There is no single accuracy number FDM dimensional capability cannot be described responsibly by saying that a printer is “accurate to 0.1 mm.” An outside edge, vertical bore, horizontal hole, thin wall, long plate and tall pillar respond to different error sources. Accuracy is closeness to nominal; repeatability is the spread between repeated parts; resolution describes the smallest distinguishable feature in a particular sense. These terms are not interchangeable. A process that repeatedly produces a dimension 0.25 mm small can be repeatable but inaccurate and may respond to a controlled correction. A process that alternates unpredictably between over- and undersize remains risky even if one cube is perfect. The objective is therefore a documented map of what a specific machine–material–profile combination can manufacture, in which direction, on which geometry and with what spread—not a single impressive measurement.

CAD size, tolerance and clearance are different decisions A 10.00 mm CAD dimension is nominal. Tolerance defines the acceptable lower and upper limits of the manufactured feature. Clearance is the intentional difference between mating parts, and the value needed for a free lid, sliding rail, hand snap, screw guide or press fit is different. Slicer XY compensation only changes a systematic process offset; it does not choose functional tolerance for the designer. Applying the same global scale to a pin and hole may even worsen their relationship because internal and external contours do not behave identically. First measure the manufacturable distribution, then choose nominal sizes and clearance for the function, and only then compensate a stable, directionally proven error. This separation prevents a production correction from becoming hidden design intent and makes future migration to another printer possible.

Why the 20 mm cube is misleading A calibration cube is a useful quick health indicator. It can reveal a gross axis error, skew, ringing or a first-layer problem. It contains no range of bores, thin walls, long distances, inside and outside corners, alternative orientations or repeated samples. Caliper jaws may also touch a corner bulge, seam or elephant foot rather than a stable wall field. A 0.1 mm error over 20 mm does not prove how a 200 mm component behaves, nor whether the error is a constant offset or one that grows with length. Use the cube for a daily check, not as a machine-capability study. A tolerance system requires several geometries and sizes in multiple directions, plus at least three repeats. The coupon should resemble the actual product family rather than a popular benchmark chosen for convenience.

What the NIST artifact teaches desktop FDM users The NIST additive manufacturing test artifact is not a direct recipe for a hobby printer, yet it provides a valuable methodological principle: characterize a system through multiple geometric and surface features. Its approach covers planes, holes, pins, thin elements, angles, circles and surface properties. A desktop workshop does not need to reproduce the complete standardized-style object. Build a modular coupon with outside X/Y dimensions, vertical and horizontal holes from roughly 3 to 20 mm, several pins, walls from one line to multiple lines, steps and Z heights. Each feature needs an unambiguous measurement method, and the collection should represent real jobs. Keep the artifact small enough to repeat routinely but not so crowded that minimum-layer-time effects or local heat accumulation become the dominant experiment.

A number without a measurement plan is not evidence Before printing, define what will be measured, where, with which instrument and how many times. On an outside dimension, specify jaw location and avoid the seam and lower flare. For a bore, decide whether a two-point caliper, gauge pin or bore instrument is appropriate. A budget caliper displaying 0.01 mm increments does not have 0.01 mm measurement uncertainty. Jaw parallelism, contact force, burrs, layer texture, temperature and operator technique all contribute. Measure every feature at least three times, verify zero, and check the instrument against a known reference for critical work. Record nominal value, individual readings, average, range, deviation and acceptance limit. Extra decimal places cannot create real confidence and should not be reported beyond the capability of the method.

Part temperature is a measurement variable A newly removed polymer part can continue cooling and shrinking. A warm measurement cannot be compared directly with a component conditioned for a day, especially for large engineering-polymer geometry. Standardize waiting time and environment—for example, measure only after complete cooling and several hours in the same room. Moisture-sensitive materials, including many polyamides, may change later, so conditioned and service states may need separate definitions. Annealing, washing, painting, machining, assembly stress and creep can also move dimensions. If a component enters service only after post-processing or at elevated temperature, final acceptance should reflect that state rather than the convenient moment after printing. The protocol must say when a dimension becomes official.

Mechanics before software correction A loose belt, rocking carriage, dirty linear guide, unsecured hotend, skewed gantry or non-square axes create errors that global scaling cannot repair. Inspect assembly to the manufacturer's procedure, drive condition, bed stability and nozzle seating. An oval hole might indicate unequal X/Y scale, but also backlash, belt condition, acceleration or squareness. Growing lateral displacement on a tall wall may identify an orthogonality problem. Before changing steps per millimetre or an axis scale, prove a linear scale error over a long distance, in more than one direction and on repeated builds. On a modern belt system with known pulley geometry, the factory motion constant is rarely the first variable to edit. Do not bake extrusion and thermal errors into kinematics merely because one cube was wrong.

Flow changes where walls are placed Excess extrusion multiplier can create wider roads, bulged corners, smaller holes and oversize outside walls; insufficient flow leaves gaps, thin walls and weak bonding. Validate flow for the actual filament diameter, feeder state, nozzle and temperature. Prusa's guidance treats extrusion-multiplier tuning as a material-flow operation, not axis calibration. Measuring a single-wall box can provide evidence, but it is not decisive by itself because actual line width, slicer assumptions and jaw contact complicate interpretation. Check top surfaces, mass, wall structure and a representative part. Pressure advance or Linear Advance manages changing nozzle pressure during acceleration; a poor value distorts corners and short segments. It therefore belongs in dimensional qualification on a high-speed printer, even when average flow is correct.

Temperature, speed and volumetric rate interact The same profile may produce different dimensions at low and high speed. At greater volumetric flow, nozzle pressure rises, extrusion may lag or fall short, and material may accumulate at deceleration zones. Excess temperature softens detail; insufficient temperature destabilizes flow and weakens bonding. Slowing down often improves repeatability but is not a universal accuracy switch because cooling, minimum layer time and acceleration alter the thermal state. Print the coupon with the same speed, acceleration and volumetric limits as the real component. If a shop uses quality, standard and fast modes, each needs its own capability data. A slow coupon cannot establish tolerance for a rapid production run, and a manufacturer's maximum-speed claim does not establish dimensional performance at that speed.

Material shrinkage and environmental stability Cooling polymer shrinks, but not necessarily uniformly or by one universal percentage. Crystallization tendency, chamber temperature, wall thickness, infill and cooling all influence the outcome. PLA can appear stable on a small object while ABS, ASA, PC or nylon develops warping and directional contraction on a broad plate. Fibre-filled filament often reduces shrinkage and improves dimensional stability, but it is abrasive and requires a new nozzle/profile qualification. Compensation should therefore belong to a specific brand, colour, moisture condition, machine and geometric range—not to a generic “ABS scale.” Draft and room-temperature variation on an open printer can widen the distribution. Production tolerance requires a controlled environment or, at minimum, documented seasonal requalification.

X, Y and Z are not one process XY contours are shaped by nozzle trajectory, road width, corner dynamics and lateral polymer behaviour. Z dimensions additionally depend on layer height, first layer, motion mechanics, shrinkage and top-surface closure. Prusa publishes different typical internal tolerance figures for XY and Z on its own properly assembled printers, but this is manufacturer-specific evidence, not a guarantee for all FDM systems. Measure columns of several heights and exclude the first-layer flare. A Z error can be a constant initial-layer offset, an accumulating scale error or a crowned top surface; each calls for a different response. Applying one percentage scale to all three axes is rarely defensible. Record axis, feature type and measurement location separately.

Three hole orientations, three behaviours A vertical-axis hole is formed as a closed XY contour on every layer; polygonization, path placement, flow and corner dynamics determine its result. The lower surface of a horizontal hole is an overhang and can flatten or sag. An oblique bore combines both mechanisms. A universal rule that holes are always 0.2 mm small is therefore unsafe. Check source-mesh resolution as well: a coarse STL can polygonize a circle before slicing, while 3MF or modern CAD-linked workflows may preserve better geometry. Threaded, bearing or sealing features may be more reliable when drilled, reamed, tapped or fitted with an insert. Qualify the printed pilot diameter and machining allowance for the selected tool instead of assuming the printer must finish every critical surface.

Thin walls, gaps and small bosses Nominal wall thickness interacts with nozzle and selected line width to determine how many toolpaths the slicer can place. Older fixed-width engines could leave gaps or omit awkward walls; adaptive-width generation is more flexible but does not make every geometry infinitely manufacturable. A 0.5 mm wall with a 0.4 mm nozzle may become one variable-width road, while a solid 0.8 mm wall uses two roads and cools differently. Gap tests should establish when a slot remains open, when it fuses, and whether finishing tools can reach it. Small pins also suffer heat accumulation, so minimum-layer-time rules and printing several copies simultaneously can change diameter. Validate the feature in its real thermal context.

First layer and elephant foot A first layer compressed too closely against the bed widens the outline, closes lower gaps and distorts bore entries. A hot, soft lower zone can enlarge the flare. Use elephant-foot compensation only after Z offset, bed plane, flow and temperature are stable. Measure the lower one or two millimetres separately from the upper wall field. If only the base is wrong, do not scale the entire model. A small CAD chamfer can keep the flare away from a mating face but does not replace a correct first layer. For a critical seat, consider another orientation or machining. Repeat the coupon at several bed locations because mesh or sheet curvature may produce a spatial dimensional map.

Orientation and seam affect the result Rotating the same part by 90 degrees, standing it on edge or changing a hole axis can alter the measurement. Orientation changes path direction, overhangs, cooling, support marks and layer anisotropy. A Z seam may contain surplus or missing material, so do not measure an outside dimension only across it. Place and mark the seam away from a critical interface. If production permits several orientations, treat each as a separate process variant. Best dimensional accuracy may conflict with strength or surface quality, requiring an engineering priority. The related orientation guide supports that functional decision; this tolerance method quantifies the consequences of the chosen build direction.

Fit coupons measure function directly A paired pin–hole or rail–groove coupon shows which designed clearance produces a free, sliding, hand-snap or tight connection. Create a stepped series, perhaps in 0.10 mm nominal increments, while letting initial measurements determine the final range. The coupon must reproduce actual wall thickness, orientation, contact length and material. A 5 mm pin does not predict a 100 mm sliding rail. Evaluate cooled parts with a consistent deburring rule and comparable force. If lubricant, screws, compliant clips or inserts are part of the product, include them. The outcome is not “the printer's clearance”; it is design data for one defined function and process state.

Total versus per-side clearance Designers often specify 0.2 mm without saying whether it is per side or the total diameter difference. A 10.0 mm pin in a 10.2 mm hole has 0.2 mm diametral clearance and theoretically 0.1 mm on each side. A rail and groove divide their total width difference similarly. Documentation should state nominal pin, nominal hole and whether clearance is diametral or per side. Actual fit suffers simultaneously from an oversize pin and undersize hole, so consider both measured distributions. If the mating components come from different machines or materials, uncertainties combine. Online rules such as “0.2 mm always works” ignore this tolerance chain.

Compensation in the slicer or CAD? Global XY scaling is appropriate only after a proportional error is demonstrated over multiple lengths and both directions. Contour or hole compensation can target internal and external geometry, but names and sign conventions vary among slicers. Functional clearance is design intent and should remain visible in CAD. A stable manufacturing offset can live in a qualified machine/material profile so other models benefit. Do not correct twice: oversizing a hole in CAD and expanding it again in the slicer produces a loose fit. Record every original measurement, correction, version and regression result. A slicer update or new path engine can change line placement, so old compensation is not automatically transferable.

Small-shop statistics: centre and spread Three repeats are a minimum signal, but a production decision requires more. The average shows central bias; range and standard deviation describe variability. A process centred in tolerance with large spread can still generate many rejects. A tight distribution with a shifted mean may respond well to correction. Do not treat several objects printed side by side in one job as completely independent: they share material, thermal conditions and startup state. Repeat across separate starts and preferably different days. Record build-plate position because the bed and chamber can have spatial temperature or geometry effects. Even a simple control chart can reveal slow drift before customer parts fail.

Acceptance criteria and decision risk Not every dimension is critical. Identify functional characteristics on the drawing: bearing seat, screw hole, sealing face, datum edge and overall envelope. Give each a measurable limit and instrument. A decorative housing curve may accept more deviation than a sensor pocket. If the required tolerance is narrower than the qualified process capability, there are three honest options: redesign with more clearance, add machining, or choose another manufacturing technology. Sorting rejects after printing is not process capability. For safety, health or regulated components, a desktop coupon cannot replace a qualified procedure and suitable metrology. Tolerance is a functional and business-risk decision, not simply a slicer setting.

Production record and requalification Retain CAD and 3MF files, slicer and firmware versions, G-code, printer, nozzle, sheet, filament maker, colour, lot, drying record, room temperature, orientation and all measurements. Version the profile. After nozzle, hotend, extruder or sheet replacement, firmware or slicer updates, belt adjustment, major service or material-lot changes, run a short regression coupon. Full requalification is justified when that check moves, a new material or geometry range is introduced, or requirements tighten. This catches nozzle wear, mechanical loosening and sensor drift before a customer part. Documentation is useful only when a bad result can be traced to a change and the previous good state can be reproduced.

Practical validation sequence First restore mechanical and thermal baseline, clean the nozzle, confirm filament condition and qualify the first layer. Second validate temperature, extrusion multiplier and pressure compensation using a controlled profile method. Third print the multi-feature coupon at least three times, then a stepped fit coupon in the real material, orientation and layer height. Fourth classify constant-offset, scale, spatial, geometry-specific and random errors. Fifth correct only the component supported by evidence and repeat with the unchanged measurement method. Finally qualify a real component containing the critical features. This takes longer than measuring one cube, but prevents false corrections, scrap and expensive reprints.

Editorial conclusion FDM is not simply inaccurate; it is strongly process-dependent. Reliable fits emerge when the designer rejects universal internet clearance values, separates accuracy from repeatability, measures representative geometry and binds the evidence to a machine–material–profile combination. NIST's multi-feature method, official modelling and flow guidance, and the workshop's repeated measurements together provide a defensible foundation. The best result is not necessarily the smallest single deviation. A stable, known distribution is more valuable because tolerances and clearances can be designed around it. When the process cannot meet the required interval, the solution is not a more aggressive slicer number but redesign, secondary machining or a different process.

Match the measuring tool to the tolerance Instrument resolution, repeatability and contact geometry should be materially better than the tolerance being judged. Calipers suit general outside dimensions and fast checks but their inside jaws are weak evidence for small bores. Gauge pins or go/no-go gauges test function more directly. A micrometer can improve outside measurement yet contact different layer peaks than the real assembly. Flatness needs a reference and indicator; angular error needs a square and measurements at several heights. Soft or thin walls can deform under force. Record contact location, direction and support, not merely the instrument name. A shop cannot promise a controlled limit that it cannot measure credibly.

The printed part belongs to an assembly chain Screws, bushes, bearings, adhesive, paint, coatings, heat-set inserts and deburring all consume or create space. Installing an insert moves hot polymer; paint can add tenths across two mating faces; tightening a screw can ovalize a compliant housing. Qualification should represent the complete assembly. Long guides also depend on straightness and twist, not only local clearance. Mark which surfaces remain as printed, which receive finishing and at what stage the tolerance applies. This prevents manufacturing from chasing a CAD dimension that the next operation intentionally changes.

Multiple printers and full build plates Two printers of the same model are not automatically one process. Belts, nozzles, sensors, hotends, firmware and service history can shift their centres and spreads. Run the same reference coupon on every machine and bind results to machine identity before deciding whether one common profile is valid. On a full plate, centre and edge parts may see different first layers, cooling and airflow; rotate sample positions. Batch qualification must use the same quantity and layout as production because minimum layer time and interactions differ from a single part.

An AI-generated mesh cannot enter tolerance-controlled production directly

The InstructMesh study's 120 single-image TRELLIS reconstructions show why dimensional control begins at the digital input. Of those models, 78.3 per cent contained more than one fabrication-relevant error, including missing openings, hollowing mistakes and wall-thickness problems. A closed and attractive render is not dimensional data, and a successful local mesh repair does not establish that a critical feature has the right nominal size or tolerance.

After an AI-assisted local edit, save a new version, record the overall scale and selected region, then perform the same digital inspection required after manual mesh editing. Measure bounding dimensions, holes, walls and mating surfaces; check self-intersections and open edges; then confirm in the slicer that the feature survives as a distinct toolpath.

On InstructMesh's 64³ voxel grid, one voxel spans about 1.5 per cent of the shape extent. This is a proportional bound: on a 100 mm model it can imply a feature scale around 1.5 mm even before additional decoding changes. That is not an official tolerance, but it explains why a fine fit should not depend on this kind of local generative correction. Rebuild critical interfaces in parametric CAD or against measurable reference geometry.

The acceptance rule is that AI repair remains a model-preparation operation. Manufacturing tolerance still comes from capability of the specific machine–material–profile process and the functional requirement. If the edited region seals liquid, carries load, mates with another part or performs a safety function, use a representative coupon and repeated measurements.

Separate digital and physical evidence

Digital evidence includes a valid mesh, documented scale, measured nominal geometry and a feature that survives slicing. Physical evidence is the finished part measured at several positions and across repeated copies, plus a functional fit or load test. Neither replaces the other: a valid mesh can print inaccurately with a poor process, while one successful build cannot verify every hidden area of generated geometry.