kerfmaster pin list · Agreed & queued (build after fidelity)
recordmixedStates what the code does today. Auditable against the code — this is the only type that can go stale.
Content last changed 2026-08-03 — computed from the item itself, not typed.
Contract — QC stage
This contract describes the code as it runs today.
takes raw-geometryhealed-geometry
makes topology
fails if A defect that is in the customer's own file reported as ours, or silently healed away.
Contract last changed 2026-08-06 — computed, not typed. Dated separately from the text above, so neither date can speak for the other.
Raised by Aristide 2026-07-30 from a screenshot of the novi part: two jogs, ~10 thou and ~9 thou, on the same element on opposite sides, that were "not healed into a smooth arc".
Both are real, both are in the source DXF, and we reproduced them exactly rather than failing to smooth them:
The Y values match to a thou, so they are the two sides of one element, exactly as described. A zoomed render of the source alone, with no healing applied, shows the step already present.
Two ~90° turns with a short straight between them is the spike signature — a real corner — not a sustained gentle turn (see Spike vs. sustained). Our rule is that true source corners stay sharp. The step is also larger than the 10 thou gate, so smoothing it would move the cut path further than the tolerance we promise.
Asking for finer healing would NOT remove these — it would preserve them harder. Tolerance controls how closely we follow the customer's geometry. Tightening it from 10 thou to 1 thou makes us follow a 10 thou source step more faithfully, not less. Tolerance cannot decide whether a piece of source geometry is wanted.
So there are two separate operator controls, and conflating them is the trap:
A third, from the same conversation: pass-through, for files accurate enough that they should not be healed at all. Pass-through must still measure and report deviation — "do not heal" must never quietly become "do not check".
Kerf is 0.007". A 10.5 thou step is about 1.5 beam widths — even cut exactly, the laser can only render it as a ragged little notch. That is an argument for cleaning these up rather than honouring them.
The clean sweep, done 2026-08-02. Every straight segment under 20 thou in the source polylines, with the kind of its two neighbours — eight, not four, and they are two different classes:
The pinned list of four mixed the two classes and missed the biggest one: 15.78 thou is larger than anything named here before, and it is the same arc-straight-arc structure, so any cleanup rule written against the 10.48 will catch it too. The pin’s own OPEN asked for exactly this sweep.