Engineered clothing

A bottom-up cost model for a lined woven coat (fabric from pattern area and marker efficiency, trims, sewing minutes times wage, overhead, brand and retail multipliers) applied to two reported designer prices. Under assumed inputs the factory cost is $186–248 for a $2,850 trench and $469–551 for a $3,850 coat, under 15% of the price in every wage scenario, and the assumed multipliers do not close the gap. A cheap-cloth replica costs $46–100 to make; only a very lean case reaches $29–34. Stylised: every cost input is assumed, prices are from search snippets, and fit and durability are not scored.

A flat skirt-tube pattern is draped by position-based-dynamics cloth on two synthetic elliptical-section bodies and scored by circumferential strain and waist ease. The best waist is close to the body’s waist perimeter (×1.01 and ×1.03); tighter patterns strain, looser ones leave ease. A pattern optimised for one body costs 2.1–2.9 on the other against own optima of 0.15–0.75 (arbitrary cost units). Stylised: the body is a stack of ellipses, the cost weights are assumptions, the hem optimum sits at the edge of the search range, and the optimum cost moves from 0.75 to 1.34 on a finer mesh.

Marker making as strip packing. On 20 random 8-box markers an exact MILP proves optimality on 19; shelf heuristics (NFDH, FFDH) are on average 20% and 17% longer than optimal, bottom-left 7.3%, and simulated annealing over placement order 0.5%. On two stylised garments (a shirt and a four-gore skirt, no seam allowance, one ply) true-polygon nesting with 180° turns beats bounding-box packing at the same annealing budget in every case, by about 5–9% for the shirt and 9–43% for the skirt. Stylised: the shapes are invented, polygon results have no exact reference, and the raster is conservative.

Follow-up to the nesting and cost-teardown projects. Adding a 1 cm seam allowance to stylised pieces lowered marker efficiency from 0.69 to 0.63 for a shirt and from 0.71 to 0.51 for a four-gore skirt (the skirt mostly loses its interlocking gores); nesting three shirt sizes in one marker was 9% shorter than three separate markers. Swapping the measured efficiencies into the teardown formula moves a $7/m² trench’s cloth cost from $26 (assumed 0.80) to $33–41. Stylised: invented outlines, one ply, small annealing budget (differences of a few cm are noise), and the coats were not nested, so the cost step is a sensitivity, not a measurement.

Cutting-order planning as an exact integer program (marker mixes and plies) over 34 nested markers of a stylised three-size shirt. Laying several garments per marker saves 9–11% of cloth against one garment per marker on 20 random orders. Mixing sizes adds a further 0.4–3.7% in sample, but re-nesting the markers the plan picked with fresh seeds leaves a +0.09% difference: the gain is within annealing noise (about 5%) and partly selection of luckily short markers. Stylised: invented shirt, linear size grading, small annealing budget, mixes of at most four garments, free setup cost.

Follow-up to the nesting and cost-teardown projects (stylised, preliminary). A coat drawn as 10 shell and 5 lining polygons, grown by a 1.5 cm seam allowance and nested by annealing at cloth widths of 110, 140 and 150 cm, gives marker efficiencies of 0.68–0.73 for the shell and 0.59–0.68 for the lining, against the fixed 0.80 assumed earlier (0.77–0.82 without allowance). Transplanted into the teardown trench model, the efficiency alone raises factory cost by about $20–30; the larger uncertainty is the pattern area, which these invented pieces do not settle. One ply, one size, my own shapes, no exact reference.

Stylised and preliminary: a synthetic bust and hip population (assumed means 95 and 100 cm, sd 8 cm, no survey data) with a fit tolerance of ±4 cm. Choosing k sizes to fit the most people is maximum coverage; solved exactly (integer program), greedily and as a regular grid on 300 people, 3 seeds. At bust–hip correlation 0.9, exact sizes fit 12–25 points more people than the best grid and 6 sizes reach 80%; greedy is within 3.7 points of exact. At correlation 0 the gap is small, and on 20,000 fresh people the exact set is no better than the grid, so the optimiser overfits small samples. Limits: Gaussian two-measurement population, pass/fail box fit, no cost of extra sizes.

Stylised and preliminary: follow-up to the size-set project on a synthetic bust and hip population (assumed means 95 and 100 cm, sd 8 cm, no survey data). A body fits a garment if it is at most 8 cm smaller and no more than stretch larger than the nominal size, in both measurements; the best k sizes are found exactly (integer program). Sizes needed to fit 80% fall from 6 to more than 8 with no stretch to 2–3 at 10% and 1 at 20%, and the loss on fresh people shrinks too. The stretch model is idealised: usable stretch is far below rated stretch, so these are upper bounds.

A Fanger/ISO 7730 PMV model and a Gagge-style two-node model extended with fabric infrared emissivity and transmittance, solar absorptance and vapour permeability. The no-optics PMV matches an open-source ISO implementation to 0.006 over 144 cases. Indoors without sun, infrared transmittance (0 to 0.8) is worth about 2.4 °C of comfort at 28 °C, more than cutting insulation from 0.7 to 0.3 clo (1.8 °C), but makes a 16 °C room feel 7.3 °C colder; in sun, solar absorptance (7.7 °C span) dominates the infrared properties. Stylised: assumed fabric properties, a parallel-path radiation model with no reference data, and recalled two-node constants.

Runs the two-node model from the thermal-comfort project against pythermalcomfort over 108 conditions (18–36 °C, 30–80% humidity, 1–2 met, 0.3–1 clo). Core temperature agrees within 0.25 °C everywhere; skin temperature differs by 0.58 °C on average and up to 2.1 °C, mostly in cool air; evaporative loss differs by up to 14 W/m² in hot humid air. The direction of change with insulation and humidity agrees in 23 of 24 cases. This compares two implementations, not either against people; causes of the skin-temperature gap are untested and the optical extension is not covered.

Rebuilds the two-node model with nine switchable choices (body area, constants, convection, clothing radiation, skin mass fraction, vasodilation gain, sweat caps, update order, step count), each set to the repository's version or to the pythermalcomfort version read from its source. With all nine set to the reference's, the model matches it to 0.008 in every output over 108 conditions. Switching one at a time, the skin-temperature gap (mean 0.58 °C) falls most with the variable skin mass fraction (to 0.36), the convection coefficient (0.43) and the vasodilation gain (0.44); body area, constants, update order and step count each change it by under 0.01 °C. A comparison of two implementations, not of either against people; the effects interact, so the attribution depends on the order.

Adds infrared transmittance, outer emissivity, solar absorptance, radiant temperature and solar load to the switchable two-node model and compares each property’s effect under the repository’s design choices and under those of pythermalcomfort. Over four stylised scenarios the properties rank identically by dry-heat-loss effect; sizes agree within 3% in a 28 °C room but differ by up to about 2× in cool air (traced mostly to the variable skin mass fraction) and in sun (convection coefficient); the one sign change is under 0.1 °C. The reference has no optics, so its optical terms are my own linearised construction without an external check. A comparison of implementations, not of people or measured fabrics.

Re-runs the switchable two-node comparison of fabric optics (infrared transmittance, emissivity, solar absorptance, insulation) for 24 h of model time instead of one hour (stylised, preliminary). In the 28 C room, with hot walls and in sun the one-hour effects are already steady. In the 16 C room the one-hour disagreement between the repository’s design choices and those of pythermalcomfort (ratio 1.9–2.7) shrinks to 0.89–0.98 at steady state, so it was mostly a transient-speed difference. A model fixed point, not a claim about people; no fabric or person was measured.

Re-runs the two fixed states of a gated textile (infrared transmittance 0.15 closed, 0.8 open) from an earlier model in the switchable two-node comparison, under two sets of design choices and at 3 h and steady state (stylised, preliminary). Opening the gate lowers skin temperature by 0.20–0.25 C in a 31 C room and 1.4–1.5 C in a 16 C room, and moves 5.5–6.2 W/m² from evaporation to dry loss at 31 C; the choice sets differ by about 0.05 C there. Limits: two fixed states with no switching dynamics, evaporative heat loss rather than sweat mass, no validation against people.

Adds a skin-temperature-triggered gate (hysteresis band, first-order actuation lag) to the switchable two-node comfort model (stylised, preliminary). In constant 31, 24 and 16 C rooms over 3 h the gate settles into one of the two fixed states with at most one change, so it cannot beat the best fixed state; a 10 min lag only delays the start (open 94% vs 98% of minutes at 31 C). Threshold placement matters more: a 32–35 C band opens under one set of design choices and never under the other. A mostly null result; it needs changing conditions to show value. No fabric or person was measured.

A stylised, preliminary follow-up to garment-twonodeswitch. One step in room temperature at minute 90 of 180 in the switchable two-node model, scored by mean |skin temperature − 33.7 C| (a proxy, not a comfort index). After a step to a hotter room the gate beats both fixed states and lies within about 0.02within about 0.02–0.07 C of an oraclenbsp;C of an oracle that knows the schedule; after a step to a cooler room it gains little, and a 10 min actuation lag can make it worse than staying closed. Transmittances are assumptions; no fabric or person was measured.

A stylised, preliminary follow-up to garment-twonodestep. One step in metabolic rate (1 to 2 met, or back) at minute 90 of 180 in the switchable two-node model, scored by mean |skin temperature − 33.7 °C|. In a 20 °C room the gate never opens; at 24 °C it beats both fixed states under this repository’s choice set and ties fixed closed under the reference’s; at 28 °C staying open is as good as any policy; after walking stops a 10 min actuation lag erases most of the gate’s gain. Bands were not tuned and results depend on the choice set. Not a claim about any real textile.

A stylised, preliminary follow-up to garment-twonodemet. A skin-temperature gate band is tuned by exhaustive grid search on rest-to-walk and walk-to-rest steps in 22, 24 and 26 C rooms, then scored unchanged on held-out rooms (23, 25, 20 and 28 C). The tuned band (a narrow 33.6–33.7 C) beats the untuned 33.2–34.2 C band and both fixed states on both held-out sets under both choice sets, by 0.01 to 0.06 C against the untuned band. It also beats the four-schedule oracle on training scenarios, because the oracle is restricted, and it chatters (up to 24 switches in 180 min) with instant, free switching, so the gain may not survive actuation lag or cost. Six training and eight held-out scenarios from one model; no real fabric or person.

A stylised, preliminary follow-up to garment-twonodeheld. The skin-temperature gate band is retuned by exhaustive grid search for each actuation lag (0, 2, 5, 10 min) and each cost per switch (0, 0.01, 0.03 C), then scored on held-out rooms. A 10 min lag cuts the tuned gate’s gain over never switching from 0.23 to 0.15 C (“between” rooms, this repository’s choice set). A switch penalty widens the band and removes the chatter. The tuned band is not always better: under the reference choice set it loses to the untuned band by up to 0.035 C in 3 of 24 cells, and once to never switching (0.052 C). Six training and eight held-out scenarios, assumed lags and penalties, one model; no real fabric or person.

A stylised, preliminary, negative follow-up to garment-twonodelag. A transmittance that ramps linearly with skin temperature is tuned by exhaustive search (72 bands) for each actuation lag (0 to 10 min) and scored on held-out rooms against the tuned on/off gate. Proportional wins in 5 of 16 cells and every difference is within 0.010 C: the search picks bands only 0.1 to 0.2 C wide, i.e. nearly on/off. Forcing a softer ramp (at least 1.4 C wide) costs 0.02 to 0.07 C. Switching is free in this run, so a switch penalty could change the answer. Six training and eight held-out scenarios; no real fabric or person.

A stylised, preliminary follow-up to garment-twonodeprop. Both gates are retuned on training rooms for each actuation lag and a per-swing penalty (one full close-to-open swing = 1 for either gate; the penalty scale is an assumption), then scored on held-out rooms. The proportional gate has the lower cost in 52 of 80 cells, by at most 0.096 C on a proxy of order 0.3–1.5 C; at the highest penalty the gain holds in the between rooms (8 of 8) but mostly reverses in the outside rooms (2 of 8). Simulated comfort, not a claim about any real garment.

A stylised, preliminary follow-up to garment-twonodepen. Both gates are retuned on a wider set of rooms (20–28 C instead of 22–26 C) and scored on rooms held out of both. At the highest penalty, proportional wins 7 of 8 cases at 21 and 27 C when tuned wide, against 4 of 8 when tuned narrow, so the reversal is partly an artefact of the training range. At 18 and 30 C the gates differ by at most 0.007 C, which this model cannot resolve. Two penalties only; simulated comfort, not a claim about any real garment.

A stylised, preliminary follow-up to garment-twonodewide. Both gates are tuned on one person (0.5 clo, 70 kg) and scored unchanged on other clothing levels and body sizes. Body size barely matters (retuning gains at most 0.002 C), clothing does: retuning an on/off band for 1.2 clo gains up to 0.375 C, while the proportional gate's gap stays at most 0.075 C. Proportional wins every cell except three at 0.3 clo and penalty 0.1. Mass and height enter the model only through body area and heat capacity; simulated comfort, not a claim about any real garment.

A stylised, preliminary follow-up to garment-twonodestep. Exhaustive search over open/closed schedules (at most 3 toggles on a 10 min grid, instant actuation) in the switchable two-node model, scored by mean |skin temperature − 33.7 C| (a proxy, not a comfort index). One toggle is enough (more improve the score by at most 0.001 C). The best toggle is not always at the step: it is early or late after a step down. The gate is within 0.02 C of the optimum after a step up but 0.06–0.11 C away after a step down (0.18 with 10 min lag in one case). One person, two assumed transmittances, a coarse grid; not a claim about any real garment.

Joins the thermal comfort model to the garment cost chain for five fabric levers (infrared transmittance, low emissivity, white, thinner, thicker cloth). Levers combine sub-additively (infrared-transparent plus thin: +2.58 °C jointly, +3.25 summed), and in each scenario only one treatment helps, so a price premium sets cost, not ranking. At assumed premiums the treatments cost $0.5 to $19 per °C on a $20–50 replica trench and $3 to $114 on a designer one, because the retail chain multiplies the premium about 6×. Stylised: every premium and the thickness pricing are assumed, not quoted, and the comfort model has no reference for its optical extension.

Weights the comfort benefit of fabric levers by wearing time across a warm room, sunny outdoors and a cool room, and adds an idealised gated infrared lever. A fixed infrared-transparent shell has positive expected benefit only if the warm room takes more than 75.5% of wearing time (against the cool room); white never harms and has the best benefit per dollar in all four mixes tried. Across 66 wearing mixes the best garment changes with the mix, and the gate is chosen in 65 of 66 mixes at the fixed shell's premium but in 18 of 66 at twice it and none at four times. Stylised: premiums, gate cost and wearing shares are assumed, and the gate is an upper bound.

Applies a Wright learning curve (unit time falls by a fixed share per doubling of output) to the replica trench coat of garment-teardown ($43 materials, 76 experienced-operator minutes). With an assumed first coat 4× experienced time and an 85% rate it takes 20 coats to get within 2× of experienced time; with 8× and 90% it takes about 9,100, so the answer is set by the two assumed parameters. The first coat stays under $50 only if the sewer’s time is valued at $1.45/h or less ($4.28/h for the 100th). In a factory run the learning surcharge is under 8% after 200 coats. Stylised: learning rates are assumed, not measured.

Combines the modelled replica costs and reported prices of garment-teardown with an assumed durability layer. With no resale and equal care the $2,850 trench must outlast a $56 replica by about 46× (52× for the $3,850 coat). With lognormal lifetimes (log-sd 0.5) and a 60-wear replica, the designer coat needs a median life near 3,100 wears to win half the time and wins about 6% of the time at 1,000; designer care of $1 per wear pushes the needed median to about 73,000. Lifetimes, care and resale are all assumed, not measured. Stylised.

Lets the wearer own two garments (a shell and a layer) and wear whichever is most comfortable in each of three scenarios, searching all lever pairs under one total added-retail budget on the replica trench. With thin cloth allowed the pair beats the best single garment in 58 of 66 wearing mixes at $50, but the apparent $20 gain is an artefact: thin cloth has a negative assumed cost that funds the second garment. With thin cloth removed the gain is 0 at $20, +0.49 °C (54 of 66 mixes) at $50 and +1.02 °C at $100, usually a warm-weather infrared-transparent shell plus a cold-weather thick low-emissivity coat. Stylised: premiums, wearing shares and fabric properties are assumed, garments are not stacked, and no real garment was measured.

Re-runs the two-garment wardrobe search of garment-wearpair with a wearer who picks a wrong option in a share e of scenarios, re-optimising each purchase at each e (replica trench, 66 wearing mixes, thin cloth removed). If a wrong pick is a random wrong option the pair’s edge holds (+0.55 C at e = 0, +0.57 C at 0.3, $100 budget), because errors hurt a single garment’s all-or-nothing choice more than an overlapping pair. If every mistake is the worst owned option the $100 edge falls to +0.09 C at e = 0.3, with a break-even e of median 0.72 and as low as 0.04. At $50 the edge is negligible either way. Stylised: error rates and premiums are assumed, no wearer behaviour was observed, and the conclusion flips between the two error models.

A thin-sheet shielding model of a conductive square weave (sheet resistance, aperture bound, metal mass) plus a Monte Carlo of random yarn breakage. Every design tried has a sheet shielding of 54–87 dB, but the aperture bound limits it: 36 dB at 2.4 GHz for a 1 mm pitch, about 15 dB at 28 GHz. Breakage raises resistance as the effective-medium estimate 1/(1−2f) up to 30% broken segments and the weave stops conducting near 50% (bond percolation). Stylised: plane wave, ideal crossings, recalled constants, no measurement; covers only the RF part of the optics milestone.

A two-layer resistor network of a conductive weave: warp and weft touch only through a contact resistor at each crossing, and yarn segments break at random. On an intact fabric the contacts carry no current, so they have no effect at all. Once yarns break they carry the bypass currents: with 20% of segments broken, the median sheet resistance is 1.74 segment-resistances with perfect contacts, 2.11 with contacts equal to one segment, 4.07 at ten and 16.7 at a hundred. For a stylised stainless weave that is 59.6 dB against 52.3 dB at ten segments (thin-sheet only, aperture limit excluded). Stylised: contact values are a sweep, not measurements; one lattice size, 20 draws, independent breaks.

Follow-up to the contact-resistance model, on a 24×24 conductive weave with 60 draws per cell and bootstrap intervals. When contacts next to a broken yarn are worn (contact resistance ×k), the median sheet resistance at 20% broken and a base contact of one segment rises from 2.05 segment-resistances (k = 1) to 3.03 (k = 10). Clustering matters more: a fold band across the current that holds the mean break fraction fixed but breaks 4× as often inside the band leaves the weave open in 57 of 60 draws at 20% broken, while the same band along the current direction lowers the resistance to 1.43. Stylised: square lattice, one lattice size, a rule-of-thumb wear multiplier rather than a measured one, no flexing physics, and no measurement.

A transfer-matrix model of dielectric quarter-wave stacks on a flat nylon-like substrate, with colour computed under an equal-energy illuminant. Six-pair stacks reach 99% reflectance with 2–3 g/m² of oxide and saturated colour, but the peak blue-shifts from 620 to 540 nm by 60° incidence (ΔE₇₆ 48), and 10% layer-thickness error moves hue by ΔE≈10 while roughly preserving chroma. Stylised: flat substrate, constant indices, no fibre scattering.

Eight-layer thin-film stacks on a flat nylon-like substrate, optimised to minimise the colour shift between 0 and 15–45° incidence while keeping chroma C* ≥ 40 and hue near a quarter-wave target. At an equal budget of 3000 evaluations (10 seeds), Nelder–Mead with restarts cuts the median mean ΔE₇₆ shift from 12.4 to 1.4 (red target) and from 25.9 to 5.6 (blue), beating random search and differential evolution, but chroma falls toward the floor and oxide mass rises about 1.8–2.5×. Stylised: flat substrate, three sample angles, recalled indices, no fabrication tolerance; no LLM involved.

The 8-layer stacks optimised in garment-stackopt are fabricated virtually with independent Gaussian thickness error (3 and 5%) and scored on 200 held-out draws. The nominal optimum (angle shift 1.4 ΔE₇₆) degrades to a mean cost of 24.9 at 3% error, worse than the plain quarter-wave stack (13.2), and only about half of draws keep an acceptable colour. A robust objective averaged over six error draws, at equal total cost, recovers part of this (red target, 3%: 11.7 against 13.2) but not the colour drift, and at 5% error or the blue target the quarter-wave stack is hard to beat. Stylised: flat substrate, no dispersion, uncorrelated errors, 10 seeds, no significance tests, two optimisers.

Follow-up to garment-stackrobust. The 8-layer stacks are trained on a CVaR objective (mean of the worst quarter of 8 sampled thickness errors) or on errors with a shared rate-drift component (ρ = 0.5), then cross-evaluated on 200 held-out draws under both error models (5% error, 8 seeds, 3000 evaluations). Negative result: no optimised design beats the plain quarter-wave stack (red, independent error: mean cost 14.1 against 16.5 for the best optimised one). CVaR trims Nelder–Mead's 90th-percentile cost (47.1 to 43.1) while raising its mean, and correlated-error training did worse. Stylised thin-film model; nothing was coated or measured.

The 8-layer stacks of garment-stackopt re-optimised on a grid of chroma floors and oxide-mass caps (Nelder–Mead restarts, 3000 calls, 5 seeds per cell). For the red target at a chroma floor of C* 40 the best mean angle shift is 0.68 ΔE₇₆ at a 4.0 g/m² cap, 2.71 at 3.0, and no run is feasible at 2.0 or below; raising the floor to 60 costs more (2.84 at a 4.0 cap). Mass binds before chroma. Stylised: flat substrate, three sample angles, untuned penalties, a coarse grid and a single optimiser, so cell values are noisy and the front is not smooth.

The constrained problem of garment-stackpareto with the layer count varied from 2 to 12 (Nelder–Mead restarts, 3000 calls, 8 seeds per cell). For the red target at C* ≥ 40 and 4 g/m², 2 and 4 layers reach a best mean ΔE of about 3.4, while 6, 8 and 10 layers reach 0.5–0.8; medians are several times higher, so the ranking among 6–10 is within noise. At 12 layers few runs were feasible, probably a search limit at equal budget rather than a property of the stack. Stylised: flat substrate, three angles, no thickness tolerance; no coating was made.

A one-joint ankle model of an exosuit-style garment: a suit adding a fraction of the biological push-off torque cuts modelled whole-body metabolic cost by 3.1, 8.8 and 14.1% at gains 0.1, 0.3 and 0.5, but a delay of 0.25 of the gait cycle removes the benefit and a stiff passive spring raises ankle cost 17%. Added mass erases the saving at 8.9 kg if each kg costs 1% of whole-body power and at 1.8 kg if it costs 5%; the break-even scales as 1/sensitivity, a number the project does not supply. Stylised: lumped muscle, prescribed gait, assumed constants; not OpenSim or MyoSuite.

Extends the ankle exosuit model with a power train: motor mass from peak torque and a torque density, battery mass from electrical energy and hours of use, and the resulting mass fed back into metabolic cost. With an assumed 8 N m/kg motor the gain-0.3 suit weighs about 9 kg, 8.4 of it motor, and raises whole-body cost 0.2–1.9%; the sign flips between 8 and 20 N m/kg, while eight hours of battery add under 2 kg. Stylised: every power-train constant is assumed, not sourced.

Extends the exosuit power-train model with where the mass sits: a kg at the shank costs an assumed multiple of a kg at the waist, and a waist layout pays a Bowden-cable loss (efficiency 0.6–1.0) and keeps 0.3 kg of cable at the ankle. At gain 0.3 with a 4 h battery the waist layout wins once the shank penalty exceeds 1.03–1.39× the waist one; at ratio 3 and 8 N m/kg the shank suit is +20.7% against +3.3% for the waist suit. The ratio is assumed, not sourced, so the result is the break-even, not a prediction. Stylised.

Replaces the assumed Bowden-cable efficiency in the shank-versus-waist model with the capstan law exp(−μθ) for friction μ and total bend θ. At gain 0.3, a 4 h battery, 8 N m/kg and an assumed penalty ratio of 3, the waist layout stays ahead until about 31 rad of bend at μ = 0.1 (about 15 rad at 0.2), far beyond a hip-to-ankle route, so energy lost to friction does not decide the layout. Hysteresis, backlash and pretension are not modelled, and μ and the ratio are assumed. Stylised.

A Cunniff-style specific-toughness scaling (U*) for nylon, aramid and UHMWPE with a 0.5 m² torso-panel mass, cost, thickness and stiffness model. At equal weight UHMWPE gives 1.49× aramid's V50 in this scaling, and needs 0.20–0.67× the areal density depending on an unknown exponent. Under a 3 mm, 1 kg suit-like limit the panel falls short of a 4 kg/m² aramid reference except for UHMWPE with an optimistic exponent. Stylised: recalled formula, assumed properties and unsourced prices, no ballistic physics and no protection claim.

Synthetic chest, thigh and wrist accelerometers and knee and elbow strain channels, six activities, a softmax classifier on spectral features and unseen simulated people. Two sensors reach about 99% and all five 99.9%, because the invented classes are easy to separate. A model trained on clean data falls to 81, 61 and 45% when the IMUs rotate by 0.3, 0.6 and 1.0 rad (a garment slipping), while training with 0.6 rad of slip holds 94–99%. Strain-only stays at 97% under slip by construction. Stylised: sinusoidal signals, no real dataset, no drift or hysteresis.