Showcase

Case studies of an AI agent using HiNC through its web API on machining data anyone can fetch from a public source. Each case follows the agent as it builds a project from what was published, runs it as a simulation and checks it against pass criteria written down beforehand, then turns the result into something a machining engineer can use — the stock the program really needs, where program zero lies, cycle time, peak cutting force. Everything here is simulated: no case was cut on a real machine.

A dilemma on these pages is any point where the published data, the engine or the documentation left the agent a doubt or a choice: a missing value, a contradiction between sources, a result that looked wrong, a tool that did not behave as documented. The project the agent built is its working material and is not published; the record of how the work was managed, every dilemma met and how it was resolved, is.

Cases

Ordered from the simplest set-up to the most demanding: three-axis parts first, simultaneous five-axis last.

Three-Axis Milling Benchmark, Parts 1–3: From Published CAM Data to an Accepted Simulation — the first three parts of the benchmark whose parts 4–10 follow, each given as a STEP model and a CAM program: a program zero the data set never states, a stated block 1 mm too short for its own programs, a depth peak that doubles at the finer resolution with nothing wrong, and a build procedure proved by four agents that saw nothing else
RPTU Benchmark Parts 4–10: Seven 3-Axis Jobs Set Up and Checked by an AI Agent — seven parts from one public benchmark: a program zero that moves between parts, a stated stock that fits none of the models, a tool 5 mm shorter than the obvious choice, material the program leaves at an island's bay tips, and a check that failed for the wrong reason
Ball-End Finishing in Five Directions: A Force Swing That Ranks Like the Measured Roughness, and a Form Error Tool Deflection Does Not Reach — convex and concave R45 faces in 42CrMo4 from a paper that publishes only charts: the agent rebuilt both parts and wrote the ten raster programs; the swing of the cutting force along a pass ranked the directions as the measured roughness did, the force came within 3–27 % of a later measurement, the modelled tool's deflection came out a thousandth of the measured form error, and a removal-rate defect in HiNC surfaced on sloped passes
End Mills Run to Breakage in Hardened 42CrMo4: Seven Set-Ups Ranked against Measured Tool Life — Ø10 mm end mills run until they broke, published with no program and no tool geometry: the agent wrote the contour programs, built both cutters from their maker's catalogue and ranked HiNC's per-cycle wear against the tools' lives; the radial depth ranks as the data do per cycle but not per volume removed, the long holder is invisible to the model, and a review of the first runs caught an entry move that had faked every load peak
Connecting-Rod Forging Die: A Forging Rebuilt from One Drawing, a Die Sized on a Photograph, and an R1.5 Ball in an R1.5 Corner — the closed-die forging die of an aluminium connecting rod from a paper that publishes the forging's drawing and a photo of the real dies but no die drawing: a first forging overturned by three reviewer agents and a second within 0.4 % of the paper's volume, the flash land, gutter and block sized on the photo, three tools programmed on an earlier case's height-field CAM that HiNC cut to +0.04 mm of the design and timed to 0.04 %, an R1.5 ball bending ten times its limit where it finishes R1.5 corners with its whole quarter arc, two feed revisions that HiNC's step data steered and that brought its largest bend in finishing from 112.4 to 39.6 µm but not under the limit, because what is left is the entry into the corner, not the feed
Heatsink Porthole Die: An Extrusion Die Rebuilt from a Paper's Drawings, and Where Milling Gives Way to EDM — a two-plate extrusion die for a twelve-fin aluminium heatsink, drawn but never modelled in its paper: the agent read both plates off the figures and checked them against the paper's porthole areas, wrote four set-ups with its own CAM, HiNC caught its rapid moves grazing the roughing walls, and a Ø3 long-neck cutter's seven hours, its bending and the bearing zones it cannot reach drew the line between milling and EDM
OpenCellular Antenna Cover: Two Injection-Mould Inserts Derived from the Part and Milled with the Agent's Own CAM — a polycarbonate cover from an open-hardware base station, published with a drawing that states the mould's requirements but no mould: the agent derived both P20 inserts (the vent holes' draft putting 582 core pins in the fixed half, the deep bosses left to EDM), wrote every toolpath with its own CAM down to a Ø1 mm cutter in the rib slots, and played both whole inserts through acceptance within 0.4 % of its planned time; HiNC caught three mistakes in the agent's own code, read the small tools' loads too high where they meet a corner head-on, and showed an engine defect that ends a run “Finished” with nothing cut
Jetson AGX Thor Enclosure: A Thin-Walled Aluminium Housing in Eight Set-ups, with the Agent's Own CAM — four parts of an open-hardware computer enclosure published as models and drawings with no process: the agent wrote its own 2.5-axis CAM and two set-ups per part, a T-slot cutter reached a groove under the wall top while the wall was still backed, HiNC caught a drill moving before its length offset, a spindle asked for 3.3 times its rating and a Ø1.5 grille cutter bending 0.14 mm, and the finished parts compared with the models showed four areas no pass had reached; the accepted programs came within 1 % of the estimated 6.4 hours
SpaceTeamSat1 CubeSat Structure: Six Al 7075 Plates, Twenty-Four Clampings, and Whether the Small Cutters Hold Up — the six Al 7075 plates of a university team's 1U CubeSat, published as models and drawings with no process: the agent counted 24 clampings from the holes' axes, wrote its own 2.5-axis CAM and sized every holder's stick-out from a height map of the part, which a cutter held deliberately short confirmed at the place it named; the original programs overloaded the spindle and bent the D6 and D2 194 and 360 µm, and a feed scaled to the stock the cutter meets, walls finished in two passes and long cutters planned on their own brought the finishing within 25 µm on walls and 50 µm in corners but for one step; the batch takes 6.3 hours, and HiNC showed edge set-ups placed 120° about a diagonal that read no force
THWS Keychain: A Siemens 840D sl Program with No Model, No Stock and No Fixture — an engraving and drilling program in Siemens syntax on a five-axis machine that only positions its rotary axes: the plate read out of a drilling cycle, an engraving depth that had to come from the set-up, and 45 % of the simulated cycle found in feed moves that cut nothing
Mars Rover Wheel: A 0.75 mm Skin between Chevron Grousers, Milled on Four Axes — Curiosity's wheel rebuilt from published numbers and a photograph, one tread pitch milled on an A table: the skin survives only with the tool kept on a radius and a grid finer than the skin, the roughing's power peaks sat where the zig-zag pocket became a full slot and a revision slowing only those passed, and the cutter turned out to pull the skin outwards rather than press it
A 17-Inch Forged Aluminium Wheel: Designed from Three Public Numbers, Programmed by the Agent, Checked by HiNC Before Any Cut — a one-piece forged 6061 wheel the agent designed itself from three fitment numbers, with both set-ups written as 3+2 and five-axis toolpaths: one draft chosen so a cylindrical cutter finishes every window wall exactly, a valve hole drilled from the tire side with the head at 72.5°, HiNC's per-step power pointing at the full-width cuts of a roughing that looked clean, and HiNC's feed optimization cutting the wheel's machining time by 45 % within the spindle's rating, mostly at the feed ceilings the agent had set
NIST Five-Axis Test Artifacts: A Cone Frustum and a Pyramid from a Paper's Drawing — two simultaneous five-axis parts from a paper that gives only drawings and cutting conditions: the agent wrote the toolpath itself, a machine cradle too narrow for the head, a home position inside the part, and a written five-axis program whose cutting-depth peaks agree with the toolpath's to 0.00001 mm
BCAM Blade: Two Five-Axis Flank Finishes Told Apart Before the Cut — a thin steel blade finished with the side of a conical cutter by a commercial CAM program and by its authors' own algorithm, from six published Heidenhain programs: the agent placed the blade with the semi-finishing program alone, worked out what “a radius of 3 mm” measures, read a tool 0 and a datum shift as the NX program's cutter on a second preset, and HiNC's whole chain read the paper's 45 measured points to 0.002 mm (NX) and 0.005 mm (G1) on average — the +0.14 mm the NX program leaves already in its tool path
NASA HECC Centrifugal Impeller: Simultaneous Five-Axis Machining from Published Blade Coordinates — a compressor impeller published only as blade sections with no unit: the agent wrote the whole five-axis tool path itself, HiNC's cutting-load model read the first path's tool stress far above its stress line, the owner spotted on the canvas lifts that cost a tenth of the simulated time, and HiNC's messages and a review caught retracts that grazed the part and entries that fed too fast

NC optimization cases

The same product cases taken one step further: the agent runs HiNC's NC optimization on the programs a case accepted — on some pages also on an earlier or a hand-revised version of them — as Workflow: NC Optimization describes, and replays the optimized programs on the same machine. Each page sets cycle time, spindle load, cutting force, tool deflection and the machined shape before the optimization against the same after it, measured on HiNC 3.2.45 and judged on criteria written down before the first optimization play, and says what the agent protected from the optimizer, such as finishing passes kept at their programmed feed or a floor that lets no step run slower than programmed. The pages follow the outline below and the order of the cases above.

Connecting-Rod Forging Die, Optimized: HiNC's Feed Optimization Takes 24.5 % off the Hand-Revised Programs, and Even the Lowest Feed Leaves the R1.5 Corner over Its Limit — the die's accepted programs and their two hand feed revisions through HiNC's feed optimization in three variants: the roughing 36.6 % shorter within the spindle-power and tool-stress targets in every variant, and the hand-revised programs, allowed only to speed up, 24.5 % shorter with their peaks unchanged; deflection targets that write 1,422 of the 1,423 over-limit steps at the corner at the 100 mm/min minimum feed bring the R1.5 ball's largest bend in steep finishing from 112.4 to 23.8 µm, still over its 12.5 µm limit, while slowing the 2 mm before and after each of those steps nearly doubles the machining time, so what is left to change is the entry into the corner or the size of the ball, not the feed
Heatsink Porthole Die: HiNC's Feed Optimization Takes 13 % off Both Plates, and No Feed Keeps the Ø3 Fin-Relief Cutter within 30 µm — the die's eight accepted programs through HiNC's feed optimization, each replayed on the same machine: both plates from 608.6 to 529.1 simulated minutes (−13.1 %), the porthole roughing about a third shorter at its feed-per-tooth ceiling with the spindle under a quarter of its short-term rating, and the finishing kept at its programmed chip while force targets trim the wall finishing's largest deflection from 122 to 78 µm; on the Ø3 long-neck cutter in the fin slots even the 30 mm/min minimum feed leaves 72–82 µm against the 30 µm allowed, at about 47 hours for the fin relief, the numbers behind sending it to EDM
Antenna-Cover Injection Mould, Optimized: HiNC's Feed Optimization Takes the Moving Insert from 12.0 to 7.9 Hours, and No Feed Holds the Ø1 mm Tool within 10 µm — the moving insert's two micro-tool programs, a Ø2 mm rest program slowed by the agent's hand corner rules and a Ø1 mm rib-slot program at the catalogue's halved slotting feed, through HiNC's feed optimization and replayed on the whole insert: 719.97 to 474.00 simulated minutes (−34.2 %), 148 of the 246 minutes saved being feed through air, the ramps and plunges the CAM had sent at the plunge feed through levels already cut, with the tools' largest stress ratio falling and the finished stock the same region by region; on a trimmed patch the optimizer found the corners without the hand rules, 20.72 against 20.61 min with them, and a target force set for the Ø1 mm tool's 10 µm limit made its trimmed program 3.8 times slower and still left a 99th percentile of 11.9 µm, so that limit is out of reach of the feed
Jetson AGX Thor Enclosure, Optimized: HiNC's Feed Optimization Takes 28 to 52 % off the Long-Cutter Set-ups, Holds the Cover's Finishing Bend within 25 µm, and Takes 37 % off a Grille That Looked at Its Limit — the bottom shell's and the top cover's first set-ups, two Ø12 mm cutters on the spindle's real power curve, and the back panel's Ø1.5 mm grille through HiNC's feed optimization in three variants: 197.9 to 121.7 simulated minutes for the three programs (−38.5 %), the roughing set by the chip ceiling the agent allowed rather than by the spindle; finishing held at its programmed chip in one variant, and in the other steered by a target force and a ceiling per line drawn from its own bend, which brings the cover's largest finishing bend from 36.5 to 24.2 µm and leaves the shell one wall step and 12 floor steps just over the 25 µm limit, at most 26.7 µm; and the grille cutter, whose stress looked at its target on the case's 0.25 mm grid, reaching its programmed chip on 99 % of its steps at 0.125 mm
SpaceTeamSat1 CubeSat Structure, Optimized: HiNC's Feed Optimization Brings the Original Program to the Spindle Target and the Deflection Limits for 3.8 % More Time, and Takes 9.9 % off the Hand-Tuned One — plate X-'s first clamping through HiNC's feed optimization at the case's 0.125 mm grid, both the program as the agent's CAM first wrote it and the hand-tuned one the case accepted: the original program's D16 roughing brought from 1.6 times the spindle's short-term rating to the 0.667 target and its D6, D4 and D2 finishing within the case's deflection limits (D2 359.9 to 49.9 µm) for 3.8 % more time, 19.9 % less than the hand-tuned program, though at chips down to 0.001 mm, so the limits hold in HiNC's model only; and the hand-tuned program 9.9 % shorter within the same limits, two thirds of it from roughing ramps raised to the full chip
A 17-Inch Forged Aluminium Wheel, Optimized: HiNC's Feed Optimization Takes 45 % off Both Set-ups, with the Finishing Passes Left at Their Programmed Feed — the wheel's six accepted five-axis programs through HiNC's feed optimization in three versions, each replayed on the whole wheel: optimizing every pass takes both set-ups from 17 min 25 s to 9 min 36 s (−44.9 %) within the spindle's rating but bends the finishing tool from 94.4 to 140.2 µm; with the final flank loops and the pocket floors left at their programmed feed and no step slower than programmed, 44.5 % remains and the finishing replays step for step as programmed, the saving resting mostly on the feed-per-tooth ceilings the agent chose

Error-term cases

Cases with constant geometric errors of the machine built in, most on a case accepted in this Showcase and some on a paper's test part built for the purpose. The agent gives axes or the spindle of the machine tool error terms, each a static branch of its own just before the axis's branch — a squareness between two linear axes, the location or the tilt of a rotary axis, a spindle out of tram, at the size a paper measured on a real machine where one printed it, at a typical size otherwise — as Error-Term Branches describes. It then plays the case's programs with the error terms off (the nominal cut, the ideal geometry of every comparison), on with the controller unaware of the errors, and off again with a Diff (the noise floor); where RTCP, a tilted plane or CL playback lets the controller compensate, also on with the errors known to it, as a controller with calibrated kinematics would be. Each page predicts the tool-tip offset and the part's deviations before the first play, checks them step by step and on the cut part, and says what the errors mean for the part; what the model leaves out is listed under Machine Geometric Error. Measured on a HiNC 3.2.46 development build with error-term branches.

A 17-Inch Forged Aluminium Wheel on a C Table 20 µm Off Centre: an Uncalibrated Controller Leaves Each Spoke About 40 µm Thinner and the Bolt Circle 36 µm Off, a Calibrated One Cuts the Nominal Wheel — the wheel's four accepted G43.4 programs of the first set-up, replayed on its table-C / swivel-head-B machine with one marked error-term branch upstream of the C axis: the C table's centre of rotation 20 µm off in +X (EX0C, a typical value), with the G54 a touch-off at C 0 would find typed in. With the controller unaware of the error, all 157,412 cutting steps put the tip on the predicted offset to within 1.6 × 10⁻¹³ mm: the roughing at C 0 is unaffected, the wall facing −X is 39.7 µm too deep, each spoke is 38–41 µm thinner as if the cutter radius were 20 µm larger, and the bolt holes, drilled at C ≡ 234°, sit 35.6 µm off centre as one set. With the mark cleared, as under calibrated kinematics, the wheel comes out as the nominal one, and only the machine's X stands 20 µm further in +X at every cutting step
BCAM Blade on a C Table 0.020 mm Off Centre: One Error Cuts the G1 Finish Thinner and Leaves the NX Finish Thicker, and a Calibrated Controller Cuts Both Back to Nominal — the blade's two accepted five-axis flank finishes, a commercial CAM program and the authors' G1 programs, replayed under Heidenhain M128 on the case's B/C table-table machine with one marked error-term branch upstream of the C axis: its centre of rotation 0.020 mm off in +X (EX0C, a typical value), with both presets typed in as a touch-off at B = C = 0 would find them. With the controller unaware, every cutting step's tip lies on the predicted offset within 2.6 × 10⁻¹³ mm, and because the two programs turn C to opposite sides, the same error leaves 0.010 to 0.027 mm of material on the NX face and cuts the G1 face 0.011 to 0.029 mm too deep (1st to 99th percentile), which by derivation would take 20 of the paper's 25 G1 points outside ±0.01 mm. With the mark cleared, as under calibrated kinematics, both cuts match the nominal ones within 5 × 10⁻¹⁰ mm
The NIST Cone Frustum on a C Axis 10 µm Off Centre: an Uncalibrated Controller Cuts Most of the Cone up to 18 µm Too Deep yet Only 3 µm Out of Round, a Calibrated One Cuts the Nominal Part — the frustum's accepted G43.4 program, replayed on its B/C table-table machine with one marked error-term branch upstream of the C axis: the C axis' centre of rotation 10 µm off in +X (EX0C), the error Moylan et al. report for the machine that cut the NIST frustum, with the G54 a touch-off at B = C = 0 would find typed in. With the controller unaware of the error, all 35,513 cutting steps put the tip on the predicted offset to within 1.14 × 10⁻¹³ mm: the side cut at C = 0 is unchanged, the cone side is cut too deep from θ ≈ ±44° round the −X side, 18.1 µm at the deepest, and a roundness check at mid-height reads only 2.98 µm of it while the diameter shrank 16.25 µm and the centre moved 10.26 µm. With the mark cleared, as under calibrated kinematics, the frustum comes out as the nominal one, and X, Y and Z stand up to 20 µm from where the uncalibrated controller with the same touch-off puts them
RPTU Benchmark Part 8 on a Machine with Three Measured Squareness Errors: At Most 1.7 µm on an ISO 10791-7-Style Test Piece, and Each Family of Features Reads One Angle — part 8 of the RPTU three-axis benchmark, the one made after ISO 10791-7, played again on its skeleton machine with the X–Y, Y–Z and Z–X squareness errors a laser tracer measured on a real vertical machining centre (24, 33 and 56 µrad), each a marked error-term branch directly upstream of its axis with its pivot at the touch-off, so the case's G54 stays. All 159,337 cutting steps put the tip on the exact three-angle formula to within 1.24 × 10⁻¹³ mm, and the cut part, measured vertex by vertex against the nominal cut, deviates by at most about 1.7 µm: the block's front and back faces slope along X by ±24.4 µm/m and its left and right faces with height by ±56.0 µm/m, the 30° square's corners change by 0.65–0.66 µm and the Ø54 boss becomes two lobes 0.733 µm out of round, so all three angles read back from the part within 1 %. The program is plain three-axis, so the errors are cut in whether marked or not; the same X–Y squareness would come to 7.82 µm on ISO's 320 mm test piece
A Roundness Test Part with the Machine's Measured Squareness Errors: Only the X–Y Squareness Makes the Circles Oval — the test part of Holub et al. (Applied Sciences 2020), a Ø300 flange, a Ø200 boss and a Ø100 bore, built from the paper's printed numbers into a base project of its own and played again with the three constant squareness errors the paper measured on its machine, as two marked error-term branches pivoted at the touch-off so G54 stays: the ballbar's X–Y squareness (22.4 µm/m) upstream of Y, and the laser tracer's squareness of Z to Y and to X (+33 and −56 µrad, one 65 µrad rotation) upstream of Z. All 5,411 paired steps put the tip on the predicted offset to within 1.24 × 10⁻¹³ mm. The X–Y squareness alone cuts each circle into an ellipse, 3.535 / 2.421 / 0.940 µm out of round, γ times the tool-centre radius; the Z squareness leaves the roundness alone and leans the axis the three cylinders share by 65 µrad, 3.25 µm over a 50 mm wall. The X–Y squareness explains 35 % / 26 % / 18 % of the roundness the paper measured before compensation, while the residual squareness after it would leave under 1 µm where the paper still measured 5.6–6.2 µm. The programs are plain three-axis, so the errors are cut in whether marked or not
Plunge against Layer Roughing of a Deep Slot on a Z Axis 0.003° Out of Square: Both Walls Lean toward +Y, 3.9 µm at 74 mm Deep, and Both Strategies Cut the Same Slot — the deep-slot study of Silva et al. (Journal of Manufacturing and Materials Processing 2024), a closed slot 12 mm wide, 140 mm long and 74 mm deep, roughed by full-depth plunging or in 20 layers of 3.7 mm and finished in five levels with a D12 end mill 85 mm out of an ER32 chuck, rebuilt by the agent with three programs of its own, the plunge roughing in 28 plunges, and played again on its three-axis skeleton machine with the Z axis 0.003° out of square with Y (15.7 µm per 300 mm, a typical value), as one marked error-term branch upstream of Z pivoted at the touch-off, so the case's G54 stays. All 73,771 cutting steps of the three runs with the error on put the tip on δ = (0, −Zp·sin θ, −Zp·(1 − cos θ)) within 1.51 × 10⁻¹³ mm, by depth alone. Both long walls lean toward +Y together, the +Y wall cut too deep and the −Y wall keeping material, 0 at the rim and 3.9 µm at 74 mm deep, with the width unchanged and no step between the finishing levels, and the plunge and layer routes come out with every wall vertex the same bit for bit, as the plan predicted before the first play. The programs are plain three-axis, so the error is cut in whether marked or not
The NISTIR 6868 Test Part Cut with the X–Y Squareness a Laser Ball Bar Measured on Its Machine: Two Rows of Holes 6.87″ out of Square, and the Slot an Ellipse 1.64 / 1.69 µm out of Round — the test part of Gilsinn and Ling's NISTIR 6868 (2002), rebuilt from the report's scanned drawing and tables into a base project of its own with the agent's four Fanuc programs, and played again with the X–Y squareness a laser ball bar measured on the production machine that cut the report's 21 parts, −6.87″, as one marked error-term branch directly upstream of Y, pivoted at the touch-off so the G54 stays. All 55,493 steps put the tip on the predicted offset, k·X in +Y, to within 2.07 × 10⁻¹³ mm; the 24 squared holes move in +Y by 0.33 to 4.66 µm and their two rows read −6.8717″ back through the report's own equations, the 150 mm square becomes a parallelogram, and the slot an ellipse 1.64 / 1.69 µm out of round, set by the cutter-centre radius rather than the wall radius the plan used, so one criterion failed as written. The simulated orthogonality lies 5.8″ from the CMM mean of −1.06″, inside its expanded uncertainty of 18.1″; the squareness explains under a tenth of the measured roundness, and the plain three-axis programs cut the same part whether the error is marked or not
Mars Rover Wheel on an A Axis Tilted 0.005° to the X Travel: the 0.75 mm Skin Tapers from 0.733 mm at the Faceplate End to 0.767 mm at the Tailstock End, and Without RTCP a Calibrated Controller Cuts the Same Skin — the Mars rover wheel's two accepted G43 H programs, which position X, Z and A without RTCP, replayed on its generic four-axis machine with one marked error-term branch directly upstream of the A axis: the line through the faceplate and tailstock centres tilted 0.005° to the X travel (EB0A, a typical value), the tailstock end 46 µm low over 531 mm, pivoted at the touch-off point on the groove floor so the case's G54 stays. All 71,562 steps put the tip on the predicted offset to within 4.07 × 10⁻¹³ mm with the machine coordinates unchanged, and the groove floor, the skin's outer face, follows x·sin α to within 0.0183 µm in forty 10 mm bins: the skin runs from 0.7326 mm at the faceplate end to 0.7674 mm at the tailstock end and stays the same round the wheel, while the walls move by tenths of a micrometre; the plan's skin criterion failed as written only because its filter took in uncut faces. A program without RTCP does not read the machine chain, so a trimmed run with the mark cleared, as under calibrated kinematics, cut the same part bit for bit; only aligning the centres, or an RTCP program on a controller that knows the tilt, would take the taper out
V-Grooves on an 11 mm Sphere with the Tool Tip 10 µm below the B Axis: Both Grooves Cut Deeper from 0 at Their Top to 7 µm at Their Ends, and with the Program in Machine Coordinates a Calibrated Controller Cannot Help — the five-axis program of Wu et al. (Mendeley Data 2023) that cuts two V grooves into an 11 mm sphere in machine coordinates, X Y Z B C with no RTCP, played again on the generic swivel-head-B / table-C machine an agent read back from it, with the B axis' centre 10 µm above the tool tip (a typical value) as one marked error-term branch between B and the spindle, and the G54 Z a touch-off at B 0 would find typed in. All 4,961 cutting steps put the tip on dz·(−sin B·cos C, −sin B·sin C, 1 − cos B) within 1.8 × 10⁻¹⁴ mm, and since the tool meets the sphere at a constant 48.6°, the groove's depth error goes by height alone: 0 at the grooves' top, −6.45 µm on the floor and −7.33 µm at the deepest point of their ends near the equator, 28 % of the program's 25 µm depth of cut, while the two pits the program's own singular blocks leave at B ≈ 0 do not move. With the mark cleared, as under calibrated kinematics, the machine coordinates stay the same bit for bit and the cut byte for byte, since a program in machine coordinates never goes through the chain: the correction has to go into the post-processor
The Cubic Machining Test of JMMP 2021 with the Eight Rotary-Axis Errors a Ball Bar Measured: Zones II and III Stand 0.43 / 0.23 µm above Zone I, and a 10 µm Tool-Length Error Shifts Them by 1.34 µm — the cubic machining test of Li et al. (J. Manuf. Mater. Process. 2021), one face in 3 × 3 zones cut by an R3 ball-end mill under TCP with the tool vertical, tilted 30°, or tilted 30° and turned 45°, built anew from the article on a generic B/C table-table machine, with the eight rotary-axis errors the authors identified with a ball bar put in as two marked error terms just before the B and the C axis, and the touch-off work offset typed in. Every TCP step's tip lies on the prediction written before the first play within 7.4 × 10⁻¹² mm; the face the ball paths sweep stands +0.433 µm higher in the four zones II and +0.234 µm in the four zones III against zone I, within 0.004 µm of the prediction, and the angular errors also tilt every face by up to 1.97 µm across a box, which zone means cannot see. A controller tool length 10 µm too long or too short moves the tilted zones by 1.340 µm against zone I, (1 − cos 30°)·eL. With the marks cleared, as under calibrated kinematics, the tip comes back within 5.85 nm, yet the tilted faces still stand 0.050 µm high, as the tool's direction keeps the errors. The Diff's box means read the 4.63 µm ridges between unswept steps instead, so six of the plan's twelve criteria failed as written, each cause measured; the authors put most of the +4.4 to +6.1 µm they measured down to the ball's edge form, which this case does not model
Jetson AGX Thor Enclosure, Bottom Shell, on an X–Y Squareness of 0.003°: Turned Over, the Two Set-ups' Faces Stand 8.85 µm Apart at the Ends of the Long Walls; Seated in Soft Jaws Machined on the Same Machine, the Mismatch Runs along X and the Seam Is Gone — the enclosure's bottom shell, its inside machined in OP10 and its outside in OP20 after the part is turned over about X, both plain three-axis Fanuc programs, replayed on the case's skeleton machine with the X and Y travels 0.003° out of square (15.7 µm per 300 mm, a typical value) as one marked error-term branch directly upstream of Y, pivoted on the spindle axis both touch-offs share so the G54 stays. All 224,167 cutting steps of OP10 and 165,184 of OP20 put the tip on δ = (X(cos θ − 1), −X·sin θ, 0) within 8.36 × 10⁻¹⁴ mm; within one set-up the faces facing ±Y only shear by ±4.42 µm, but on the part turned over the shear changes sign, so the faces OP10 and OP20 cut stand 8.85 µm apart at the ends of the long walls, a step at one end of each long wall's seam where the part is placed in the design frame, while seated in soft jaws machined on the same machine, an assumption about the fixturing the agent played as well, the seam is gone and the mismatch moves to the wall ends and the holes' ±x sides. The tightest size, a 1.2 mm pad wall between a hole cut in OP20 and a pad cut in OP10, is predicted to use at most 7.6 % of its ±0.1 mm general tolerance, and the machine would have to be 13 times as far out of square to use it up; the programs are plain three-axis, so the error is cut in whether marked or not
A CubeSat Side Panel on a Z Axis 0.2 mrad out of Square with Y: Milled Flat, Its End Faces Lean Whole; Drilled on Edge, Its Holes Lean while Their Mouths Stay — three of the five clampings of the STS1 CubeSat's side panel CS_Yminus, replayed on the case's three-axis skeleton machine with one marked error-term branch upstream of the Z axis: the Z travel 0.2 mrad out of square with Y, the factory-acceptance limit of a light milling machine, and ten times that as a display value. Milled flat, every face toward ±Y shifts and leans, the end bars' end faces by ∓2.03 µm on their upper band and ∓2.77 µm on their lower band; drilled on edge with the touch-off on the end face, the hole axes lean 2.000 µm per millimetre of depth at 2 mrad while the mouths stay within 0.13 µm, and the tool tip meets the formula within 2.54 × 10⁻¹³ mm on every step.

How a case page reads

Every case page uses the same eight headings, in this order, so two cases can be read side by side.

The case — what the agent was given: the published files and what they state, and what they leave out — the stock, where the part sits, the controller, the material, the tool beyond its diameter.
What the agent built — the set-up item by item — machine, fixture, stock, target model, program zero, controller, material, tool and holder, spindle, mission — each value marked as read from the source, derived from it, or chosen by the agent.
How the agent managed the work — what counts as a pass, written down before the run; coarse runs before fine ones; other agents rebuilding the case from the source files and the agent's written instructions and nothing else (a blind build); other agents set to find errors in its written claims (adversarial review); sharing one server with other agents; and where a person stepped in — what the owner noticed or ruled, and what the agent changed as a result.
The dilemmas — the body of the page, the small ones included. Each is told as its situation, the risk had it been missed, how it was noticed, the resolution, and the evidence that the resolution held.
Results and benefits — the measured numbers with the HiNC version they were measured on; what they tell a machining engineer, a teacher or student, and someone weighing whether the approach is worth using, including what the work cost in run time and memory.
Honest limits — what is assumed rather than measured, and what the engine does not check.
What a reader can take to their own case — the checks that carry over to other data.
Source and licence — the original source with its link and the search terms that find it again if the link moves; its licence or terms of use, named and linked; the attribution line as the licensor asks for it, including that the material is provided without warranty; what was changed from the original and what was left as delivered. This site offers no download: the reader fetches the originals from the source, and where the licence allows, the company site keeps a backup copy of the original files.

Every value on a case page says where it came from — read from the source, derived from it, chosen by the agent, or measured in a run — and measured numbers name the HiNC version they were measured on. Every picture shows the cutter in its holder at a stick-out a shop would use, because a cutter drawn hanging out of the spindle with no holder would break in practice. No case comes from a client.

The methods behind the cases

The agents build and check a case with the same public pages any caller of the web API reads: