Braille contractions

Measured against liblouis 3.29.0 (en-ueb-g2), the reference implementation braille displays and embossers run, over the 5,000 most common English words: 99.7% of words identical, and 99.9% once weighted by how often each word actually occurs. That comparison runs on every build, so this figure cannot drift away from what the translator does. The method, and all fourteen disagreements, are published in full.

• All 65 shortforms (ab for about, bec for because, o'c for o'clock)
• All 33 initial-letter contractions as standalone words (one, day, time, their, question)
• All 14 final-letter contractions (-ment, -tion, -ness, -ful, -ity, -ance, -ation and the rest)
• Shortest-form selection — where several spellings are valid, the one with fewest cells is chosen
• All 23 alphabetic wordsigns (but, can, do, every, from, ...)
• All 19 strong groupsigns (and, for, of, the, with, ch, sh, th, ...)
• Strong wordsigns standing alone (child, shall, this, which, out, still)
• The lower groupsigns ea, bb, cc, ff and gg, with the rule that keeps them between letters (read, rubber, bigger — but not sea or eat)
• The con- and dis- prefixes, including the spellings that are not the prefix (contact but not cone, display but not disc)
• The lower wordsigns be, enough, his, in, was and were standing alone
• Initial-letter contractions inside longer words as well as alone (b-one, ph-one, net-work, part-y), in the positions where that agrees with the reference more often than not. There are four positions, because a trailing plural s is its own case: the one contraction ends a word 96% of the time, ends it before an s 85% of the time, and sits in the middle of one 16% of the time, so it is allowed in the first two and not the third
• The be- prefix, which applies only where be is a syllable of its own — begin and become take it, best and bench do not. That distinction is about pronunciation, so it follows the reference implementation's own patterns: 98.4% agreement across all 3,676 English words beginning be
• Words spelt like a shortform, which are written so they cannot be read as one — agst drops its contraction, and ab, which has none to drop, takes a grade 1 word indicator instead
• Shortforms carrying an ending: friends, friendly, friendless, greatest, declared, receiver. The endings allowed are the ones measured safe, because shoulder is not should plus an er

Not covered:

• A shortform buried inside an unrelated word (newsletter, aboutface, greatgrandmother) — the reference contracts about 600 of those from a list of individual words rather than by rule, and this engine spells them out instead: longer, but never a contraction that reads as the wrong word
• The last words where an initial-letter contraction must not be used inside them (money, reverse) — position removes most of these but not all, and what remains is about meaning rather than spelling
• Four abbreviations the reference gives a grade 1 indicator from a word list rather than a rule (blvd, bldg, llc, llb) — over the 8,624 dictionary words whose cells open with a shortform, those four are the only ones that get it, so there is no rule here to follow
• The groupsigns ed, er and ow standing alone as whole words, which the reference writes as a single cell while spelling out ar, en, gh and ing in the same position — this engine spells all of them out, which is longer and unambiguous
• Context rules that depend on syllable boundaries or pronunciation
• Typeform indicators for bold, italic and underline

Good enough to read, learn from, and check your work against. Not a substitute for a certified transcription — if the output is going to be embossed for someone to rely on, have a qualified transcriber check it.