Status: internal measurement against a simulator, not peer reviewed, and not validated against a molded part — no press trial has been run. Tier: two negative results — one kills a method, one kills a hypothesis we would have been happy to be right about, and the second one is useful.
1. Summary
Two pre-registered experiments, both of which failed, in different directions.
A learned pitch correction loses to upsampling. An operator trained on 151 fine boxes to turn a coarse fill solution into its 2× version scores 0.767 against 0.830 for the coarse field simply upsampled. It is worse than doing nothing clever. Its kill rule — no improvement over its own input — fired.
A tile head needs only the entry face. An operator trained on the downstream halves of parts matches the whole-part operator within the declared bars while using no melt-state information at all. The hypothesis under test was that the inflow state matters; it does not, here. At this pitch the freeze-margin fields are local.
2. The box operator
The idea was cheap and standard: keep the coarse solve, learn the residual between it and its refinement, apply the correction. It is the shape of every super-resolution result in the last decade.
| Spearman against the 2× truth | |
|---|---|
| coarse field, trilinearly upsampled | 0.830 |
| learned box operator (151 fine boxes) | 0.767 |
The kill rule was written before the run — no improvement over its own input — and it fired cleanly. We report it because the negative is informative about why: 151 boxes is a small corpus for a residual that, per the discrepancy-field note, is strongly structured but not simple. The operator learned enough to move the field and not enough to move it the right way, which is the worst available outcome — upsampling at least preserves the ordering it was given.
This does not show that pitch correction is impossible. It shows that this one, at this corpus size, is worse than the null.
3. The tiling spike
Here the pre-registration had two bars, and the split between them is the whole result.
| Spearman | worst-decile IoU | |
|---|---|---|
| whole-part operator, downstream cells | 0.922 | 0.543 |
| tile head, same downstream cells | 0.891 | 0.463 |
| no-inflow control | identical to the oracle on every channel | — |
Bar one — "does a tile head work at all?" Passed. 0.891 / 0.463 against 0.922 / 0.543 is inside the declared tolerance.
Bar two — "does the inflow state matter?" Failed. The control that was handed no melt-state information came out identical to the oracle that was handed all of it — on every channel, including the arrival margin, which is the channel most obviously downstream of what the melt was doing when it crossed the cut.
A failed bar that makes the engineering easier is still a failed bar, and the reason to say so is that we predicted the opposite in writing.
4. What follows
Tiling along the flow direction is possible with only the entry face handed over. That is a materially cheaper interface than passing a melt state across every cut, and it is what makes tiling a candidate for large parts at all.
5. The caveat, which is in the record and is not small
The cut design kept the melt state nearly constant across the cut. So "the inflow state never matters" is not what was shown. What was shown is that when the state barely varies, a model cannot extract signal from it — which is close to a tautology and is why the bar was written as a falsifiable prediction rather than assumed.
The experiment that would settle it cuts where the state varies sharply — across a gate, a thickness transition, a weld region — and has not been run.
6. What this does not establish
- 151 boxes. The box operator's failure is a statement about this corpus size, not about residual learning.
- The tiling result is single-direction and single-cut-design, per §5.
- One material, one press slate (VICTREX 450G, Sodick LP20EH3 defaults), one pitch. "Fields are local at this pitch" is the claim; locality at finer pitch is untested.
- No molded part has been measured. Both truths are solver output.
7. Records
Pre-registered: docs/2026-09-14-n8-multiresolution-boxes-preregistration.md
(box operator, including the kill rule) and
docs/2026-09-12-tiling-spike-preregistration.md (both tiling bars and the cut
design caveat). FutureMold engine at commit fb1023bf.
Colophon — authorship and AI assistance
Author (responsible natural person): Anderson Brunsvold, Future Micro Mold.
Both experiments were specified, scripted and run by an AI agent (Claude, Anthropic) in the FutureMold repository under the author's direction; the prose is AI-drafted and human-reviewed. Both kill rules and both tiling bars were written into their declaration files before the runs, which is why a result that contradicted our own prediction is published as such.
No result here has been reproduced by a second implementation or a second party, and none of it has been checked against a molded part.