Research
What we measured
Six questions about braille that have numerical answers, and did not have published ones. Four are measurements this site takes of itself on every build; the other two are what those measurements turned up about the reference implementations everything here is measured against.
This is not a blog. Nothing here is published to a schedule, and there will not be another piece until there is another thing worth reporting.
What these have in common is that the number already exists. This site's Grade 2 output is compared against the reference implementation on every build; its page estimator lays out a whole novel to calibrate itself; its contraction rules were set by surveying thousands of English words against the reference rather than by reading a rulebook. Those measurements were sitting in the repository doing their job silently. The write-ups are the method and the reading.
Every figure on these pages is derived from the committed data at build time. None of them is a number typed into a sentence, so when the engine improves the articles change with it instead of quietly becoming false.
The six
How accurate are free braille translators?
Every free translator claims Grade 2. This one is measured against the reference implementation on every build, and the 14 words it still gets wrong are listed.
How many braille pages will my book be?
Transcription vendors answer three to four pages per print page and then ask for the file. Here is the arithmetic, measured by laying out a whole novel.
Which shortforms can carry an ending?
The usual teaching gives one list of endings for all 73 shortforms. Measured against the reference over 628 words, that list is wrong in both directions.
Where a contraction may sit inside a word
The rule is that the letters must keep their meaning, which spelling cannot decide. Over 11,014 words, position decides most of it — and the middle of a word is not one case but two.
Six bugs we found in the reference braille tables
liblouis is the implementation this whole site measures itself against. Building Arabic and Portuguese engines against it surfaced six rows where it departs from its own stated rule or from the published standard — each with a file, a line, and a test that fails the day it is fixed.
The Hindi table most software ships is out of date
Bharati Braille 2.1 was released in January 2026 and liblouis updated its Devanagari table a month later. The released version still carries the 2014 one, which writes 250 of the 5,000 commonest Hindi words differently.
Why publish the failures
The conformance piece lists all 14 words this site's translator gets wrong, in a table, with the reference's answer beside its own. That is an unusual thing for a tool to do about itself.
It is also the only version of the claim that means anything. “Accurate Grade 2 translation” is what every competing tool says, and it is unfalsifiable. A number with a method and a list of failures can be checked, argued with, and — if it is wrong — corrected by someone who knows better. That is worth more than the fraction of a per cent the failures cost.
If you read braille and something here is wrong, say so.
Reuse any of it
The text on these pages is CC BY 4.0 and the figures are facts. Quote them, put them in a lesson, argue with them in print. The underlying tables are also available as JSON if you would rather have the data than the prose.