Ten of fourteen styles within 0.25 mm of the vendor die.

Because it would not have told you anything you need. A folded render answers “what does this box look like”. The questions that cost money are “will this flat cut”, “is the tuck the right depth for this caliper”, and “will my die maker’s software open the file”. None of those are answered by a render.

Building one properly would have taken the time that went into solving 62 vendor dies. We made that trade deliberately. If a 3D view is what you need, Packaging Dieline is not the right tool and we would rather say so here than after you have paid.

You get told, and you still get your file. The validity envelope switches from Verified to Extrapolating and names the exact problem — which dimension, your value, and the range the reference dies actually covered. For example: W = 200 mm is outside the range these 3 dies sampled (30–60 mm). Geometry is extrapolated.

The engine does not refuse. It stops claiming. The published deviation for that style simply does not apply to your flat any more, and Packaging Dieline will not pretend otherwise. In that state, cut a proof.

It is AC1024, with $INSUNITS = 4 and $MEASUREMENT = 1, so it lands in millimetres rather than unitless drawing units. Curves are LWPOLYLINE segments carrying true bulge values, not polygons pretending to be arcs, and circles are real CIRCLE entities. Layers are named cuts, folds and bleeds, which is what die makers expect to receive.

We write the file, read it back, and compare the geometry to what we generated. The worst disagreement in that round trip is 0.00804 mm. Nothing is rasterised at any point in the pipeline, so there is no resolution to be too low.

You do not have to believe any of that. Open the file in a text editor and look — the steps are listed in the exports section above.

It means we obtained real production DXFs — files that converters actually ran — and solved each style's geometry against them path by path, until the generated flat landed on the reference. The deviation we publish is the largest distance between our geometry and the reference die, across every sampled size and every path. Not an average, not a typical case. The worst point we found.

The die count is how many independent dies a style was solved against. It matters as much as the deviation: 0.017 mm across 4 dies is a stronger claim than a tighter number across one. It is also why the pillow box, at 2 dies, is flagged even though its figure is small.

Gable box, 1.16 mm. Our worst style that still has a verifier we can run — and the most instructive number on this page. It was published at 4.94 mm, which was a six-die figure. Adding a pair of W = 190 dies showed the die set had never reached the width where the engine actually broke: on all eight dies the honest baseline was 15.07 mm. Re-derived against all eight, it is 1.16 mm worst case, with sheet identity passing 8 of 8 at a worst residual of 0.0592 mm.

The old explanation was wrong too. We described a width-dependent law breaking “at W ≥ 152 mm”. It is not a break, it is a clamp: the lock depth is sec + min(sec, 80), one constant, exact to 0.0592 mm on all eight dies. That puts the knee at W = 160/161, not 152, and the same saturation shows up independently in the lock cutout and the V-tab. So the engine is not degrading above 152 mm — it is following a law we had mislocated. Model 112310, 8 dies, 5 distinct calipers (0.27, 0.35, 0.50, 1.00, 1.50), the widest caliper spread of any family in the set. Usable, with care: still cut a proof before a run.

Two things are still open, and we would rather say so. The caliper step threshold is bracketed only to (0.5, 1.0] — a die at t = 0.7–0.9 would close it. And the V-tab knee is bracketed between W = 152 and W = 180; a die at W = 160–170 would pin it.

Reverse tuck end, 1.11 mm. Fine for layout and quoting. For plate, proof it, particularly on squat boxes where H ≤ W — that is where the lip and corner radius shift.

Auto-lock bottom, 3.71 mm, no runnable verifier. The largest figure we publish. It came from a solve we can no longer re-run, because those dies are not on hand. We treat it as unverified in the app and you should treat it as a starting geometry to correct, not a finished die.

123-bottom tuck top, no die at all. A draft. It follows the tuckend tube rules and nothing more. Use it to discuss a structure. Do not send it to plate.

No. The free tier's style list is published on this page, before signup, and it does not shrink. Exports are never metered on any tier — not counted, not capped, not watermarked, not downsampled. If a structure is not included in your tier, you will see that in the style picker before you type a single dimension, not at the download button.

Whenever we solve more dies. Two figures moved in the August 2026 patch: reverse tuck end went from 10.87 mm to 1.11 mm, and gable from a published 4.94 mm to 1.16 mm. The gable one is worth reading twice: the 4.94 was measured on six dies, and adding two wider ones put the honest eight-die baseline at 15.07 mm before the re-derivation brought it to 1.16 mm. The number moved because the evidence grew, not because the engine got luckier. All the old values are still printed in the table, because a tolerance history you can read is worth more than a tolerance you have to trust.

If a number ever moves the wrong way, it will be published the same day, in the same table.