Any assembly is held in space by whatever surface the builder chose to trust. That choice is the datum, and it gets made whether or not anyone writes it down. When it is not written down it gets made at the bench, usually by whoever is holding the part, and it changes from part to part without anyone noticing.
The two candidates in most fabricated work are a machined face and a formed face. One is cut, the other is shaped. They behave differently under load, they hold tolerance differently, and they cost differently, and the difference shows up in the finished assembly rather than in the individual part.
A machined face is flat, and flatness is what a locator wants. It usually carries a known relationship to a hole or a bore, because both were produced in the same setup or in setups controlled against each other. That makes it a reliable origin for measurement and a reliable seat for clamping.
The cost is that machining has to happen somewhere in the route, which means the part is set up again after forming or welding. It also means the machined face has to be protected afterwards: a burr rolled over an edge, or a spatter ball left on the surface, destroys the very property that made it useful.
A formed face is usually the largest surface on the part and often the one the designer had in mind. It is available immediately, needs no extra operation, and spreads clamping force over a wide area, which is gentle on thin material. For a bracket or a panel it is often the only sensible choice.
What it does not give you is flatness in the machined sense. Forming leaves curvature, springback and thickness variation, and a face that looks flat on the bench may rock on a straightedge. Used as a datum it introduces a small variation into every part, which is acceptable when the assembly has clearance to absorb it and unacceptable when it does not.
Every forming operation moves material. A bend stretches the outside of the corner and compresses the inside, and the flat either side of the bend is never quite as flat as it started. A pressed panel is worse, because the draw pulls material from the flange and the flat areas carry residual stress that releases when the part is trimmed or welded.
The practical consequence is that two nominally identical formed parts can sit differently on the same three points. That is not a defect, it is the nature of the process. It becomes a defect only when the drawing assumes a formed face is a reference surface without saying how flat it has to be.
Most real assemblies use both. A machined pad on a billet part sets the position and a formed panel hangs from it. The order matters: locate from the machined feature, let the formed feature follow, and make the fastening between them tolerant enough to absorb the difference.
Reverse that and the assembly inherits the variation of the formed face at the point where the machined feature is supposed to be exact. It still bolts together, which is what lets the mistake survive. It shows up later as a hole that has to be opened out, or a joint pulled into place with a clamp before welding.
The commonest version of this problem is a thin formed panel fixed to a heavier machined frame. The frame is the accurate part and the panel is the flexible one, so the frame should set the position and the panel should be allowed to conform. Fastening the panel at every point with no allowance turns it into a structural member and transfers its variation straight back into the frame.
The fix is a hierarchy. Fix the panel hard at one point or along one edge, and let the remaining fastenings have enough clearance to locate without forcing. Where the panel has to be welded rather than bolted, the sequence does the same job: tack from the fixed point outward and let the free edge close as the panel settles. Brackets built on that principle sit in the other parts range.
Two checks cover most of it. Put the part on a flat plate and see whether it rocks, and where. Then measure from the candidate datum face to the feature that matters, twice, turning the part over between measurements, to see whether the number follows the part or the setup.
The rocking test is the more informative of the two, because the position of the rock says which corner is out and therefore which part of the process is responsible. A part that rocks on a diagonal is usually twisted, one that rocks along an edge is usually bowed, and the two have different causes and different fixes.
Plenty of drawings arrive with no datum marked at all. In that case the fabricator has to choose, and the choice should be recorded on the drawing or the traveller so the next part is made the same way. A datum that lives only in one person's head is not a datum, it is a habit.
The safe default is the most stable feature the part has: a machined pad if there is one, otherwise the largest flat, otherwise a pair of holes produced together. Everything else is then measured from it, and variation is attributed to the feature rather than to the choice of origin. That is what makes a dimensional complaint answerable.
The datum question is cheap to settle at quoting and expensive to settle after the first article has been measured.
None of this is exotic. It is the difference between an assembly that comes together on the first attempt and one that needs a file and a clamp. The titanium fabricated components range shows parts built from a machined pad outward, cut and formed sections show the other route, and the stainless welding pages cover how much a joint is expected to move once the datum is set. Where a part is quoted from a sample, the customization process starts by gauging exactly these faces, and assemblies with mixed datums are covered under aluminium welding as well.
