The S2000 digital dash preview already had a geometry gate. That told me where the pixels were. It did not tell me whether the lettering, colour or bar behaviour matched the available references.
The actual React instrument SVG, rasterised with simulated telemetry. This is a bench preview, not a photograph of a fitted dash or a browser interaction test.
I corrected the face in separate passes, then added an acceptance check across six tracks: graphics, colour, position, size, motion and font. Each needs something independent to compare against: named regions in the lit photograph, the supplied glyph plate, or the dimension lock. A newly captured image cannot be its own proof that the face is right.
The useful tests deliberately break it. One erases a tach numeral and updates the capture's hash so it still looks fresh. The font check must reject it anyway. Others restore the former LCD colour or fill the thin overrun strokes into a solid region. They fail their corresponding checks.
Motion needed a different test. Both renderers were progressively filling a fraction of the next tach cell, while the lock specified complete 200 rpm steps. The replacement compares actual rendered pixels: 3,200 and 3,399.9 rpm must look identical; 3,400 must change. The same invariant is checked near zero and in the overrun range. A still screenshot would miss it.
The full native gate and the five static SVG tracks passed. The implementation also passed 175 Python tests and 26 web tests, with the native acceptance and React render checks included in CI. That gives me a more useful reason to refresh a screenshot baseline: the reference checks passed first.
The remaining boundary matters. A still photograph cannot establish factory response timing or calibrated display brightness. The traced alphabet is a reconstruction, and the SVG capture leaves browser CSS and interaction outside its checks. Physical fit is still separate work.