A designer coat sells for thousands of dollars. How much of that is cloth and sewing, and how close can a $20 to $50 build get? And with a few hundred dollars of real budget, what can materials science, mechanics and electronics do for clothes?
Those are the two questions of the engineered-clothing research track in this repository. Its brief is section 6 of research/PROJECTS.md. It has produced 45 small projects so far, each a directory under research/garment-* with code, tests, a committed results file and a README. This post pulls out the findings that seemed most useful, with the numbers taken from each project's README.
Who did this. The work was done by AI coding agents (Claude) running unattended in scheduled sessions, with the repository's owner setting the direction and guardrails. No tailor, textile engineer or physiologist has reviewed it.
Read this first
Nearly every one of these READMEs is labelled "stylised", and the label is earned. Cost inputs are round-number assumptions. The two retail prices came from search-result snippets, not fetched pages. Fabric properties are assumed, not measured. The bodies are stacks of ellipses or synthetic gait curves. Nothing was sewn, bought, worn or tested. Read the results as sensitivity studies: which levers matter inside a model, not what a real garment does. None of this is professional engineering, safety or purchasing advice.
1. Factory cost is a small slice of a designer price
garment-teardown builds a bottom-up cost model (pattern area, marker efficiency, cloth price, trims, sewing minutes times wage, overhead) and applies it to two coats with reported prices.
| reported price | modelled factory cost, mid wage | share of price | |
|---|---|---|---|
| trench coat | $2,850 | $204 | 7% |
| wool-blend coat | $3,850 | $493 | 13% |
Across three wage scenarios the factory cost stays under 15% of the price. Fabric dominates it, and a 25% change in cloth price moves the trench's cost by 16% against 3% for the same change in wage. The README is candid about what it cannot explain: its assumed brand and retail markups price the trench at about $900 to $1,200, well short of $2,850, so the rest is an unexplained residual.
The replica costs $46 to $81 (trench) across the wage scenarios. Only the leanest case, very cheap cloth and low-wage or home labour, reaches $29. A follow-up, garment-costperwear, asks how long the $2,850 coat must last to beat a $56 replica per wear: 45.6 times as many wears with no resale and equal care. With the replica lasting 60 wears, the designer coat breaks even on a coin flip only at a median life of about 3,100 wears. Care cost is the lever that most helps the replica.
2. Cutting waste: a cheap heuristic is not enough, a cheap search nearly is
Laying pattern pieces on cloth is a strip-packing problem. garment-nesting compared shelf heuristics, bottom-left placement, simulated annealing and an exact mixed-integer program on 20 small random markers. The exact solver proved optimality on 19.
| method | mean gap to optimum |
|---|---|
| shelf packing (NFDH) | 19.8% |
| bottom-left, largest first | 7.3% |
| simulated annealing, 500 decodes | 0.5% |
Packing true polygons instead of bounding boxes shortened markers by 5% to 43%, most for skirt gores that interlock when turned 180 degrees. At an assumed $7/m², that is worth $0.92 to $6.04 of cloth per garment.
Two follow-ups temper this. garment-nestcost found that with a 1.5 cm seam allowance its coat shell reaches only 0.68 to 0.73 efficiency, below the 0.80 the teardown assumed. garment-cutplan found that putting several garments in one marker saves 9.3% to 10.7% of cloth, but the extra gain from mixing sizes was mostly the optimiser picking lucky markers: re-nested with fresh seeds, the "optimal" plan ended 0.09% longer than the simple one.
3. Sizing: a regular grid is fine until measurements correlate
garment-sizeset asked how many sizes fit a synthetic population within ±4 cm. When bust and hip are strongly correlated (rho 0.9), eight optimised sizes fitted 0.871 of a fresh 20,000 people against 0.677 for a regular grid. When they are uncorrelated, the optimised set was no better than the grid on fresh people, because it overfitted the 300-person sample. garment-sizestretch adds stretch: at rho 0.5, fitting 80% of people takes 8 sizes with no stretch, 4 at 5% stretch, 2 at 10% and 1 at 20%.
4. Thermal comfort: which fabric property matters depends on where you are
garment-thermal extends the ISO 7730 comfort index with four fabric properties. Its no-optics core agrees with the open pythermalcomfort implementation to within 0.0062 PMV over 144 conditions. Inside the model:
- In a 28 °C room, a cloth that passes 80% of body infrared feels like a room 2.4 °C cooler. In a 16 °C room the same cloth feels 7.3 °C colder.
- Outdoors in sun, white against dark cloth is a 7.7 °C span, while infrared transmittance moves comfort by 0.3 °C.
The fleet then checked its own two-node body model against an open implementation. garment-twonodecheck found core temperature within 0.25 °C, but skin temperature off by 0.58 °C on average and up to 2.1 °C. garment-twonodeoptics found the rankings of fabric properties survive the swap, while the sizes move by up to a factor of two. And garment-wearmix weights the levers by where a garment is worn: a fixed infrared-transparent shell pays only if warm rooms take more than 75.5% of wearing time (when the rest is a cool room). White cloth never hurts in the model and is the best lever per dollar in all four wearing mixes it compares.
5. Exosuits: the motor, not the battery, decides
garment-assist models an ankle-assist garment against the metabolic cost of carrying it. The answer rests on one unsourced number, the cost per kilogram of added mass: break-even is 8.9 kg at 1% per kg and 1.8 kg at 5% per kg. garment-assistpower adds a motor and battery. At an assumed 8 N·m/kg motor, the suit weighs about 9 kg, 8.4 kg of it motor, and raises whole-body cost (+1.0% with a 4-hour battery). At 40 N·m/kg it saves 5.7%.
6. Protective and smart fabrics
garment-protective uses a recalled Cunniff-style scaling. It finds that a suit-like panel (at most 3 mm and 1 kg) reaches 0.50 to 0.84 of a 4 kg/m² aramid reference for aramid, and 0.60 to 1.18 for UHMWPE, depending on an unknown exponent. It is explicit that this is modelling only. A garment is armour only after certified testing, such as to NIJ Standard 0101.07.
garment-smart trains an activity classifier on synthetic sensor signals. The one informative result is fragility: if the garment slips (sensor rotation with a standard deviation of 1.0 rad), accuracy falls from 99.9% to 45.2%. Training with slip brings it back to 94.4%.
For RF-shielding cloth, garment-optics finds the weave's gaps, not yarn conductivity, set shielding above about 1 GHz: a 1 mm pitch gives 36.4 dB at 2.4 GHz and about 15 dB at 28 GHz.
Does the work reproduce?
A separate project, swe-audit, re-ran a random sample of 40 research projects. Two garment projects were in it, garment-nesting and garment-markercost. Both reproduced their committed results except for one timing line.
Caveats beyond the models
The track also shows a fleet habit worth naming. Fifteen of the 45 projects are garment-twonode* follow-ups, and the track's own measurement work in swe-value asks whether chains like that are worth as much per token as new milestones. Several are real checks of earlier results, and those checks found real gaps. Others are small variations.
What would move these results from models to evidence is listed in each README as an owner step: sew a muslin, cut a real marker, measure a fabric's infrared transmittance or shielding. None has been done yet.
For how the agent fleet itself runs, see Throughput per unit of supervision.