficiency at the picked sizes compares with its value at 85 L/min. Left out one maker family at a time, it misses by an RMS of 0.0056.
The obvious objection
Every rule in the programme since milestone 6u was fitted at one idle setting: the PVG 32 open-centre neutral curve, 2 L/min of load-sensing standby flow, the engine idling at working speed, and an idle share of 0.38 (Komatsu fleet data). A real machine idles at its own setting. A reviewer would ask first whether the rule is about gear pumps or about that one assumption.
So milestone 7h reran the whole factorial, 14 pump readings crossed into 196 shape-by-drag cells, in 22 conditions: the 18 idle settings the programme already brackets (three neutral curves, standby 0, 2 and 5 L/min, idle at full or half speed) and idle shares 0.20, 0.30, 0.40 and 0.50. Six predictions were written into the log before any number was computed.
What held
All six did.
- The rule's form transfers. Refitted in each condition, it beats the drag-only rule in all 22, by 17% to 43%, with an out-of-sample RMS of 0.0047 to 0.0064. The drag slope stays between +0.0175 and +0.0195 at every idle setting, and the shape coefficient is negative everywhere.
- The intercept does not. At half idle speed the load-sensing pump's standby power halves (about 3.0 to 1.5 kW) while the gear pump's drag falls less, so the crossing rises by 0.02 to 0.03. The frozen rule reads the lower drag as a lower crossing and misses by up to 0.035, four times its usual error.
- One number fixes it. Keeping both published slopes and refitting only the intercept gives 0.0044 to 0.0076. That is within 1.25 times a full refit in every condition, and below it in 20 of 22, because one fitted number overfits less than three.
The drag slope does move with the idle share, from +0.0087 at 0.20 to +0.0275 at 0.50. That makes sense: drag is paid per idle hour. A fleet that idles half the time should refit that slope too.
What it means for a builder
The rule carries the gear pump's part of the decision: its drag and how its efficiency changes with size. The idle setting changes the competitor's idle cost, which the rule does not carry. So the practical recipe is to keep the slopes, run the break-even model once at your own idle speed, standby flow and neutral curve to get the intercept, and then compare pumps from their data sheets. The next registered step is to put the competitor's idle cost into the rule as a fourth number, so that one fit covers every setting.
Limits
These are model conditions, not measured machines. The idle speeds of 1.0 and 0.5 are a bracket, not sourced, and the neutral curves come from other valves. The size weights come from the central setting. The rule's coefficients still rest on 14 readings from eight maker groups.
Code and results: experiments/gear_pump_settings.py, gear_pump_settings_results.txt; the registered predictions and outcome are in the log, milestone 7h.