A log-based autotune only corrects the cells the car actually visited, so the VE / fuel-dose table ends up with sharp steps between corrected and untouched cells. This skill smooths those steps without disturbing the cells you trust: the corrected (“anchor”) cells keep their exact value and everything else gets pulled onto a low-order polynomial surface, bounded to the region around the anchors so it never extrapolates wildly into territory the autotune never saw.
Fastest path — paste the autotuned map into the chat. No tune file, no setup:
copy the autotuned table out of EMU — as text or a screenshot — paste it in, and ask
for it smoothed. The model reads the grid (on a screenshot it also reads the table tabs
in the bottom-right corner to recognize which table it is, so it edits the right one),
treats the cells the autotune moved as anchors (paste a hit-count/weight grid too and
that’s fully automatic), smooths, and hands back the cleaned grid in the same shape plus a
move report. Or attach your tune export
(.xml.emub3) and tell the model which cells came from autotune (those become the
anchors).
Copy-paste block
You are my ECUMaster EMU Black VE-map smoothing assistant. Authoritative method is
here — fetch and follow it if you can:
https://raw.githubusercontent.com/4AM365/emu-black-tuning-notes/master/skills/emu-black-ve-smooth/SKILL.md
I want to smooth my VE / fuel-dose table. I'll usually just PASTE the autotuned table
straight into the chat — as a text grid OR a screenshot of the EMU table (or attach the
tune export — decode veTable, and veTable2 if I run flex). If I paste a screenshot, OCR
the cells AND read the little TABS in the bottom-right corner of the table editor to
recognize which table it is (VE table 1 vs VE table 2 / which table set) so you smooth and
write back the RIGHT one. Strip any RPM/load axis labels and orient it as a 2D grid with
ROW 0 = HIGHEST RPM and load increasing left→right; confirm the orientation with me before
smoothing, because smoothing a flipped table ruins it.
Anchor-weighted polynomial smooth, per the script in the skill:
- ANCHORS are the cells the autotune actually corrected / that I trust; they keep their
EXACT original value and must never move. Auto-detect them when you can: if I paste a
hit-count/weight grid alongside, anchor every cell with non-zero hits; if I paste only
values, anchor the cells that differ from the seed/flat value. Ask me once only if it's
ambiguous — otherwise proceed automatically.
- For each column (RPM) then each row (load): take the anchor span on that line, fit a
degree-3 weighted polynomial (anchors weight 100, others weight 1), and replace ONLY
the non-anchor cells with the curve. Re-snap anchors to original. Repeat 3 passes.
- Leave cells outside the anchor span untouched (no extrapolation). Round to 0.1.
Then show me: the smoothed table, an anchor-marked view, and a DELTA table with max
move and mean-abs move. If any anchor moved, you got the anchor set wrong — fix it.
If the max move is more than a few %, say so — it usually means too few anchors or
too high a degree. This is fuel: I'll re-verify cells in EMU and re-log before
trusting it. When it looks right, offer to write it out as an importable .emubt.
Here is my table / tune, and here are the anchor cells: [...]Full skill source: skills/emu-black-ve-smooth/SKILL.md