Design lab
Does the simulator generalize, and can search beat a hand-tuned design?
The bench (P1–P6) measures reliability on the pills used during concept development. This page evaluates three additional questions: which inputs drive the outcome, whether the geometry rules generalize to held-out pill shapes, and whether a search over mechanism dimensions improves on the analytic baseline when both are evaluated on separate held-out seeds. The final section lists the limits that still require physical testing.
1 · What actually drives the outcome?
Variance-based sensitivity (η², one-way ANOVA effect size) computed directly from the bench data already on this site (P3–P6), no new simulation. For each concept and factor, η² is the share of the variance in dose-exact outcomes explained by that factor's level; 0 % means the factor made no difference in this sweep, 100 % means it explained all of it. This tells you which knob is worth turning first, and, just as usefully, which one is not the problem.
2 · Held-out generalization to new pill geometries
Each concept's geometry (funnel angle, pocket/slot auto-sizing, lane height, singulator gap, cup diameter…) is calculated from the pill's length and width rather than from a lookup table or fitted outcome. The five cases below are outside both the eight product presets and the 25-size P2 sweep: an oversize tablet, a micro tablet, a nearly round caplet, an extreme-aspect-ratio capsule, and a nearly round softgel. They were reserved from concept development and debugging. This is a within-simulation generalization test: it evaluates whether the sizing formulas hold on new geometry, not whether the simulated physics matches a physical machine (§4).
3 · Search vs. the hand-tuned baseline: train/test evaluation
The shuttle's pocket sizing is an analytic formula (§2), not a fitted model. A random search over pocket length × depth was scored on 8 search seeds. The selected candidate and the analytic baseline were then both evaluated on 16 held-out seeds that were not used during the search. The held-out result is used for the comparison.
4 · ProDoseBench: a reproducible protocol
The six protocols behind this whole site (P1–P6, plus the sensing Monte-Carlo and the chute sweeps) are a fixed, versioned, seeded specification, not one-off numbers. Anyone with the repository can regenerate every chart on this site byte-for-byte from the same seeds, or point the same trial code (js/trial.js) at a different mechanism.
What's fixed
Pill geometry model, bottle geometry, the physics engine and contact solver, the trial harness (js/trial.js, shared by Node and the browser), the metric definitions (single-pill, dose-exact, jam, damage) and the random seeds for every protocol.
What's swept
P1 reliability (8 pills × 12 fills), P2 size range (25 sizes), P3 adjustment (per-concept setting grid), P4 last-pills / fill level, P5 motor limit (tuned machine), P6 stress test (tight clearance, rigid lips).
Reproduce it
node bench/run.js # P1–P6 → data/bench.js node bench/sensing.js # confirmation Monte-Carlo node bench/chute.js # transport & collection node bench/challenge.js # held-out geometry (§2) node bench/optimize.js # search case study (§3)
Extend it
A new concept implements the same design interface (bounds, control, sense, compat…) and registers in P.DESIGNS, the bench, sensing lab and this page pick it up with no other changes.
5 · What this cannot show
These results are limited to the simulation model. The following questions still require physical hardware and measured data.
Sim-to-real transfer
Nothing on this site has been checked against a physical machine. Friction, restitution, crush strength and drop ratings are typical handbook values for each pill class, not measurements of specific tablets, and the contact model takes the larger static friction of the two materials in contact (Matter.js's convention), which is a conservative bias for low-friction linings. A calibration/validation split against real hardware (build on a subset of pills, predict on a held-out subset, and measure the gap) is the next test. It requires physical pills, a prototype rig, and camera or sensor ground truth that have not yet been collected for this project.
Physical fault injection
The simulator can stall a motor or lose a vacuum seal because a parameter is set that way. It cannot tell you what a real stepper does when a real gear tooth actually skips, or what a real belt does after a thousand hours of wear. That requires a built mechanism.
Sensor ground truth
The sensing lab's six confirmation methods are evaluated against the simulator's own pill positions, a fair test of the method, but the noise model (12 mg load-cell noise, idealized photodiode shadows) is assumed, not measured from a real beam or scale.
Materials data
The materials table on the Labs page uses typical handbook ranges for each material family. Those values should be replaced with supplier-specific datasheets and measured properties before final component selection.