An editable model you cannot read is not editable, it is unlocked. Changing one dimension means knowing which sketch owns it and what rebuilds downstream. That is reading, and nobody automated reading. A tree nobody can read gets rebuilt from scratch, quietly.

Section a bore off axis and it reads oval, and the fit will agree with you. Slice a cylinder at an angle and you get an ellipse. Minor axis is the diameter. Major axis is tilt. Fit the cylinder first, then section square to its axis. Wear is something you prove.

The part on your bench is a sample of one, and you are about to call it the nominal. Two castings off one pattern disagree before anyone fits them. Add wear and you measured one part's past, not the design. A second part costs a shipping label. The nominal is the job.

The same bore can arrive as a cylinder in one STEP file and a spline surface in the next. A cylinder carries a diameter and an axis. A spline carries control points that happen to be round. Hole recognition finds one of them. The other gets contoured, not drilled.

A tree that errors on the first edit is better built than one that quietly does not. The dangerous one absorbs it. A fillet solves to nothing, a sketch flips to the mirror solution, a hole lands on the face behind the one you meant. Quiet trees ship wrong parts.

The hard part of AI in CAD is not making the model, it is signing it off. Generated geometry still needs someone to say the bore is right. Checking a tree you did not build is slow work. The work did not disappear. It moved to the person with the most to lose.

A bolt pattern is one feature with two numbers, not six separate holes. Fit all six centres to one circle and the errors argue with each other. Pitch diameter and clocking come out tighter than any single hole. Model the holes from the pattern, not where they scanned.

The hardest call on a reverse engineering job belongs to whoever holds the scanner. Coverage, spray, which faces got a second pass: all settled before anyone asked what locates the part. A scan you did not plan is a guess about scope, in a file you now have to quote.

An STL triangle states its facing twice, and the two can disagree. One is the stored facet normal. The other is the vertex winding. Most readers ignore the normal and recompute from the winding. Some trust it, so a mesh flips in one tool. The file holds two opinions.

History free modelling is not faster, it moves the work to the next edit. Push a face and it updates. The file never says why, so the next change starts with calipers. The part comes back, because the mate moved. A tree costs more to build and less to own.

A long part scans bowed, and every frame in it was fine. Frame to frame alignment passes its error forward, and it piles onto the long axis. Targets break the chain: each frame references one thing, not its neighbour. The bow is in the scanning order, not the part.

The same scan gives two different bore axes depending on what you fit first. Fit the bore alone and the axis goes where the points lead. Fit it square to the seating face and the face decides. Both are honest. One is a measurement, the other is a datum scheme.

The reverse engineering job worth taking is a fixture, not an obsolete spare. A spare has to match a drawing nobody kept and a fit you can only infer. A fixture has to match the part already on your bench. Off by a hair, a fixture gets shimmed. A spare gets scrapped.

A mesh imported at the wrong scale looks completely correct. STL has no unit field. Not millimetres, not inches, just numbers. The importer picks, and shape is the one property scale does not change. Nothing in the file will catch it. One caliper dimension will.

A sketch dimensioned to a model edge is a bet on that edge surviving the next edit. Change the feature above it and the edge gets a new identity. The sketch breaks somewhere you never touched. Dimension from planes. A projected edge is convenience on credit.

Every sharp edge in your scan has a radius, and it is the scanner's, not the part's. A spot wider than the corner averages both faces into one band. Nothing samples the apex. A crisp edge and a small break scan alike. The corner is a conclusion, not a measurement.

The AI worth having in CAD grades the model, it does not draw it. Geometry is cheap to make and expensive to trust. Making it faster just makes more to verify. Grading films badly, which is why every demo generates. Ask one what it does when it is unsure.

A symmetry plane fitted to a worn part sits half the wear off centre. The scar is on one side only. The fit splits the difference between a good face and a worn one. Mirror from there and the good half inherits it. Fit symmetry to faces nothing ran against.

The hardest calls in reverse engineering get made furthest from the machine. Which face took the load, which bore locates it, which radius is only how the cutter left it. Whoever ran it knows all three. It goes to a desk, where a guess gets three decimal places.

The most useful thing in a reverse engineered model is a text note, not the geometry. Which faces were measured, which were guessed, which dimension got rounded to a drawing. The tree holds none of it. Six months on the geometry is obvious. The reasoning is gone.