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. The answer comes as a threshold efficiency, E50, that a selection rule predicts from numbers on a data sheet. Two milestones back (7n and 7p) a one-number idle rule beat today's two-number rule, 0.0047 against 0.0056 left-out RMS. It only did so when each pump's idle drag was averaged over the pump sizes the fuel optimiser picks for that pump. A builder cannot run that optimiser.
The idea
A publisher who has run the model on the pumps it has could print their picks once, as a table. A builder looks up the row whose idle drag curve looks most like their pump's and averages their own curve over those sizes. Milestone 7q chose one number for "looks like": the bend, drag at 79 L/min over drag at 64 L/min. It tried three tables: the single nearest pump, the three nearest, and all of them pooled. Seven predictions were written down before any number was computed (LOG.md, 7q), and the previous milestone's results were reproduced byte for byte first.
What came out
| Rule (held-out pump's drag read over...) | Left-out RMS, 308 diagonals |
|---|---|
| its own picks, leak-free (7p) | 0.0047 |
| two-price rule, no picks (today's form) | 0.0056 |
| the three nearest pumps' picks | 0.0060 |
| every other pump's picks, pooled | 0.0061 |
| the nearest pump's picks | 0.0075 |
All three tables are worse than the rule they were meant to replace. Only three of the seven predictions held.
Why
One pump explains it. Danfoss's idle curve bends with a ratio of 1.781. The other 13 variants sit between 1.281 and 1.497. The optimiser sizes Danfoss down, to a median of 58.5 L/min. Every table row puts its picks near 72 L/min whichever donor it comes from, because none of them bends. Read over those picks, Danfoss's steep curve is charged 1.467 times the drag it actually pays, and its error goes from 0.0061 to 0.0198. Picking the "nearest" donor is worse than pooling, too. With one outlier class, the nearest donor is just the most extreme of the rest, and a single donor brings its own shape noise along.
With Danfoss left out (a post hoc check, not registered), the three-nearest table scores 0.0051 against the two-price rule's 0.0056. That is about half the own picks' margin. So the table idea works where pumps resemble the ones it was built from. It fails on the one pump the own picks were needed for.
What it means
The own-pick gain lives in the pumps the optimiser sizes down, and a table can carry it only if it has a row for that drag class. Milestone 7o's drag tilt failed for the same reason. A second gear-pump family with a bending idle curve is now what any builder-side version of the one-number rule needs. Until one exists, the builder's form stays the two-price rule: drag at 85 L/min and load-sensing standby, each priced per idle hour.
Code: experiments/gear_pump_table.py; results: gear_pump_table_results.txt.