How the stitch-out prediction works, and what it cannot know

The studio shows a percentage next to every design. This page explains exactly what it is, so you can decide how much to trust it.

What is measured on your shape

After the artwork is traced, the engine builds the satin border column by column: several hundred columns at 0.4 mm spacing around each piece, with corner fanning at sharp tips and short stitches on tight inner curves. For each column it records two things:

• d₁, the depth of the column's inner end inside the fabric outline. On a straight edge of a 4 mm satin it is close to 2 mm. At a tip it can fall under 1 mm.
• n, the outward direction of the edge at that column, needed to know whether a placement shift moves the fabric edge in or out at that point.

The 1st percentile of d₁ over all columns is reported as the worst inner-end depth. It is the single number that most often decides the verdict.

The error model

Four things go wrong between the cut file and the finished piece. Each is modelled as a normal distribution with a sourced one-sigma value.

| Error | σ | Where the value comes from | |---|---|---| | Cut size (piece slightly small or large) | 0.25 mm | Laser kerf measurements on fabric and felt, 0.15–0.3 mm; drag-knife cutters on fabric are comparable or worse | | Hand placement, careful (spray adhesive, two marks) | 0.30 mm | Practitioner estimates; placement is the largest term | | Hand placement, normal | 0.50 mm | as above | | Hand placement, quick and rough | 0.90 mm | as above | | Shift during tack-down | 0.15 mm | Fabric creep under the presser foot on a running-stitch tack-down | | Satin pull-in after compensation | 30 % of 0.25 / 0.40 / 0.15 mm | Hatch and Wilcom pull-compensation defaults; residual is a conservative assumption |

Placement error is applied as a random direction and magnitude for the whole piece, so one edge moves in while the opposite edge moves out. The cut error is applied uniformly around the piece. The root sum of squares of the sourced errors at normal care is about 0.68 mm, which agrees with Brother PE-Design's 1.0 mm standard cutting allowance once you add the bite requirement.

The pass rule

For one trial the engine draws a cut error, a placement shift and a direction, then checks every column:

• Caught: d₁ + edge offset must be at least 0.8 mm (the satin bites enough fabric).
• Covered: the fabric edge must not extend past the satin's outer rail (half width minus residual pull-in).

A trial is clean when at least 99 percent of columns satisfy both. Requiring both conditions is what stops the allowance search from simply making the fabric bigger: a bigger piece is caught everywhere but shows past the satin. An earlier version of the model used only the first condition and inflated allowances to 1.3 mm; the second condition corrected it.

Three hundred trials give the percentage. The allowance is chosen first by a coarse search (0.2 to 1.0 mm, 120 trials each), then the winner is scored with 300 trials and again with careful placement so you can see what better technique buys you.

Reading the verdict

• Pass, 90 percent or more. Produce it. Standard care will do.
• Borderline, 70 to 89 percent. It stitches with careful placement: temporary spray adhesive, two alignment marks, no rushing. Read the warning to see which feature limits it.
• Fail, under 70 percent. The geometry is against you. The warning says why: a tip whose d₁ is too low, a feature narrower than the satin, or a design under 80 mm where the same millimetre of error is a much larger share of the shape.

The studio also shows the comparison across satin widths, because widening the satin from 3 to 4 mm typically lifts a woven-cotton design from 74 to 92 percent at normal care, and 5 mm reaches 99 percent.

What the model does not know

The honest list:

• Your stabiliser and hooping. A poorly hooped knit shifts more than 0.15 mm during the tack-down. A cut-away stabiliser under a stretchy fabric reduces it.
• Fusible backing. Fabric with a fusible web pressed on behaves closer to felt: cleaner cut, less fray, lower pull-in. The model treats it as the fabric you selected.
• Needle and thread. A 75/11 sharp versus a 90/14 ballpoint changes how the satin lies on the edge. Not modelled.
• Fabric grain. A piece cut on the bias stretches. Not modelled.
• Your specific cutter. The 0.25 mm cut error is a middle value. A well-calibrated ScanNCut with a fresh blade does better; a worn Cricut blade on loosely woven fabric does worse.

Because of this, the studio asks you to report how the piece came out. Reports with the outcome, machine, fabric, cut method and allowance are the data that will replace the practitioner estimates above with measured ones.

Validation that has been done

• The satin geometry was compared with the Ink/Stitch engine on ten shapes; our inner-rail depth is within 0.2 mm of Ink/Stitch's and always on the conservative side.
• The file writers produce byte-identical output to pyembroidery for DST, PES, JEF and EXP on three reference designs.
• The trace pipeline was run on 85 synthetic SVGs with analytic ground truth and 273 real icon SVGs.
• 225 real embroidery files from Brother, Janome, Wilcom, Premier+, Embrilliance and independent designers were read and round-tripped through every writer.

None of that is a physical stitch-out on your machine. Treat the percentage as a well-founded estimate that ranks designs and settings correctly, and test on scrap first.