Let one deliberate human correction teach the system — but only for narrowly trusted error shapes

Context

Bills of materials arrive as spreadsheets typed by people, and the system categorises each line from its description. Misspellings — a missing letter, two letters swapped — land lines in the wrong category or none. Hand-adding each known misspelling as a rule fixes yesterday's typos, never tomorrow's.

Decision

During the import review a person already does, the system suggests a spelling correction next to the category dropdown — never instead of it. It trusts only two typo shapes against a short, curated list of real keywords: a single inserted, deleted or substituted character, or a genuine swap of two adjacent letters. When a person confirms a suggestion, that word is stored and applied automatically on every future import. One confirmation is enough, because this is a direct yes-or-no answer from someone looking at the line, not an inference that needs repeated evidence.

Why

A looser rule looked simpler: accept anything within an edit distance of two, which a transposition costs. Checked before shipping, it matched PALTE to VALVE exactly as closely as to PLATE — two unrelated substitutions scoring the same as one swap — which would have made the right suggestion ambiguous. Narrow shapes keep suggestions rare and right, which is what makes a single confirmation safe to learn from.

Trade-off

Typos outside the two shapes — two unrelated wrong letters, a missing syllable — get no suggestion and still need a person to pick the category by hand. The system learns less per import, in exchange for almost never proposing a wrong word.

trusted typo shapes

edit distance exactly 1, or a single adjacent transposition

confirmations needed to learn a word

1

false positive caught before shipping

PALTE ≈ VALVE at the same plain edit distance as PALTE ≈ PLATE