Replacing A Part

Establish what is authoritative, before quoting any number

Write down, in one place: which physical part is on the robot right now, which files claim to describe it, and which of those are archived or superseded. Then apply the ordering.

• A physical measurement beats a file. Always. A file can predate the part.
• Files agreeing with each other is not evidence if they share an ancestor.
• An archived file is not a reference. It is a record of what was once tried.

List the interfaces before deciding what to measure

A replacement part is defined by what it has to mate with, not by what it looks like. Every interface becomes a fixed constraint; everything else is design freedom. List them first, then the measurement set follows automatically.

• Every fastener: count, spacing, diameter, thread or clearance
• Every mating surface: which face touches what, and in which direction load passes
• Every clearance: cables, moving parts, the workspace it must not intrude on
• Every datum: what you will measure from, stated once and used throughout

Measure spans between named points, not "dimensions"

The most reliable readings are outer-surface to outer-surface, because calipers can reach them. Name each point, say which part it is on and which side, and give its span partner. Where two points are not level, measure the component along the axis that matters — square, not diagonal.

• Label every number measured, derived, or chosen — and never let a chosen number get described as measured later
• Take each critical span twice, and say if the two readings differ
• A photo with the caliper in place is worth more than the number alone

Cross-validate every critical number by a second, independent path

One number, two derivations that do not share a source. If they agree, proceed. If they disagree — stop. Do not average them. Do not pick the one that fits the plan. Go back and find out why.

• Independent means different origin: a caliper reading vs a file, a file vs a photo with a scale, an assembled measurement vs a sum of parts
• Two files are not independent if one was generated from the other
• When physical and file disagree, the physical wins — but record the disagreement rather than silently overwriting

Write the constraint set down, in three groups

Define the acceptance criteria before building, not after

For each criterion: which check, which threshold, run against which file. Include the ones an automated tool cannot catch — those need a named manual step, or they will be skipped.

# the gripper's set, as an example of the shape
automated  check_mesh.py    connected shells = 1, watertight, manifold
automated  check_ribs.py    run on the TOOTHLESS body, not the final part
manual     texture bands    vertex count per 5 mm z-band, body vs toothed
manual     hairline gaps    slice at y=0, list every closed loop under 6×6 mm

The last two exist because both defects passed check_mesh cleanly. A watertight mesh can still have a texture band silently clipped away and a hairline gap that the eye finds and the script does not.

Record where every constant came from, in the source

A number in a CAD script with no provenance will eventually be "corrected" by someone reading a file. Put the origin next to the value, and where a correct value looks wrong, say so explicitly.

// 孔距 = 10.000mm — operator's caliper reading on the real hardware.
// ★★ Do NOT "fix" this to 12.0 from the STLs ★★  Someone did on 2026-08-21.
// Every upstream STL says 12.000 — they all descend from one upload that
// does not match the machine. Ruled on 2026-08-19. The measurement wins.
HOLE_SPAN = 10.000;

python policy_safety.py --provenance does the same job on the software side — one command that prints where each threshold came from. Any part with more than a handful of constants deserves the equivalent.