Status: internal measurement against a simulator, not peer reviewed, and not validated against a molded part — no press trial has been run. "Label quality" below means agreement between two grid pitches of the teacher solver. Tier: a negative result — the obvious corrective move made things worse, and the reason is worth more than the move would have been.
1. Summary
Half our training corpus carries labels that are not pitch-converged — the fast-filling half, per the pitch-convergence note. The obvious move is to throw it away and train on clean labels only.
Doing that makes the operator worse on the clean regime itself: Spearman 0.939 against 0.955 for the model trained on everything, measured on the 84 slow-fill heldout parts — the very parts the clean-label model was supposed to be better at.
Down-weighting the unconverged half instead of dropping it changes nothing (0.955, a tie).
At roughly two thousand parts, this operator is data-limited, not label-limited.
2. Method and pre-registered bars
Three arms, same architecture, same heldout split:
- all-parts — the incumbent, 2,084 parts
- slow-only — the 700 slow-fill parts, converged labels only
- weighted — all parts, unconverged half down-weighted
Bars fixed before training: adopt a variant at +0.010 Spearman, declare no-harm at −0.010. Scored on the 84 slow-fill heldout parts.
3. Results
| arm | Spearman (84 slow-fill heldout) | worst-decile IoU | verdict |
|---|---|---|---|
| all parts (incumbent) | 0.955 | — | — |
| slow-only, converged labels | 0.939 (−0.016) | −0.043 | breaches no-harm |
| all parts, unconverged down-weighted | 0.955 (±0.000) | — | ties |
The slow-only arm breaches the no-harm bar on Spearman and loses considerably more on localization — 0.043 IoU, which is the number that matters for "where is the coldest region" and is four times the Spearman loss.
4. What the fast parts are contributing
They are not contributing label accuracy; their labels are, by construction, wrong in a way we have measured. What they contribute is geometry.
The operator has to learn a map from shape to field. Every part is a lesson in shape regardless of whether the field attached to it is converged, and a noisy target over unseen geometry still constrains that map more than a clean target over geometry the model has already seen 700 examples of.
The weighted arm tying rather than winning is the cleanest evidence for this reading: if the unconverged labels were actively poisoning the fit, reducing their weight would have helped. It did not move at all.
5. What this implies
The instinct to clean a corpus before scaling it is expensive here and, at this size, backwards. The ordering of effort is: get more geometry first, and revisit label quality when the shape distribution is saturated and the operator stops improving with data.
It also means the pitch-convergence finding, which is a real constraint on what we can ship, is not automatically a constraint on what we can train on. Those turned out to be separate questions, and we had assumed they were the same one.
6. What this does not establish
- This is a statement about ~2,000 parts, not about the method. At ten thousand parts the corpus may well become label-limited and the same experiment flip. The finding is a regime, not a law.
- One architecture, one heldout split, no seed variance reported here. A 0.016 gap is small enough that seed-to-seed spread deserves to be quoted alongside it, and is not in this run record.
- The 84-part heldout is small for a 0.010 bar.
- One material, one press slate (VICTREX 450G, Sodick LP20EH3 defaults).
- The teacher is a solver, and on the fast half it is a solver we know to be unconverged. No molded part has been measured.
7. Records
Pre-registered: docs/2026-09-14-n7-converged-regime-preregistration.md (both
bars fixed there before training). Run data:
runs/20260914-n7-converged-regime/. FutureMold engine at commit fb1023bf.
Colophon — authorship and AI assistance
Author (responsible natural person): Anderson Brunsvold, Future Micro Mold.
The three arms were specified, trained and scored by an AI agent (Claude, Anthropic) in the FutureMold repository under the author's direction; the prose is AI-drafted and human-reviewed. The adopt and no-harm bars were written into the pre-registration before training, which is why a −0.016 result is published as a breach rather than as noise.
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.