About BrailleChart

Why this exists

A braille translator that gives only its output asks to be taken on trust. The published reference material is authoritative, but it is in long documents that are slow to search for a single cell.

This site is the reference and the tools in one place, with the dot numbers always visible so you can verify any output against a published table.

Who publishes it

BrailleChart is published by SentioAurum, an independent web studio that builds a small set of free web tools. No braille authority stands behind this site, and it does not speak for one. What it offers instead is that each claim can be checked. Each code follows a published standard, and the standard is named where it is used: the Rules of Unified English Braille, the Comisión Braille Española's Documento técnico B 2, the Grafia Braille para a Língua Portuguesa, the Unified Arabic Braille reference and Bharati Braille. Every figure comes with its method, and every claim about someone else's software comes with the command that reproduces it.

How pages are made and checked

No braille cell on this site is typed by hand. Each one is computed from its dot numbers or written by the site's own translation engine for that code, because a mistyped braille character is invisible to a sighted proofreader. The explanations around the cells are written for each page. They are drafted with an AI model, Anthropic's Claude, working under the publisher's direction, and none goes up on the model's word. Every braille example is written by the site's own engine and read back by the checks; every number is held in a ledger that fails when one changes unnoticed; and a sentence about history, people or practice needs a named source, linked where it is used. Mistakes still get through, and the corrections log on this page lists each one found, with the day it was put right. Wherever a sentence says what the translator does, a check reads it back against the translator: every braille example on the English pages, the words each contraction page says are written without its sign, the braille on the English, Spanish, Portuguese and Arabic phrase pages, and the Spanish number examples from the standard.

Seven checks run before the site changes. The Grade 2 translator is compared with liblouis 3.39.0 — the open-source translator whose project names the screen readers NVDA, Orca, BrailleBack and JAWS among its users — over the 5,000 commonest English words, and matches on 4,999; on the one word left, the rulebook sides with this site. 1,898 worked examples from the Rules of Unified English Braille are written cell for cell. Of the 1,043 entries in the rulebook's own Word List, 1,036 are written as the rulebook gives them, and the other seven are listed with the reason for each. The translator and the reference are compared again over a further 46,717 words, where they differ on 35, and the examples on the contraction pages are checked against what the translator actually writes. The Spanish, Portuguese, Arabic and Hindi translators are measured too: against their published word lists, against the examples in their own standards, and against liblouis on numbers and punctuation in context, where every difference that remains is recorded with its reason. The method and that one word are published in full.

What it does not do

Grade 2 output here covers the contraction system — wordsigns, groupsigns, all 75 shortforms, and the initial- and final-letter contractions — and where a rule turns on pronunciation it follows the rulebook's own examples. It does not write typeform indicators for bold, italic or underline, and a compound word that is not on its list of joins is contracted straight through; the contractions page lists what is not covered. It is reliable for learning and for checking your own work. It is not a certified transcription service, and anything being embossed for someone to rely on should be checked by a qualified transcriber.

Five codes are published here — Unified English Braille, Spanish, Portuguese, Arabic and Hindi — and each lives in its own section with its own tables. They are never mixed: a cell is looked up in the code of the page you are on. Every other language's braille assigns the cells differently again, and is not covered.

Accessibility

Every braille cell carries a text description naming its raised dots, so a screen reader announces "dots 1, 4 and 5" rather than skipping past an image. Every reference table is real HTML with proper row and column headers. Nothing is conveyed by colour alone, the whole site works at 200% zoom, and every reference page works with JavaScript disabled — only the interactive tools need it, and they say so.

If something here is inaccessible, that is a bug and worth reporting.

Privacy

Every tool on this site runs entirely in your browser. Text you type into the translator, the name generator or the bracelet maker is never transmitted anywhere — there is no server to send it to.

Google Analytics is used to count visits, to count that a tool was used — which tool and which braille code, never a character of what you typed — and to note links followed to other sites. That is the only third-party script this site adds, it is absent from the embeddable widget, and there is nothing else: no advertising network, no session recording, no cross-site tracking. Cloudflare also injects two scripts into every page — its bot detection, which profiles the browser, and its Web Analytics beacon — which this site does not control and which the privacy page describes rather than glosses over.

Licence

The braille tables are facts and belong to everyone. The written explanations are CC BY 4.0 — reuse them with attribution. There is a JSON API if you want the data directly, and an embeddable translator if you want the tool on your own site.

Every cell diagram on this site — the picture on each letter, number and contraction page — is CC BY 4.0 as well. Use them in a worksheet, a slide, a video or a print-out, commercially or not, and credit BrailleChart. Each one has a download link under it, and every one is drawn from the same verified tables the rest of the page is measured against, so the dots in the picture are the dots in the code.

Corrections

When something here is wrong, it is corrected in place, everywhere it appears, including the structured data search engines read, and the correction is listed below with its date. Corrections to the translator itself are listed too, because they change what the pages show. The checks that found most of them are published: how this site is checked.

• 20 October 2026. This page and the contractions page called liblouis the reference implementation that braille displays and embossers run. The project's own page does not say so; three studies had been corrected for the same words on 11 October and these two were missed. Both now say what the project says: that the screen readers NVDA, Orca, BrailleBack and JAWS are among its users. The four translator pages and the guide to how braille is written said that a .brf is the file embossers and braille displays read, with no source. They now cite ViewPlus, whose embosser driver sends a .brf straight to the embosser, and Bookshare, which offers its books as .brf for a braille note taker or refreshable display.
• 18 October 2026. Sections 3 to 8 of the rulebook were then gone through for examples the checks did not hold. Two more gaps were found. Fractions printed as one character, ½ or ⅜, were left out, so 5⅜ came out as 5; they are now written as the rulebook writes them, the number sign, the top number, dots 3-4 and the bottom number, and read back. And letters joined by hyphens were always given an indicator each: c-h-e-e-s-e had six. Rule 5.9.1 asks for the fewest cells and indicators that keep words contracted, and the rulebook writes c-h-e-e-s-e with one indicator for the whole run; the translator now does, wherever that is shorter, and the decoder reads such a run back as letters to its end. 68 more examples are held, 1,898 in all.
• 18 October 2026. The decoder read the numeric space of rule 6.6 as letters. Braille writes a space inside one number, as in 3 245 000 or a telephone number, as dot 5 before the next digit, and the decoder gave back 3bdejjj for the rulebook's own example, and day or ever where the digit was 4 or 5. It now reads all five of the rulebook's examples back as printed. The translator writes a numeric space where print puts a no-break, narrow no-break or thin space between two digits. An ordinary space it leaves ordinary, as the rule asks where the space is not clearly part of one number, so a list such as rooms 101 102 103 stays three numbers.
• 18 October 2026. The rulebook's accents (section 4.2) were then read in full. The translator wrote no letter with a macron, a breve or a caron, nor ł or ħ: háček, Białka and Shāh Jahān were written with those letters missing. It now writes each letter whose print form is one letter and one of the rulebook's modifiers, 63 more of them, and the 39 examples of sections 4.1 to 4.3 as the rulebook gives them. An accent typed as a separate mark after its letter, as some keyboards and copied PDFs give it, was also left out, in English, Spanish, Portuguese and Arabic: it is now joined to its letter first. The translator page gave Ł as an example of a character with no spelling; it now gives the rupee sign ₹, which has none in the rulebook's list of currency signs.
• 18 October 2026. The English translator was read against sections 3 and 4 of the rulebook, its general symbols and its accented and Greek letters, which the checks had never held it to. Of 59 worked examples there it wrote 37 as the rulebook does. The rest needed signs it left out: Greek letters, the minus sign, ratio and proportion, the prime and double prime, the dagger, the ditto mark, the female and male signs, the natural, flat and sharp, the French franc and the naira, eng and schwa, and the capital sharp s. It now writes all 59 so, and the checks hold them. Two faults came to light on the way. A word in grade 2 with an accented letter that might have read as a shortform word made the translator stop: “Im Frühling” broke it. And an accented vowel was not counted as a vowel, so the f-r of Frühling was taken for “friend” and given an indicator the rulebook's own example does not have.
• 18 October 2026. Every page was opened with JavaScript switched off. The menu, the reference tables and every page's text work without it, and each tool says that it needs it. But on 22 tool pages the tool's buttons (copy, clear, download, the picker's six dots, the finder's examples) were still drawn, and did nothing. Where a browser can say that scripts are off, they are now taken away.
• 18 October 2026. The English, Spanish and Portuguese translators had the fault found in the Arabic one the same day. A character the translator does not write, standing between a number and one of the letters a to j, took away the letter's sign, so 3☃b was written ⠼⠉⠃, which reads 32. A character that is not written is not there in braille, and the letter now keeps its sign. Separately, all five translators wrote U+FEFF, the invisible mark a pasted file can begin with, as a space, which split the word it stood in. It is now written as nothing and named among the characters left out, as other invisible characters already were. That notice also showed an invisible character as nothing at all, so it read “Left out:” followed by a blank; it now gives such a character's code point, U+FEFF.
• 18 October 2026. The Arabic translator was read against the mathematics lessons of Mada's Arabic braille portal, the reference for unified Arabic braille. It wrote no sign of arithmetic, comparison or currency: 4+5 came out as the two numbers side by side, which reads 4 5, and of the lessons' 71 worked examples it wrote 21 as the lessons do. It now writes all 71 so and reads each one back. It also writes the 50 signs of Mada's tables, each as liblouis's Arabic table (which Mada wrote) writes it, and fractions printed as one character. Where Mada's own pages disagree, on less than and greater than and on the comma in a number, the Arabic numbers page says so and says which it follows. Two faults were found on the way. When something the translator does not write stood between a number and a letter whose cell is a digit's, the letter lost its letter sign, so ٣,ا was written as 31. And on three Arabic pages, 38 runs of dot numbers that followed an Arabic word were drawn backwards, dots 1-2 as 2-1. Both are corrected.
• 18 October 2026. Every quotation on the site was looked for at the source it names, and every rule of the rulebook a page cites was read beside the page. All the quotations are there, word for word. One citation was wrong. The study of running text said this site writes the apostrophe in ’Tis under rule 7.6.15. That rule is the fallback for software that keeps no list, and by it the mark would be an opening quote. The apostrophe comes from rule 7.6.6, whose own example is “’Tis late.”, and from 7.6.14, which asks software to keep a list of such words. The page and its published list of differences now say so.
• 17 October 2026. The pages were read for sentences that are true only until something else happens. Six research pages, in all five languages, described liblouis 3.39.0 as current. It is, today; but liblouis publishes a release each quarter, and the next would have made those sentences false without a word of the pages changing. Each now gives the date the release came out, 1 September 2026, as the project itself records it.
• 13 October 2026. Every page was asked, in a browser that draws it, a set of questions about how it is laid out: whether anything sticks out sideways on a small screen or on a sheet of paper, whether every scrolling table can be scrolled to all it holds, whether small links crowd one another, whether each control shows its outline when the keyboard reaches it, whether anything is cut off when a reader widens the spacing of text, and whether anything still moves when a reader has asked for less motion. Two things were wrong. On 57 pages, the first page of this site among them, a table stood inside two scrolling boxes, one within the other, which showed as an empty bordered band under the table; each table now has one. And the page estimator's word count accepts numbers up to 100,000,000 in a field that showed five digits; it is wide enough now. Two of the questions at first answered that nothing was wrong when the measuring could not have seen a fault, so each question was tried on a fault put in on purpose before its answer was believed. The tree of roles and names that a browser hands to a screen reader was then read for every page, and nothing in it was wrong; but a screen reader also changes its voice only where a page says the language changes, and this site was not saying so. The names of the other four languages at the foot of every page, the English readings on the Spanish, Portuguese and Arabic pages of dot numbers, and the Hindi and Arabic letters quoted in this log were all unmarked, and would have been read as words of the page's own language. They are marked now. Reading those columns side by side found one more thing: the Spanish, Portuguese, Arabic and Hindi pages said that dot 4 means nothing in English braille, where the English page says it opens the two-cell signs @, $ and &. It does, and all five now say so. Last, the dates the pages state about themselves were read against the history of the files. Five research articles said, in the data they carry for search engines, that they were last modified on 2 or 4 October; their words had changed on the 10th and 11th. Four more gave only the day they were published, in August, and had changed in October. Each now gives the day its words last changed, and the tool that dates the sitemap writes that day into the page as well. And the tools were timed on whole books for the first time. Writing braille kept step with the length of the text. Reading it back did not: the decoder took 71 seconds over a novel in English and 40 in Portuguese, the page frozen the while, because at every cell it looked back at everything it had written. It now reads a novel back in five seconds, and gives the same print as before for every one of 344,592 pieces of braille tried on both. The translator, for its part, redrew every cell of a long text at each keystroke; the drawing of a long text now waits for the typing to pause, and what is copied or downloaded is ready at once as it always was. The tools were also given what nobody types on purpose. Markup that would run a script if a tool wrote it into the page as markup was typed into every field of every tool, in all five codes: each tool showed it as text. Each engine was given 30,000 strings made from the awkward corners of Unicode, and none threw or wrote anything but braille. One thing gave way: the practice drill keeps its progress in the browser, and if what came back was not the record it had kept, the first answer raised an error and nothing was recorded again. It now starts afresh instead. And the stylesheet was read for anything an older browser cannot understand. One thing was found: the bar at the top of every page, which stays in place as the page scrolls, took its background from a function that Safari before 16.2, Chrome before 111 and Firefox before 113 do not know. In those browsers it had no background at all, and the text of the page ran through the menu. It has a plain background first now.
• 12 October 2026. What had been left unchecked the day before was checked. The picture on each letter, digit, mark and contraction page was read from its pixels: all 350 show the cells of their page, and so do the twelve section images and the four charts. The contractions chart is headed “The 19 one-cell groupsigns”, and this site had taken the phrase up: its table, its answer to how many contractions there are, and its data file called the nineteen strong groupsigns. The rulebook has twelve strong groupsigns; the nineteen are those, the five strong contractions, and two of the ten lower groupsigns, in and en. The page now says so, its answer gives the count — 182, in the rulebook's nine families — and the data file, which held 162 of them, holds all 182 under those families. The privacy page said that typing into a tool makes no request. It makes one: the count that a tool was used, which the same page describes further down. Measured in a browser, that request carried the tool's name and the braille code and nothing typed. The page also named the theme as the only thing kept in the browser, where the practice drill keeps its progress there too, and called every counted download a cell diagram, where a printable chart is counted as well. Every page was laid out at the width of a phone and none ran off the side; two tools did, when given a word longer than the screen, and their words now break. Every copy button copies what its engine writes, but each then said “Copied” in English, on the button and to a screen reader, on the Spanish, Portuguese, Arabic and Hindi pages too; each now answers in its page's language. That change renamed the site's script, and the old file was taken away in the same release: for about two minutes a page still held in a cache asked for a script that was not there. The old file is back beside the new one, and stays until no cached page can ask for it. Then what a screen reader needs of a page's structure was read on every page as a browser builds it: its language, its headings, a name for every control, a caption and a scope for every table, every reference from one part of a page to another. Nine header cells on one page had no scope; nothing else was wrong. Colour was measured in both themes: 14,174 pieces of text on 46 kinds of page, none below the 4.5 to 1 that text needs. The outline of an unraised dot was below what a shape needs — 1.4 to 1 in the dark theme and 1.3 in the light — and it is the outline that says whether one raised dot is dot 1 or dot 4. It is now at least 3.2 to 1 against every background a cell stands on. And the five printable charts were read from the PDF files themselves, dot by dot: all 356 cells are the cells they should be. The two English files had been held to nothing until this day. Worked by keyboard alone, the practice drill sent focus back to the top of the page after every answer: it redraws itself, and the redraw destroyed the button that had been pressed, so each question began with a walk through the header. Focus now stays in the drill, and each result and the next question are spoken from a region the redraw leaves alone. On a narrow screen the closed menu only clipped its links, so the Tab key went into links nobody could see; a closed menu is now closed to the keyboard too. And on paper the outline of an unraised dot was fainter than on a screen — 2.85 to 1 in the Spanish and Portuguese charts, 1.83 in the English ones — where nothing else tells an a from a comma. All five PDFs now print it at 3.45. The page that documents the data files was being kept by caches for a day, under a rule meant for the files themselves, so the correction made to it on the 11th was still not what a returning reader was shown on the 12th; it is now kept no longer than any other page. The markup of every page was parsed: six line breaks on the Arabic study page were written so that a browser reads each as two, and six end tags on the page estimator closed nothing; nothing else was out of place. Without scripts, the menu on a small screen could not be opened, and three kinds of button did nothing; where a browser can say that scripting is off, the menu is now simply shown and those buttons are not. And every page was laid out under its print rules at the width of a sheet of paper: one table, the 63 cells in five codes, was too wide and would have been cut off at the edge. It now fits A4 and US Letter. The privacy page said both of Cloudflare's scripts are added to the embedded translator as to every page. Both are added; the widget's own security policy lets only one run, and a browser refuses the analytics beacon there, which the page now says. And the file written for AI agents, which describes the data files, was a day behind them: it named eleven of the fourteen fields, and said every entry's character can be built from its dots, which is not so for a digit, whose character carries the number sign. Last, every page of the live site was opened in a browser that kept nothing from any earlier visit, and what the browser reported and what it was left holding were read. Cloudflare's bot-detection script, which the privacy page described, also leaves two cookies on every page, one of them set to last a year, and the page did not say so; the page that hands out the embedded translator said the widget sets no cookies of its own, which a reader would take to mean that none are set. Both now say what was measured and what Cloudflare's documentation says of it. And the rule that lets other sites frame the widget also covered the page that documents it: that page could be framed like the widget, and its own visits were refused to analytics by the widget's policy. The rule now names the widget alone. The pages were then drawn under two conditions nobody had tried. With a contrast theme switched on — forced colours, which Windows offers to readers with low vision — every cell on this site and in the embedded translator was drawn as its own complement: a raised dot was painted as a background, a contrast theme repaints every background in the page's colour, and so the raised dots vanished while the unraised ones kept their outlines. The letter a was five rings and a gap. Printed from the browser's print dialog as it opens, without background graphics, the same thing happened on paper; the PDF charts, which are made with backgrounds on, were not affected. Raised dots are now drawn in the reader's own text colour under a contrast theme and are printed whatever the dialog says, and the dot picker, the menu button and the mark on the current page, which also spoke in backgrounds, are drawn there too. Three more things were seen in the same pictures. The outline of an unraised dot in the embedded translator was 1.49 to 1 against its page, and the edges of the page estimator's number fields and of the dot picker's buttons were 1.35 to 1; all are now 3 to 1 or more. A table's caption stood against the corner of its border. And on a narrow screen the translator ended a line in the middle of a word; a word's cells now stay together. The bracelet tool left out the space between two words: “love you” was laid out as seven cells with the e and the y side by side, which reads as one word. A space is now drawn as the blank cell it is, and counted. And a run of dot numbers such as 3-4-5-6 could be cut in two at the end of a line: on the chart as laid out for paper the number sign read “3-4-5-” over “6”, and on a narrow screen 546 runs on 136 pages were split. In tables, in lists of cells and in labels they are now kept whole; in a sentence they still wrap with the sentence.
• 11 October 2026. The letter pages in Hindi, Arabic, Spanish and Portuguese each say which other codes put a letter on the same cell, and nothing had ever checked those sentences. No Arabic or Hindi page named Portuguese, and four Hindi pages (आ, औ, छ, ख) said no other code used a cell that Portuguese or Arabic does. The sentence is now written from the five translators on all 159 pages; 84 of them changed. Eight of the ten Hindi pages on an aspirate pair said the two cells share no dot, which is true only of ब and भ and of क and ख. The pages for ड and ढ said ड़ and ढ़ take the nukta cell in front of the letter; ड़ has a cell of its own, dots 1-2-4-5-6, and ढ़ is dot 5 before that cell, as the nukta page already said. Counted again from the published word lists: the Arabic alif is in 2,477 of the 5,000 words, not in “more than half”; ज़ is not the most used Hindi nukta letter (ड़ is in 157 words, ज़ in 64); the page for ऋ gave 3 and 147 where the list has 1 and 30; झ begins 14 of the 39 words that hold it, not most of them. The Hindi pages said the newest liblouis still ships the 2014 table; it has carried Bharati 2.1 since 3.37.0, as the study says. Sentences about why a code was designed as it is, and about what readers confuse or learn first, were removed from the Hindi and Arabic letter pages, because no source for them was named. The Perkins Brailler guide now links the two Perkins histories it was written from, and has lost the details they do not contain. The study of where contractions sit said this site writes all 225 of the rulebook's examples for them; the rulebook file holds 226, and all are written. The table of endings each shortform keeps lacked the row for “braille”. Ten English letter pages said the alphabet “runs out” and that this is why five cells are words rather than letters; no source says why, so they now say only what adding dots 3 and 6 to each letter gives. Later the same day: every braille example on the site that says what it is was read back through the translators: 1,141 labelled examples, 177 dot labels, the 1,013 table rows that hold braille in Spanish, Portuguese, Arabic and Hindi, and the 63 cells of the five pages that list them. Nearly all held. What did not: the Arabic translator dropped the Arabic decimal and thousands separators, so ١٫٥ was written as the cells of 1 and 5 and read back as 15 — the tool did say the character was left out, but the Arabic numbers page used ٣٫١٤ as its own example. They are now written as the decimal mark (dots 2-5-6) and the comma (dot 2) of the unified Arabic reference, and a comma inside a number is no longer reported as left out. The table of 63 cells said dots 1-2-4-5-6 has no reading in Hindi; it is ड़. The Hindi numbers page said five consonants share a cell with a digit; seven do, and three vowels. The Hindi printable table said 19 punctuation marks and lists 21. The punctuation page said every single-cell mark is one of the first ten letters moved down a row; six of the eight are, and the apostrophe and the hyphen are not. The Arabic name page said Arabic differs from “all three other codes” in having no capital sign; Hindi has none either. The checks themselves were published the same day, at how this site is checked: the seven scripts and the files of expected differences. Five files they read are not republished, because they are other people's work — two word lists and three files drawn from the Rules of Unified English Braille — and the page says where each comes from. The signs pages were then read against the 2010 ADA Standards and the packaging guide against Directive 2001/83/EC. Every figure and clause number held. What the standard does not say was taken out: the reasons the pages gave for its rules, what sign-makers do “most often”, and “the stricter code generally governs”. One statement was wrong: the line spacing is not the tightest range in Table 703.3.1 — the dot base diameter spans less — it is the narrowest for its size. The guide's lift section named no clause; it now gives 407.4.7.1, 407.2.3.1 and the standard's table of six button messages, each of which the translator writes in the number of cells the standard counts. Five pages said Louis Braille's code had no w and that it was added later, and named no source. The one found says otherwise: Edward T. Morman's introduction to the National Federation of the Blind's edition of the 1829 book has w as the last sign of its fourth series, out of alphabetical order because French uses it only in loan words. The page for w now says that and links it. The home page listed the decimal comma among the signs Spanish and Portuguese write differently from UEB; all three write it dot 2, and the question mark is the fourth sign that differs. It left à out of the Portuguese letters Spanish does not have, counted four studies where there are eight, and said its cells are the size a fingertip feels, which no screen can promise. The FAQ said published books use Grade 2 and that every modern system renders braille characters; neither had a source, and both are gone. Last, the tool pages were read with each behaviour they describe tried on the translator. The translator page gave McDONALD'S as a word closed by the capitals terminator; it has none, because the apostrophe ends the run. The decoder page said the Grade 1 setting spells contraction cells out as letters; it leaves out, and names, the cells that are no letter. The name page said a Spanish and a Portuguese reader would each read the other's é as a different letter; neither code has a letter on the other's cell. The page estimator's table of braille pages per print page predated the present measurement: 900 words is 4.7 braille pages in Grade 2, not 4.6, and three and four braille pages need 580 and 773 words, not 582 and 776. It also cited the formats standard's rule on volumes by the wrong sub-clause and gave a reason the standard does not give. The bracelet page counted separators its pattern does not have. The finder page counted 174 contractions, a number nothing else on the site has. The alphabet page said the first ten letters moved down a row “become the punctuation marks” — six of the ten do — and called the alphabet three groups of ten. Sentences about what learners confuse, what people expect and what other tools do were taken out of seven pages, and “Grade 2 replaces words with single cells” became “in fewer cells” wherever it stood. The Spanish and Portuguese pages that are not letter pages were read the same way. The Spanish decoder page said it returns one of ¿ and ? and does not guess, and the punctuation page that the opening mark cannot be returned; the decoder decides by position and returns both, so ¿Qué? comes back as ¿Qué?. Both hubs called the alphabet three groups of ten above a third group of five letters, and built that group from the second with dots 3 and 6; they now say which accented letters come out of the same two rows. The Portuguese decoder page said a decoder that guessed º for the cell of o would spoil one word in three; 1,166 of the 5,041 lines of its list end in o, nearly one in four. The Portuguese hub said Portugal and Brazil publish their braille code together; the Grafia says they drew up its first version together in 2001, and the 2018 Brazilian edition differs in the quotation marks. Both numbers pages gave lift buttons as an example with no source, and both translators promised a contracted grade to come; those sentences are gone. In Hindi, a word beginning with ण was read back with the ण gone and nothing said — णमो came back as मो — because dots 3-4-5-6 at the start of a word was always taken for the number sign. The decoder now takes it for the number sign only when a digit's cell follows. The Arabic punctuation page said the Arabic comma and the comma inside a number are one shape; they are dot 5 and dot 2. Three studies called liblouis the translator that screen readers and embossers run, with no source; they now give the list on the project's own page. The Arabic and Hindi studies now name the word list they are measured on, and the Spanish one says the English table would fail on more than one word in six, not nearly one in five. The two other novels on the page estimator were measured again from their texts. Pride and Prejudice had been given as fitting 2.0% more words on a contracted page than Alice; that count had laid out the 154 illustration captions of the 1894 edition as if they were the novel, and without them it is 3.4%. Frankenstein's 6.4% held, measured from its first letter and not, as the page said, its first chapter. Both texts are now published with the study. Laying Frankenstein out found a gap in the translator: it left out the ligatures æ and œ, and said so, 21 times in that book. It now writes them as the rulebook does, the two letters with the ligature indicator between them (4.3.1), the same cells as the reference for every word tried. Two entries of the rulebook's Word List that were listed as differences for that reason are now right: 1,036 of 1,043, with seven listed. With three novels to hand, the translator and the reference were compared over all of them in running text, 223,107 words with their capitals, punctuation and quotation marks, which no word list has. They differ on 22 words, and the accuracy study lists them with the rule for each. The first run found one more fault here: a plural possessive inside single quotation marks — ‘to pay three months’ rent and to lose the produce of your garden.’ — was written with the closing mark after “months”. It is an apostrophe when the quotation plainly closes later, and is written so now. In Arabic, a Latin semicolon or question mark typed into the text was left out and named; both are now written as the Arabic marks, as the reference writes them. A Latin comma outside a number is still left out and named: the unified Arabic reference gives it dot 2, which after a letter is the cell of the fatha. The same comparison was then made in the four other languages, over one whole book each: Don Quijote, Dom Casmurro, Kalila wa-Dimna and Sevasadan. In Portuguese it found that the translator left out a vowel with a grave accent that has no cell of its own — flòr, què, in the spelling of 1899 — and named it; the Grafia (54.2) writes a diacritic before the letter, and so does the translator now, for è, ì, ò, ù, î, û, ä, ë, ï, ö and ñ. That put two names on the name page into Portuguese braille that the table had as outside it. In Arabic it found that the reference leaves a shadda after its letter whenever a vowel mark stands between them, which is the order Unicode normalisation gives; this site writes it before the letter, and the 5,000-word list had never shown the difference because no word in it is typed that way. Each language's study now gives its figures, and the Arabic fault is on the page of faults in the reference tables as a twelfth, with the test that will fail the day it is fixed. The checks themselves lost two sections for two releases this day: a helper that put an earlier section back cut away the ones after it. Both are restored and a gap in the numbering now fails. The two names were caught by a check that had not been lost, on the first full run after the Portuguese change. The last pages not yet read were read: the Spanish and Portuguese alphabet pages, and the Arabic and Hindi pages for names, phrases, vowel signs and the nukta. Both alphabet pages said their letters are ten base shapes repeated; two Spanish letters and three Portuguese ones are not, and the pages now count them. The Portuguese alphabet page gave 16% where the study has 15.8%, and listed the dots of eight of its thirteen letters in an answer that promised all of them. The page for ü said the cell continues to exist in Portuguese braille; the 2018 Brazilian edition of the Grafia has no ü in its table and writes it with a diacritic, while the reference table still gives it a cell, and the page now says both. The Arabic study listed five differences from the reference where, since this day, there are six. Every table of names and phrases in Arabic and Hindi was recounted from the engines, and all held. On its Grade 1 setting the decoder leaves a contraction's cell out, and its notice said this site could not read the cell as braille; it now says the cell is no letter and that the Grade 2 setting reads it. Reading a page finds what is wrong on it, and not what no check would notice going wrong later. So every number of two or more digits in the prose of the English pages was changed by one, a page at a time, and the checks were run. Of 1,845 numbers they noticed 464. About 830 of the rest stood in sections a generator writes, and the checks now run those generators and fail if any of 192 files is not what its generator writes today. The others — a clause number, a figure read in a standard, a count given once in a sentence — are now held in a ledger of every number on every page, in order, and a page whose numbers are not the ledger's fails. The ledger does not show that a number is true. It shows that none has changed unnoticed. The ledger was then tried on every run of digits on every page, 85,025 of them. It held the text a reader sees and nothing of what a reader hears or a program is given: the label of each drawn cell, the text of each image, the dot numbers in each page's structured data, the data files. Nothing else had read those either, so they were read against the engines, and this is what was wrong. In Spanish, Portuguese, Arabic and Hindi the translator and the name page drew their cells wrong: each engine put a cell's dots in as 1 to 6 where the grid takes them row by row, so dot 2 was drawn where dot 4 belongs, 3 where 2 does, 4 where 5 does and 5 where 3 does, and b was drawn as c. An unraised dot had no outline. The braille characters, the dot numbers, the copy and the download were right throughout, and the English tools were not affected; the drawn cells on those eight pages had been wrong since before the first record kept of those files, on 1 September 2026. No page held the fault, because the cells are drawn when the tool runs, and every check until this day read pages. The checks now ask each engine for all sixty-three cells and read the drawing as they read a page's. The image texts of the Spanish and Portuguese letter a said “los punto 1” and “os ponto 1”. Four Hindi letters carried a malformed dot code in their structured data. A label on the alphabet page said “dots 6” for each space between words, and the check that should have caught it built its own label the same way. The card of the study of faults in the reference tables reads 9, the rows still open; its image text said 8, its caption 5, and a sentence added earlier this day said it reads eleven. Every cell in the twenty-three data files under /api/ was right, but the Arabic file listed five differences from the reference where the study has listed six since this day, and the summary written for programs said five too, and gave 18% and 16% for shares the studies give as 17.8% and 15.8%; seven Spanish letter pages gave the 18%. Then the tools were run in a browser, in every language, and what they draw was measured against what their engines write. On the Arabic translator and name page the row of cells ran right to left, the first cell at the right, so the line read backwards, and so did the translator's list of dot numbers; a line of braille now runs left to right on every page. The cards under a name were one to a character, and so were not the cells drawn above them wherever braille does not follow print a character at a time: the Hindi halant is written before its consonant, ज्ञ is one cell, a nukta typed after its letter was dropped, the Arabic shadda is written before its letter, and a word in capitals has its two capital signs at the front. The name is now cut only where a cut changes nothing, and the cards read in order are the row. The six dots to press on the Arabic and Hindi cell finders stood in a row, where the other three lay them out as a cell. The decoder read a slash as the space between two words and then reported it as left out, and no page said that a space between dot numbers separates cells and not words; all five decoders now say to type a slash. The page estimator marked every word count over 100 as invalid, wrote “1 pages”, and still said the trade's rule of thumb is three to four braille pages to one, which the page had dropped the day before for want of a source: the sweep that removed it read pages and not scripts. Under an Arabic name the count of cells put the singular after two to ten. The rest held: every cell on every page as a browser lays it out, 9,405 of them, stands where its label says, and the translators, decoders, finder, bracelet and practice drill gave what their engines give.
• 10 October 2026. The contractions page said all eight lower groupsigns share a cell with a punctuation mark; seven do, and gg shares its cell with the word “were”. The page estimator and the page-count study said transcription agencies quote three to four braille pages for each print page, with further figures for a first-grade reader and a textbook index. No source for any of it could be found, so the pages now give the ratio as their own table does: from 1.3 to 4.6 in Grade 2, depending on how full the print page is. The Spanish phrases page said c and i differ by a single dot; each has two, and the one in the left column moves. The Portuguese and Arabic phrases pages still spoke of each phrase having a page of its own. Statements about which phrases people ask for most, what most signs get wrong and how people learn were removed from the home, phrases, practice, signs, bracelet and finder pages, because nothing published supports them here. Three guides, on what braille is, how it is read by touch and how it is written, were rewritten so that each says only what a source named at its foot says, or what is computed from the cells. The history guide named no sources. Checked against Campsie’s study of Barbier’s papers, Britannica, Perkins and the International Council on English Braille, it had several things wrong: it said a pupil demonstrated Barbier’s method at the school (the source says someone at the school), that Barbier returned to France under Napoleon (the date is not known), that Braille was totally blind by five and ill for sixteen years (neither is in the sources), that Dufau was a teacher who forced the director out (he replaced him as director), and that Unified English Braille was adopted between 2012 and 2016 (it was ratified in 2004). It now lists its sources and says only what they say. The Portuguese letter pages each said in how many of the 5,000 commonest words the letter appears, and 32 of the 39 figures were too high: the page for a said 3,041 where the published list has 3,008. All 39 are now counted from that list, as the Spanish and Arabic ones already were; three Hindi figures were one out. The Hindi pages and the write-up of the Hindi table called फ one of the commonest consonants in Hindi; in the list it is 21st of 32. The full-stop page called it the commonest mark in English prose; the measured book has 2,427 commas and 986 full stops. The Spanish and Portuguese punctuation pages said a cell with an empty top row is almost certainly a sign; no letter of either alphabet has an empty top row, so it is one. And on the letter, contraction, number and practice pages in every language, sentences about what readers feel, confuse, forget or get wrong most were removed, because no source for them was ever named. The contraction pages gave each word a place among “the most common words in English”, and 31 of them had it too high: the places had been counted with the list's 26 single-letter entries left out, so “their” was called 56th where the list has it 58th, and “character” 1,560th where it is 1,586th. They are now the word's position in the 5,000-word list the accuracy study measures against, and say so; that study now names the list and where it comes from. Nine counts of how many of those 5,000 words use a sign did not match what the translator writes — ea 150 for 148, -tion 209 for 206, -ity 42 for 40, dis- 27 for 26, and be- 24 where 15 words take the groupsign — and the page for “many” placed it third in its group where it is fourth. All of these are counted again from the list each time the site is checked. The contraction finder explained some words on the Shortforms List wrongly while writing them correctly: 122 that have something in front of the shortform were shown as that shortform alone (thereafter as “after”, boyfriend as “friend”), and a listed word with s added, such as newsletters or girlfriends, was said to hold no contraction. The cells were right in every case; the explanation now shows each sign where it stands, and it is checked against 47,659 words.
• 4 October 2026. The page estimator and the page-count study translated a book one word at a time, so a closing single quotation mark was written as an apostrophe and a passage in capitals was not seen as one. Each paragraph is now translated whole. Over the 26,439 words of the study’s book that adds one cell to each grade — 120,501 and 86,072, where the pages said 120,500 and 86,071 — and moves no line or page. The research index said every figure was derived from committed data when the site was built; the figures are measured again by the checks that run before each change, and it now says that.
• 2 October 2026. The home page, three Portuguese pages, the Spanish translator page and the brief for answer engines said Spanish braille closes a number at the decimal comma and opens it again. It does not: one number sign carries the whole number, as in Portuguese, and the translator has written it that way since 17 September. The Portuguese translator wrote an ordinal such as 1º with ordinary digits, which reads as “1o”; the standard puts the digits in the lower position. The brief for answer engines also still listed limits fixed in September. Measured for the first time against the whole Word List in Appendix 2 of the rulebook, the English translator wrote 31 of its 1,043 entries wrongly, among them fruity, asthma, Esther and captainess, and it gave evert the sign for ever. Twenty-two of the 31 are corrected and the nine left are listed with the reason for each; until then the rulebook check held only examples the translator already wrote correctly. The 75 shortform pages said which longer words keep a shortform by counting the words the reference contracts; each now names the words on the rulebook's Shortforms List, and the paragraph on recognising the sign was corrected on 28 of them. The Hindi phrases page drew two phrases with no gap between their words. Two Portuguese share images had dropped an accent, and the footer of the printable charts printed a question mark where a dash belonged. The brief for software agents under /.well-known/ still said shortforms, Arabic and the Spanish accented letters were not covered, and said the signs for in and en may not begin a word; they may, as in inside and enjoy. The 115 share images of the pages folded into their hubs on 28 September were still being served, and the phrase ones cut each phrase at eight cells and ran its words together; they have been removed. Until this date only the English translator was measured before a change went out, and the word lists the other four were measured on hold no digits and almost no punctuation. Running numbers and punctuation through each of them found what the word lists could not. Portuguese left out the dot 5 that a letter from a to j takes after a number, so 15h was written with the cells of 158; it wrote 5.2.1 as one number, gave curly quotation marks the cells of another kind, and never used the one-cell brackets the standard puts beside numerals. Hindi wrote the dash with the hyphen's single cell, the slash with the cell of a vowel and the ellipsis as three full stops, and its decoder could not read back a number with a point or comma in it. Spanish left out the raya, the ellipsis character, square brackets and single quotation marks, wrote a date such as 02.10.2024 with decimal commas, and dropped º outside an ordinal without saying so. The Arabic decoder read every full stop back as the Urdu one and lost the opening mark of «…». All are corrected. The Arabic and Hindi measurement pages said their word lists were in the repository when they were published nowhere; both lists were rebuilt from their source, reproduced every published figure, and are now under /data/. One Hindi figure could not be reproduced, a count of 205 character probes, and has been replaced by a list of 122 that is published. The name page said names are never contracted in braille; in a Grade 2 text they are, like any other word, and only the tool on that page spells them out. Thirty-two pages carried questions and answers in their structured data that were not written on the page itself; each now shows them. The Hindi dot-lookup said no sign uses the cells of the danda, the comma and nine other marks, and the English one said dot 4 opens no sign this site writes, when it opens @, $ and &. The API page still spoke of two codes, and the Arabic data counted four gaps in the reference where it lists five. Seventy-three shortform pages kept a one-line summary of the old rule of thumb after their text was rewritten from the Shortforms List. The decoder page gave 145-15-123-123-135 as its example of dot numbers, which reads “dello”; h is dots 1-2-5, and the example now reads “hello”. The phrases page promised a letter-by-letter breakdown it no longer had; each phrase now lists its cells and their dots.
• 28 September 2026. Nine English letter pages illustrated the plain letter with a word that is itself a contraction, such as “was” on W and “the” on T. The alphabet page and the home page said the third decade is the second plus dots 3 and 6; it is the first plus dots 3 and 6. The Arabic letter pages carried a paragraph on the history of Arabic and Hindi braille that said more than its sources support; it has been removed. The Spanish translator wrote a code that opens with a letter, such as b1d4, without the number sign the standard puts first.
• 27 September 2026. The 33 contraction pages still described where each sign may be used by the positions the translator had followed before 18 September, and fourteen of them contradicted what it writes now. This page, the FAQ and the translator page listed limits that had been fixed earlier in the month. Three sentences said the translator disagrees with the reference on seven words; the figure has been one since 18 September.
• 18 September 2026. The translator decided the initial-letter contractions by where their letters fell in a word, not by the rulebook's own rules. As a result it spelled “one” out in stonework and contracted “under” in laundering, both against the rulebook. The Spanish and Portuguese letter pages counted words against an earlier list, not the one the translator is now measured on.
• 13–14 September 2026. The shortform list had 73 entries. It included o'clock, which Unified English Braille dropped, and it lacked braille, oneself and thyself. The privacy page did not describe the Cloudflare beacon. Forty-nine pages said the measurements are re-run on every build after that check had stopped running.

Get in touch

Corrections are the most useful thing anyone can send. If a cell, a rule or a contraction is wrong here, it is worth more to know than anything else on this page — write to contact@braillechart.com.

If you read braille, and especially if you transcribe it or teach it, we would like to hear from you. The site is checked against liblouis and the Rules of Unified English Braille. That catches wrong output, but it cannot say whether any of it is useful: whether the practice drill teaches in a sensible order, whether the cells read well with a screen reader, whether the tone is right. Nobody who reads braille has reviewed it yet. One sentence would be more than we have now.