A longbow tiller’s finish cards say linseed, not pitch-black. Heavy color jitter makes every tiller look tarred while the notes stay the same. What should the steward do?
Select an answer to reveal the explanation.
Short Explanation
If the recipe says honey glaze and the photo looks like burnt tar, the pair is broken. Drop the jitter that erases the dye the text names—even if you gained files.
Full Explanation
Longbow tiller finish cards that name linseed (not pitch-black) become false when heavy color jitter makes every tiller look tarred while notes stay the same. The steward should drop that color transform because the visual arm no longer matches the named finish attribute—Domain 5 alignment analysis over file-count vanity.
Keeping the heavy jitter because file count is the only accuracy metric fails because mismatched color versus named finish poisons attribute learning. Rewriting Domain 2 fusion to ignore color entirely fails as an architecture overreaction to a bad augmentation choice. Replacing the notes with Riva TTS so color no longer matters fails the modality job: speech synthesis does not restore visual–text attribute alignment on finish cards.
Exam caveat: attribute-bearing captions constrain safe color augmentations. Transforms that erase the named appearance are rejected even if they “diversify” pixels.
Operational check: compare jittered stills to finish vocabulary (linseed versus tar-like); remove or soften color transforms that cross attribute boundaries; re-validate a holdout of civic tiller cards for finish-name consistency before training. Document the allowed color jitter range beside the finish lexicon so municipal woodshop models keep reading linseed as linseed.