Status: internal measurement against a simulator, not peer reviewed, and not validated against a molded part — no press trial has been run. Everything below is a statement about the fidelity of a numerical method to itself. Tier: a mechanism — a convergence criterion that turns out to be governed by the process variable rather than the geometric one.
1. Summary
We label voxelized parts with a dimensional 3-D fill solver at a coarse pitch — 2–3 cells across a wall — and derive a channel freeze margin: how far the melt in each cell sits above the material's transition temperature at the end of fill. The product claim we wanted was the field's relative structure, which cells are coldest, not its absolute values.
Refining the grid 2× preserves that ordering on parts that fill in 0.15 s or longer (Spearman 0.93) and destroys it on parts that fill in under 0.04 s (0.39). The split is by fill time. Wall thickness in cells — the quantity a person reaches for when asked "is this grid fine enough" — is not what governs it.
On all 50 parts refined, refinement moves the field warmer, never colder.
2. How we got here, in the order it actually happened
This matters because the first answer was wrong, and the way it was wrong is the ordinary way.
Five parts, refined 2×, said the ordering survived: 0.92. That is a comfortable number and it was measured honestly. Twenty-five more parts said it did not survive in general: pooled over 30 unselected parts the coarse-vs-2× rank agreement was 0.872, under the 0.9 bar we had set.
So we had a failed bar and a pooled number hiding something. The exploratory split was by fill time, which correlated +0.88 with agreement — the strongest candidate among the process and geometry variables available.
That was exploratory, so it does not count as a result. We then drew 10 new slow-fill and 10 new fast-fill parts, fixed the bars in advance (slow ≥ 0.9, fast ≤ 0.7) and ran them:
- slow-fill: 0.928, worst part 0.83 — passes
- fast-fill: 0.392, best part 0.65 — passes (the bar was a ceiling)
Both pre-registered predictions held on parts that had not been looked at.
3. Results
Pooled over all 50 refined parts:
| fill time | n | Spearman coarse vs 2× | worst-decile IoU |
|---|---|---|---|
| ≥ 0.15 s | 19 | 0.928 (min 0.83) | 0.72 |
| 0.10–0.15 s | 4 | 0.867 | 0.55 |
| 0.04–0.10 s | 8 | 0.790 | 0.39 |
| ≤ 0.04 s | 16 | 0.407 (min 0.14) | 0.36 |
The degradation is monotone across the bands, which is what a real mechanism looks like and not what a threshold artifact looks like.
The absolute zero contour does not survive at any fill time. "Frozen here" — the below-datum region — agrees with its own refinement at IoU ≤ 0.33 on 50 of 50 parts, and the below-datum fraction shrinks on every one of them. The relative ordering is usable in the slow regime; the absolute contour is not usable anywhere in this data, and we do not ship it.
4. Why fill time and not wall thickness
At fast fills, the margin field at coarse pitch is the front's numerics rather than cooling physics. There is not enough residence time for conduction to write structure into the field before the cavity is full, so what the coarse grid records is dominated by how the advancing front was discretized — and that is exactly the thing refinement changes.
The direction of the error follows: refinement always moves the field warmer, because the coarse front over-reports contact with cold cavity wall.
5. What this implies, and what we did about it
Real parts fill in about a second. That is inside the converged regime, and it is the reason this did not surface as a product defect. Half of our synthetic corpus is not — synthetic geometry is free, so the corpus drifted toward parts that fill far faster than anything a press would run.
The shipped operator now estimates fill time from cavity volume and injection rate and refuses below 0.04 s rather than returning a field it cannot support.
Two transferable points:
- A convergence study sized at five parts can return a clean number and still be wrong about the population. The failure here was not the measurement, it was the n.
- When a pooled convergence number sits just under its bar, the useful next move is to look for the variable that splits it, and then to spend fresh parts confirming the split rather than reporting the exploratory correlation.
6. What this does not establish
- No molded part has been measured. This is agreement between two grid pitches of the same solver. A converged field can be converged and wrong.
- One material, one press slate. VICTREX 450G, Sodick LP20EH3 process defaults. Whether the 0.15 s boundary moves with material transition temperature is untested.
- 2× is not convergence. Agreement between a coarse grid and its 2× refinement bounds the discretization error from below. Rung 3 of the pitch ladder is running and is not in this note.
- The 0.04 s refusal threshold is set at the bottom band's edge, not at a measured knee.
7. Records
Pre-registered before the confirmatory run:
docs/2026-09-10-v1-v2-validation-declaration.md (sections V2, V2b, V2c, V2d).
Run data: data/verification/v2-pitch/. FutureMold engine at commit fb1023bf.
Colophon — authorship and AI assistance
Author (responsible natural person): Anderson Brunsvold, Future Micro Mold.
The experiments were designed, scripted and run by an AI agent (Claude, Anthropic) working in the FutureMold repository under the author's direction; the prose is AI-drafted and human-reviewed. The bars quoted in §2 were written into the declaration file before the confirmatory parts were drawn, which is the only reason the confirmatory numbers are reportable 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.