# Verifying the accuracy claim This site tells readers its Grade 2 output is 99.98% identical to the reference implementation, and lists the words it gets wrong. That claim is the reason to trust anything else here, so it has to be checkable — by anyone, not just by whoever wrote it. The original check ran inside the forge build. That source no longer exists, and the site is now maintained as its own built output, so the measurement lives here instead, beside the thing it measures. ## Running it ``` powershell -ExecutionPolicy Bypass -File verify/get-liblouis.ps1 node verify/verify-ueb.js ``` The first command downloads the reference implementation once (~5 MB) and verifies its SHA-256. The second runs the measurement. Exit code 0 means the published claim holds; 1 means it does not. Before a deploy, run all seven — each exits 1 on a failure: ``` node verify/verify-ueb.js node verify/verify-rulebook.js node verify/verify-defects.js node verify/verify-dropped.js node verify/verify-pages.js node verify/verify-wider.js node verify/verify-languages.js ``` ## What it actually checks 1. **The reference is the right one.** liblouis **3.39.0** specifically — the version the published figures are measured against. UEB tables change between releases, so a different build would produce a different number and quietly invalidate the comparison. The harness refuses to run on any other version rather than reporting a figure that looks fine and isn't. Moving to a new release is a decision, made and dated on the page (17 September 2026: 3.29.0 to 3.39.0). 2. **The match rate still meets what the site prints.** The floor is not hard-coded here; it is parsed out of `research/how-accurate-are-braille-translators/index.html` — the page that makes the claim. A number the engine cannot meet fails this check instead of sitting unchallenged on the site. 3. **The page's figures are the measured ones.** The count of identical words, the total and the number that differ are all read from the page and compared with the run, so a figure the harness cannot read fails instead of being skipped. 4. **The differing words are still the same words.** `expected-differences.tsv` holds the one, with the exact cells from both sides. If a word starts or stops differing, or its cells change, the run fails and names it. A drift that keeps the *count* at 1 while changing *which* word differs would otherwise pass unnoticed. 5. **The UEB Shortforms List.** The 5,000 commonest words hold few shortforms and fewer of the longer words that keep one, so the list is measured on its own: the 75 shortforms, 537 longer words — the 531 the earlier Appendix 1 listed and the six the 2024 edition adds (aftermentioned, aftertheater, beforementioned, mustardy, unrepaid, unsaid) — and each of those with s added: 1,065 entries. Every one must also read back through the decoder as the same word. None still differ, and `expected-shortform-differences.tsv` is empty — a word appearing in it is a regression. The 2024 list also counts every word rule 10.9.3 admits; the engine applies that rule, but no word list can enumerate it, so it is checked by the rulebook's own examples instead. ## Files | File | What it is | |---|---| | `verify-ueb.js` | The measurement. Loads `tools/engine.js` and compares it to liblouis. | | `verify-defects.js` | Checks the ten liblouis defects still open on /research/reference-table-defects/ (nine of its eleven rows, and the shadda after a vowel mark, added 11 October 2026) from both ends, the table lines each cites, and that the two Portuguese rows liblouis 3.39.0 fixed are still fixed. | | `verify-rulebook.js` | Checks the 1,898 worked examples in `ueb-rulebook-examples.tsv` (Rules of UEB, 2024) are still written cell for cell, and the 1,892 marked `reads_back` still decode to their print: standing alone, wordsigns beside punctuation, quotation marks, the apostrophe, the question mark, punctuation that could be read as a contraction, numeric mode and letters after numbers (Section 6), and capitals (Section 8) — the capitals word indicator, the terminator, capitals passages and contractions beside a capital sign; shortforms in longer words and the letters that would be misread as one (10.9.3, 10.9.8, 10.9.9); strong groupsigns standing alone (10.4); every initial-letter contraction the rulebook gives a rule of its own (10.7.1 to 10.7.9); and bridging — no groupsign reaching across the join between the parts of a compound word (10.11). | | `ueb-rulebook-examples.tsv` | Those examples: section, print, the rulebook's braille, and whether the decoder reads it back. They are examples the engine writes; the measurement is the Word List below. | | `ueb-word-list.tsv` | The rulebook's Word List (Appendix 2): all 1,043 example words of its contractions section, each with the rulebook's braille. `verify-rulebook.js` writes every one and fails unless the entries that differ are exactly those in the next file. | | `expected-word-list-differences.tsv` | The 9 Word List entries the engine writes differently, with the rulebook's cells, the engine's, and the reason. | | `verify-wider.js` | Compares the engine with liblouis over `wider-list.txt` and fails unless the words that differ are exactly those in `expected-wider-differences.tsv`. Also holds two counts the pages publish: liblouis writes 1,030 of the 1,043 Word List entries as the rulebook gives them, and 500 of the 504 words beginning be are written identically. | | `wider-list.txt` | 46,717 words: every token of letters in the first 50,000 lines of hermitdave/FrequencyWords `content/2018/en/en_50k.txt` (OpenSubtitles 2018, CC BY-SA 4.0), in frequency order. | | `expected-wider-differences.tsv` | The 35 of those words the engine and liblouis write differently, with both sides' cells. | | `verify-languages.js` | Spanish, Portuguese, Arabic and Hindi. Re-measures each engine against its published word list in `/data/` (and the counts its research page prints), against the examples in its own standard (Grafia Braille 14, 16, 18, 37.1 and 38; Documento técnico B 2, 6 and 14.3), and against liblouis on about 120 to 160 patterns of numbers and punctuation in context; then checks that each engine reads back what it wrote. | | `expected-pattern-differences.tsv` | Every one of those patterns where an engine and liblouis differ, or where an engine does not read its own braille back as the print, with the reason. A line that stops being true fails the check; so does a difference that is not here. | | `verify-dropped.js` | Checks that the tools' "left out" notices never fire on supported text, that skipped cells are reported, that the blank cell U+2800 and line breaks read as themselves in all five decoders, that ¿ ¡ carry the cells in UEB Appendix 3, and that dot numbers typed in Arabic-Indic or Devanagari digits are read. | | `verify-pages.js` | Checks that the pages say what the translator does: every braille example marked `data-print` on an English page is exactly what the engine writes, every word a contraction page lists as written without its sign is written without it, and each of those pages' general rule holds when probed. The other checks guard figures; this one guards prose. It was written on 27 September 2026, after all 33 contraction pages spent nine days describing position rules the engine had dropped — fourteen of them contradicting it — and it fails in 172 places when run against that older engine. | | `get-liblouis.ps1` | Fetches liblouis 3.39.0, pinned and hash-checked. | | `corpus.txt` | 5,000 English words in frequency order: the first 5,000 lines of `google-10000-english.txt` in first20hours/google-10000-english, which is cut from Peter Norvig's word counts of Google's Trillion Word Corpus (public web pages). Compared with that file on 10 October 2026: identical, sha256 of the lines joined by LF `07ffadeaf0de99199de61137589aab0b65977385c92cf9606165ecc258a49d93`, and `verify-pages.js` holds the hash. Committed on purpose. | | `expected-differences.tsv` | The one known difference, with both sides' cells. | | `ueb-shortforms-list.tsv` | The ICEB Shortforms List: each shortform and the longer words that keep it. | | `expected-shortform-differences.tsv` | The Shortforms List words that still differ. Empty: a word appearing in it is a regression. | | `liblouis/` | The downloaded reference. Not committed — re-fetch it. | The corpus and the expected results are committed so the ground truth can be inspected directly, rather than taken on trust from a script that claims to have computed it. The binary is not, because it is 5 MB of build output that anyone can re-fetch from the project that produced it. ## Single letters The corpus holds 5,000 entries, 26 of them single letters, and all 5,000 are compared. A standalone letter that is also a wordsign — b is "but" — is written with the grade 1 indicator (`⠰⠃`), as the reference writes it; a, i and o are not wordsigns and take none. Until 14 September 2026 the engine did not do this and the harness left the 26 letters out. ## Result as of 19 September 2026 ``` reference : liblouis 3.39.0, unicode.dis,en-ueb-g2.ctb corpus : 5000 words (frequency order) identical : 4999 differing : 1 match rate : 99.98% site claims : 4999 of 5000, 99.98%, 1 words differing shortforms : 75 (ICEB Appendix 1 lists 75) list words : 1065 (shortforms, listed longer words, and those with s) identical : 1065 read back : 1065 of 1065 differing : 0, matching expected-shortform-differences.tsv exactly with s : 300 of 300 identical, 300 read back 3 yrs 6 mths: ⠼⠉ ⠰⠽⠗⠎ ⠼⠋ ⠍⠹⠎ (rulebook 10.12.2: ⠼⠉ ⠰⠽⠗⠎ ⠼⠋ ⠍⠹⠎) PASS ``` The 99.98% on the page is a floor: the run must meet it, not equal it. On 1 September 2026 the same corpus gave 4,960 of 4,974 words (single letters excluded) and 14 differences. The two that went, newsletter and newsletters, are words on the Shortforms List. Before 14 September the engine's list had 73 shortforms, including o'clock, which UEB eliminated, and missing braille, oneself and thyself; it did not use shortforms inside the longer words ICEB lists. Later on 14 September cds joined the differences. The engine now keeps a shortform's cells when s is added (Rules of UEB 10.9.5), so c-d-s also spells coulds, and it marks the letters with a grade 1 indicator as rule 10.9.7 requires — the rulebook writes "3 yrs 6 mths" as ⠼⠉ ⠰⠽⠗⠎ ⠼⠋ ⠍⠹⠎. liblouis 3.29.0 wrote all 75 shortforms with s the same way but left that indicator out, so there the reference was the one out of step. On 17 September 2026 the reference moved to liblouis 3.39.0, the current release, and six words left the differences because the engine now follows the rule the reference already did: blvd and llc (a grade 1 indicator, rule 10.9.8), ed and er (strong groupsigns used wherever their letters occur, 10.4.1), experienced (the sign for ence in "enced", 10.10.6) and money (the sign for one in a word ending "oney", 10.7.6). cds left because 3.39.0 now writes the indicator. One word joined: ar, which the rulebook writes as its groupsign ("-ar" is ⠤⠜) and 3.39.0 spells out. On 18 September 2026 four more left — multimedia, reverse, severe and coupons — when the engine took up the rulebook's own rules for the initial-letter contractions (10.7.2 to 10.7.9) in place of the positions it had measured against the reference. Three of those rules are spelling and are applied as written: "one" not after an o, "ever" only when the letters are not preceded by e or i, "under" not after a or o. The rest are pronunciation — "had" when the a is short, "time" when the letters are said as the word — and there the rulebook decides by example, so the words it writes without the contraction are held in the engine with their inflections. The same change closed eight of the ten Shortforms List differences (the hereinabove, hereinafter and hereinagain families, and greatgrandmotherhood). Later the same day the engine took up bridging itself (10.11): a groupsign may not reach across the join between the parts of a compound word, nor across a prefix where the letters are "ea". That closed reaction (re|action), whereas (where|as) and both remaining Shortforms List words (parent|hood), which is why the list now matches on all 1,065 entries. Which words are compounds is a fact about English rather than about braille, and no spelling rule settles it. A splitter reading "fat" and "her" out of father bars a contraction the rulebook uses; one told that a prefix always makes a join breaks mistake, renew and agreeable, which 10.11.5 contracts straight through. Both were measured before being rejected: the splitter took the 12,900-word benchmark from 16 differences to 44 and broke 26 rulebook examples. So the engine holds the words instead — every word the rulebook names in 10.11 with the position of its join, and eight found in the benchmark — and a compound it has never seen is contracted as before. One word remains, ar, and there the rulebook sides with the engine: it writes "-ar" with the groupsign and 3.39.0 spells out a-r. The corpus no longer separates the two, so the next measurement this needs is a larger and harder one. ## 2 October 2026: the Word List and the wider list The larger and harder measurement. `ueb-rulebook-examples.tsv` had been built by keeping the rulebook examples the engine wrote correctly, so it could show a regression and could not show a gap. Appendix 2 of the rulebook is the complete list of the example words in its contractions section, 1,043 of them, and the engine as deployed that morning wrote 31 of them wrongly. Twenty-two were gaps in rules it claimed to follow, and are fixed. Nine remain, each with its reason in `expected-word-list-differences.tsv`: two ligatures the engine does not write, three examples printed in bold or italics, two abbreviations and one abbreviated word whose reading depends on meaning, and Al, where the Word List and rule 10.9.7 disagree. liblouis 3.39.0 writes 1,030 of the 1,043. The same day the engine was compared with liblouis over 46,717 words from a subtitle frequency list. They differed on 160. Most were compounds with a groupsign across the join, which the engine now holds; evert was a plain error (10.7.4). They now differ on 35, and `verify-wider.js` fails if that set moves. ``` word list : 1034 of 1043 in Appendix 2 written as the rulebook gives them differing : 9, matching expected-word-list-differences.tsv exactly word list : the reference writes 1030 of 1043 in Appendix 2 as the rulebook gives them be- words : 500 of 504 identical words : 46717 (frequency order, OpenSubtitles 2018) identical : 46682 differing : 35 (0.07%) ``` ## 2 October 2026: the other four languages Until this date only English was re-measured before a deploy. The Spanish and Portuguese figures rested on `/data/es-reference.tsv` and `/data/pt-reference.tsv`, which nothing re-ran; the Arabic and Hindi word lists were not in this repository at all, though both research pages said they were. They were rebuilt from the source the pages name (hermitdave/FrequencyWords, 2018: `ar_50k.txt`, first 5,000 words of Arabic letters and harakat; `hi_full.txt`, first 5,000 in the Devanagari block) and every published figure came back: 4,999 of 5,000, 370 vowelled, 159 with lam-alif; 4,997 of 5,000, 1,257 with a halant, 313 with a nukta, 250 written differently by the 2014 table. One figure did not: 205 Hindi character probes, whose list is lost. `/data/hi-characters.tsv` replaces it with 122 that are published. A word list has no digits and almost no punctuation, so none of this could show the faults that mattered. Numbers and punctuation in context did: Portuguese wrote 15h as 158 (no dot 5 before the letter, Grafia 16 c), Hindi wrote the dash as a hyphen and could not read 1.5 back, Spanish dropped the raya, the Arabic decoder returned the Urdu full stop. `verify-languages.js` holds all of it now. ``` es word list ......... 5032 lines, 5032 identical to liblouis, 5032 read back pt word list ......... 5041 lines, 5041 identical to liblouis, 5038 read back ar word list ......... 5000 lines, 4999 identical to liblouis, 5000 read back hi word list ......... 5000 lines, 4997 identical to liblouis, 4996 read back hi characters ........ 122 probes, 122 identical pt standard examples . 40 of 40 written as the standard gives them es standard examples . 6 of 6 written as the standard gives them ``` ## The reference-table defects ``` node verify/verify-defects.js ``` /research/reference-table-defects/ names five rows in liblouis 3.39.0's Arabic table, three in its Spanish table and one in its Portuguese table (new in 3.39.0) that depart from their own rule or the published standard, and two rows of its Portuguese table that did in 3.29.0 and are fixed in 3.39.0. For each one this checks both halves the page states — the reference still gives the answer quoted, and this site's engine (`tools/engine-ar.js`, `tools/engine-pt.js`) gives the other — plus that the table lines cited really hold what the page says (for example, the 36 shadda rules on lines 97–132 of `ar-ar-g1-core.uti`). If liblouis fixes one, the run fails and the page is what must change — which is how the two Portuguese fixes were caught. For those two it checks the fix still holds. `verify/liblouis-3.29.0/`, when present, is the previous reference, kept only to compare releases; nothing reads it by default. ## Not covered Nemeth and music braille are **not** verifiable this way. liblouis 3.29.0 shipped `nemethdefs.cti` — a definitions include — but no standalone Nemeth translation table, so there is nothing to measure against. Do not publish Nemeth tables on this site until there is a reference implementation to check them with. The Spanish, Portuguese, Arabic and Hindi engines are measured by `verify-languages.js`. ## The book: which contractions do the work (4 October 2026) `measure-usage.js` translates the whole of `data/alice-body.txt` in Grade 2 and asks the engine, through `BR.explainText()`, which sign it used for every word where the word stands. A sign's saving is the letters it stands for less the cells it takes. The script stops if the savings of all signs, less the grade 1 indicators Grade 2 adds before letters standing alone, do not equal the difference between the Grade 1 and Grade 2 transcriptions exactly: 34,447 - 18 = 34,429 = 120,501 - 86,072. node verify/measure-usage.js # summary node verify/measure-usage.js --write # rewrites data/alice-contractions.tsv `verify-pages.js` (section 7) measures it again and fails if the data file, the research page `/research/which-contractions-do-the-work/`, the "In a whole book" section of any contraction page, or the cells, lines and pages printed on `/research/how-many-braille-pages/` and `/pages/` say anything else. Found while reconciling: `layoutBraille()` translated one word at a time, so it could not tell a closing single quotation mark from an apostrophe. It now translates each paragraph whole, which added one cell to each grade of the book and moved no line. ## Generators kept with the checks Three sections of the site are written by scripts rather than by hand, and the scripts are kept here so they can be run again when the engine changes (run each without `--write` first to see the result): node verify/build-usage-pages.js --write # /research/which-contractions-do-the-work/ and "In a whole book" on 147 contraction pages node verify/build-digit-pages.js --write # "What is particular to N" on the ten digit pages node verify/build-phrase-cells.js --write # the cell-by-cell line under each phrase on /phrases/ node verify/build-punctuation-grade2.js --write # "In Grade 2, the same cell" on the eight punctuation mark pages (18 October 2026) node verify/build-grades-guide.js --write # /learn/grade-1-vs-grade-2/ node verify/build-learn-guides.js --write # /learn/what-is-braille/, /learn/how-to-read-braille-by-touch/, /learn/how-braille-is-written/ The three guides built by `build-learn-guides.js` state nothing about people, teaching or history that is not in a source named at the foot of the page; the sources and the date each was read are listed at the top of the script. What they say about cells is computed and asserted there. ## Printable tables as PDF (10 October 2026) node verify/build-chart-pdfs.js # prints /es/chart/ and /pt/chart/ to PDF with headless Chrome The files are the pages themselves, printed with the site's print stylesheet, so their cells are the cells the suites already check. `verify-pages.js` fails if a chart page changes after its PDF was printed (`verify/chart-pdfs.json` records what each was printed from). Arabic and Hindi are not offered yet: the page prints correctly, but the text layer of the file comes out with decomposed letters, which a screen reader would misread. ## The book, letter by letter (10 October 2026) `measure-letters.js` counts, for each of the 26 letters over `data/alice-body.txt`, how often it is printed, how often Grade 2 still writes it as its own cell, which contractions take the rest, and what its cell is doing each time it appears. It stops unless spelled + inside contractions equals printed for every letter, and every appearance of a letter's cell has a job found for it. node verify/measure-letters.js --write # data/alice-letters.tsv node verify/build-letter-usage.js --write # /research/what-grade-2-does-to-each-letter/ and "In a whole book" on the 26 letter pages `verify-pages.js` (section 8) counts it again and fails if the data file, the study or any letter page says anything else. ## Letter-page word counts (10 October 2026) Every Spanish, Portuguese, Arabic and Hindi letter page says in how many of the 5,000 commonest words its letter appears. Nothing counted those again until 10 October 2026, when 32 of the 39 Portuguese figures turned out to be too high and three Hindi ones one out. `verify-languages.js` (section 5) now counts all 154 from `data/-reference.tsv` and fails on any that differ. ## Places and counts from the 5,000-word list (10 October 2026) The contraction pages say where a word stands in the 5,000-word list, how many of the 5,000 are written with a sign, and how many carry an initial-letter contraction inside them. Nothing counted these again until this date, and 53 of 97 did not come back: 31 places had been counted with the 26 single-letter entries left out (their 56th for 58th, character 1,560th for 1,586th) and were stated as places "in English"; nine counts no longer matched the engine (ea 150 for 148, -tion 209 for 206, be- 24 where 15 words take the groupsign); the page for "many" put it third in its group, and it is fourth. A place is now the position in `corpus.txt`, as on the accuracy study, and the pages say which list. `verify-pages.js` (section 9) counts every place, count and comparison again, holds the hash of the list, and fails on a page that gives a place "in English". The same section checks `explainWord()`, which the contraction finder prints: over the 5,000, the wider list and every listed shortform word with and without an s (47,659 words), the segments must spell the word and add up to the cells the translator writes. Until this date they did not for listed words with something before the shortform (thereafter was explained as the shortform "after" alone, 122 words like it; unfriendly had its letters out of place), for a listed word with s added (newsletters had no explanation), or for a word respelled under 10.6.10 (bein). The cells were right throughout: over 189,244 word forms and the whole book the translator's output is unchanged. ## The letter pages of four languages (11 October 2026) `verify-languages.js` section 6. Each Spanish, Portuguese, Arabic and Hindi letter page says which other codes put a letter on the same cell; the sentence had been written once and never checked. No Arabic or Hindi page named Portuguese, the hamza forms and क्ष/ज्ञ were missed, and four Hindi pages (आ, औ, छ, ख) said no other code had a letter on a cell that Portuguese or Arabic uses. The sentence is now generated from the five engines (the letters of a code are the letters that have a page in its alphabet) and the suite builds it again for all 159 pages. The same section holds 47 sentences on the Hindi and Arabic pages about cells and about the word lists, each rebuilt from the engines or from `data/hi-reference.tsv` and `data/ar-reference.tsv`: the dots of the ten Hindi aspirate pairs and what each pair shares (eight pages had said "no dot"), ड़ as its own cell ⠻ and ढ़ as ⠐⠻, the commonest conjuncts, how often न् and anusvara are written before a dental, the Arabic letter ranks, the sixteen Arabic letters whose cell is a letter in English, Spanish and Hindi too. It also fails on any Hindi or Arabic letter page that says why the code is as it is ("यह संयोग नहीं है", "الهيكل اللاتيني"): no source for those was ever read. Run against the pages as they stood on 10 October it fails in 213 places. Section 5's pattern for the Hindi count read "शब्दों?" — optional anusvara, not optional plural — and so could not match "१ शब्द"; it is "शब्द(?:ों)?" now. `/learn/perkins-brailler/` was checked the same day against the two Perkins documents it names (History of the Perkins Brailler, J. Seymour-Ford, April 2002; Howe Press & the Perkins Brailler, McGinnity, Seymour-Ford and Andries, 2004) and Perkins' brailler store, and now links all three. `verify-languages.js` section 7 rebuilds the computed sentences of the Spanish and Portuguese letter pages (letters one dot away, the decade of the cell, the cells English braille spends on the letter): 155 sentences, all of which held when first counted. ## The two older studies (11 October 2026) `/research/contraction-positions/` and `/research/shortform-endings/` report surveys taken against liblouis 3.29.0 over a word list that was not kept, and say so; those counts cannot be taken again. `verify-pages.js` section 10 holds what can be: each percentage is the ratio of the counts beside it, the totals add up, the table of endings is the engine's own `shortformEndings`, 72 shortforms take an ending and three (about, almost, him) take none, and the number of 10.7 examples is the number in `ueb-rulebook-examples.tsv`. It found two things: the page said 225 examples where the file has 226, and the endings table had no row for "braille". `verify-pages.js` section 11 does the same for the 26 English letter pages: 146 computed sentences (the dots of the letter, the letters one dot away, the family of three, the digit, "N cells instead of M" beside each practice word). All held when first counted. Ten pages carried the sentence the contractions hub lost on 10 October — that the alphabet "runs out", which is why and, for, of, the and with have cells — and it is gone from them too; the section fails if it comes back. ## Every example that says what it is (11 October 2026) `verify-languages.js` section 8, over all five languages: - (a) every `.braille-line` and `.braille-text` whose label names its print ("“una” en braille: 3 celdas") holds the cells the engine of that language writes, and the count in the label is the number of cells — 1,141 examples; - (b) every `.braille-inline` whose label gives dots holds those dots — 177; - (c) every table row holding braille in Spanish, Portuguese, Arabic and Hindi is borne out by the print beside it, the dot numbers beside it or the cell count, and any dot numbers in the row are the dots of its cells — 1,013 rows; - (d) numbers: the Arabic decimal separator U+066B and thousands separator U+066C are written between digits as the decimal mark (2-5-6) and the comma (2) the unified Arabic reference gives (Mada, braille.mada.org.qa, read this day), and reported as left out anywhere else. Until now they were dropped and ١٫٥ read back as 15. liblouis' ar-ar-g1 does not define either character and writes its code point; both lines are in `expected-pattern-differences.tsv`; - (e) the 63-cell tables on the five dots pages name, for every cell, exactly the letters the engines write with it — 756 columns (found: Hindi ड़ on dots 1-2-4-5-6 was missing); - (f) the counts in the headings of the Hindi and Arabic printable tables are the rows below them (found: 19 punctuation marks announced, 21 listed). `verify-pages.js` section 12 rebuilds what the eight punctuation-mark pages and their hub say: which letter each mark is when moved down a row (six of the eight are one), and which marks stand one dot apart. Read against liblouis 3.39.0 the same day and found true: the Hindi halant page's account of the four two-letter rules in `devanagari.cti` (lines 317–322) and of कर्म and भर्ती; the Arabic pages' account of the tatweel (two cells of dots 2-5) and of the letter sign missing before fa. ## Signs, the home page and w (11 October 2026) `verify-pages.js` section 14 holds the signs section to the 2010 ADA Standards as read on ada.gov that day: the five ranges of Table 703.3.1 in inches and millimetres on two pages, the clause numbers and figures of 703.2, 703.3, 703.4 and 216 on the pages that give them, and Article 56a and 54(a) on the packaging guide. Those figures are the standard's, not computed; the check exists so that one changed by accident fails. Two things are computed: the six elevator messages of Table 407.4.7.1.3, which the translator must write in the cells the standard counts (stop 3, alarm 4, open 3, close 5, main 3, phone 4), and the cells each legend on the capital-letters page costs in capitals and as the standard allows. It also fails if a page gives a reason for a rule that the standard does not give — the pages had a "Reason" column, and "which is why braille is set by machine". `verify-languages.js` section 9 rebuilds what the home page says about the three Latin-script codes: that the capital sign, full stop, question mark and brackets differ from UEB's in both Spanish and Portuguese; that the decimal comma does not (the page had said it did); which accented letters the two share, which they write differently and which only Portuguese has (à was missing); and that the studies it counts are the cards on `/research/` (it said four; there are eight). The same section holds five pages and `llms.txt` on the letter w. They said the code had no w and that it was added later, with no source. Edward T. Morman's introduction to the National Federation of the Blind's edition of Louis Braille's 1829 book says w is the tenth sign of the fourth series there, out of alphabetical order because in French it appears only in loan words. The pages now say what is computed — w is j with dot 6 added; u, v, x, y and z are a to e with dots 3 and 6 added; a to i with dot 6 added are ch, gh, sh, th, wh, ed, er, ou and ow in UEB — and name that source for the history. The check fails if "added later" comes back or the source goes. ## The tool pages and the hubs (11 October 2026) `verify-pages.js` sections 15 to 18 try, on the engine, what the tool pages say the engine does. - 15, translator and decoder: ABCs takes the capitals terminator and McDONALD'S does not (the page had it as an example of the terminator; Rules of UEB 8.4.2 and 8.6.3); BSc, a passage of three words in capitals, two words that are not a passage; no contraction across a capital sign; a letter sign on a to j after a digit, a decimal point or a comma, and on no other letter; the `.brf` in ASCII, 40 cells by 25 lines; dots 2-3-6 read as a quotation mark before a word and a question mark after one; what the Grade 1 setting does with contracted braille ("shop" comes back as "op" with the first cell named as left out — the page had said the cells are spelled out as letters); and that more than half of the 63 cells have more than one reading in the lookup's table. - 16, name, bracelet and practice: the 24 names (letters, and cells one more); the seven accented names in three codes; José in Spanish and Portuguese, alike but for the é, and each decoder reading the other's as "Jos" with the cell named (the page had said each reads a different letter); the Spanish é as the UEB cell for "the"; the bead count of a six-letter word with and without a capital; and the drill's rules as its script has them — three wrong answers from the cells one dot away first, five boxes, local storage. - 17, page estimator and finder: every figure on `/pages/` and its study that follows from the layout of the book — the worked table, 0.124 lines a word, 26.2 cells a line, 29% and 22%, the five ratios of braille pages to a print page, and the words three and four braille pages need (the table predated the present measurement: 4.6 for 4.7, 582 and 776 for 580 and 773). The constants in `tools/estimator.js` must be the book's as laid out today; its two cells-per-word figures were one cell behind and were rewritten. BANA Braille Formats §1.6.1 is held as read on brailleauthority.org. The finder's preference rules are tried: thence in three cells (10.10.2), ence before a, d and r (10.10.6), one-cell groupsigns before two-cell contractions (10.10.7). - 18, alphabet hub: six of the first ten letters are punctuation marks when moved down a row, and the alphabet is ten, ten, five and w. Not re-measured: the two other novels the estimator page mentions (2.0% and 6.4% more words on a contracted page). Their texts are not in the repository. ## Spanish and Portuguese: hubs, numbers, punctuation and tools (11 October 2026) `verify-languages.js` section 10 rebuilds from the two engines what those pages say beyond their tables: - the hubs' account of the alphabet — the second decade is the first with dot 3, u v x y z are the first five with dots 3 and 6, and which accented letters come out of those two rows (Spanish á é ú from h i j, and ñ ü w from g h j with dot 6; Portuguese ç é á ú, and â ê ô à ü õ w). Both hubs had "three groups of ten" above a group of five; - the punctuation pages — no letter has an empty top row, most marks in the table have one, and an apostrophe between two letters is the full stop's cell and is read back as a full stop; - the Spanish decoder — ¿Qué?, ¡Hola! and ¡¿qué?! come back as written (two pages said the opening mark could not be returned), and all double quotation marks come back as « »; - the Portuguese decoder — the share of the reference list ending in o (1,166 of 5,041: the page said a guessed º would spoil one word in three), and 1º read back from the lowered digit; - the tables of names on both name pages, letters and cells. Held to a source, not computed: the Portuguese hub's sentence on the Grafia Braille para a Língua Portuguesa (first version drawn up by Brazil and Portugal under a cooperation agreement of 2001 — presentation of the 3rd Brazilian edition, MEC, 2018). ## Arabic and Hindi: numbers, punctuation and tools (11 October 2026) `verify-languages.js` section 11 tries on the Arabic engine what `/ar/numbers/`, `/ar/punctuation/`, `/ar/translator/` and `/ar/decoder/` say: the seven letters that stand on digit cells and take a letter sign after a number; ١٢٣٤ and 1234 as the same five cells; lam and alif as one cell; the shadda written before its letter and read back onto it; the 24 marks of the punctuation table; and that the Arabic comma is dot 5 while the comma inside a number is dot 2 (the page had called them one shape). Later the same day the Latin ; and ? were added as the Arabic marks' twins — the unified Arabic reference prints the semicolon's row with ";" itself, and liblouis writes both so — and four rows left `expected-pattern-differences.tsv`. The Latin comma outside a number is still not written and is named as left out: the reference calls it the mathematical comma and gives it dot 2, which after a letter is the cell of the fatha, and that is what liblouis writes for it. Section 12 does the same for Hindi. It also holds a fix to the decoder made that day: dots 3-4-5-6 at the start of a word is the number sign only when a digit's cell follows, and ण otherwise. Until then णमो came back as मो, the ण dropped without a word. A word beginning with ण and then a letter on a digit's cell (णब) is still read as a number; the code has no letter sign to tell them apart, and the pages say so. The 5,000-word list round-trips as before (four malformed lines fail before and after). Also in section 10: the share of each list's first 5,000 words that carry one of the code's own letters (890 in Spanish, 791 in Portuguese), as the two research pages give it. ## Two more novels, and the ligatures (11 October 2026) `data/pride-body.txt` and `data/frankenstein-body.txt` are the story text of Project Gutenberg eBooks #1342 and #84, cut as `data/alice-body.txt` was: from the first heading to the last line of the novel, nothing of Project Gutenberg's or the edition's around it, blank lines between paragraphs kept single. In Pride and Prejudice (the George Allen edition of 1894) the 154 illustration blocks are taken out, brackets matched, and the first chapter's heading, which the edition sets inside one of them, is written CHAPTER I. Frankenstein starts at Letter 1. `verify-pages.js` section 17 lays both out as the estimator lays out Alice and holds the two pages to the result: 199.5 and 205.4 words to a contracted page, 3.4% and 6.4% more than Alice; cells to a word 3.26, 3.33 and 3.30; 33, 57 and 95 words to a paragraph. The pages had 2.0% for Pride and Prejudice. That figure reproduces exactly if the illustration blocks are left in and laid out as text, which is how it was made. It fails if either text holds an illustration block or the words Gutenberg, or if the translator leaves any character of them out. It did leave one out. `tools/engine.js` had no æ or œ: 21 in Frankenstein, one in Pride and Prejudice, each named as left out. They are now written as the rulebook gives them (4.3.1 to 4.3.3): the two letters with the ligature indicator, dots 4-5 then 2-3-5, between them; no contraction takes in a joined letter; a capital Æ under a single capital sign carries a second sign before the indicator, and in a word in capitals it does not. `verify-ueb.js` compares 34 words with liblouis 3.39.0 (identical in Grade 2, and read back in both grades) and holds the rulebook's three examples. Two rows left `expected-word-list-differences.tsv` (ædile, færie): ``` word list : 1036 of 1043 in Appendix 2 written as the rulebook gives them differing : 7, matching expected-word-list-differences.tsv exactly ``` ## Running text: three novels against the reference (11 October 2026) `verify-wider.js` now also translates the three novels in `data/` whole, in Grade 2, with this translator and with liblouis 3.39.0, and compares them word by word: 223,107 words with their capitals, punctuation, quotation marks and numbers in context. `lou_translate` breaks a long input line into several output lines, so a paragraph cannot be given to it whole. (The first attempt did that, and nearly every paragraph of the two longer books came back out of step; it measured nothing.) Each paragraph is cut at spaces into pieces of at most 300 characters, preferring the end of a clause, and the reference's pieces are joined again. This translator is given the paragraph whole, because what closes a quotation may be many words on. ``` running text: 22 of 223107 words in three novels differ (0.010%), each in expected-context-differences.tsv ``` `expected-context-differences.tsv` has the 22: book, the words around, the word, both spellings and the rule that decides it. Eleven are an apostrophe opening a word (’Tis, ’em), which the reference writes as a closing single quotation mark (7.6.15); six are two chapter headings of Alice, where the reference makes CHAPTER, the numeral and the A that begins the title one capitals passage (8.5.4); two are the apostrophe in General----’s; two the long dash of Frankenstein's "——" (7.2.4); one is the closing mark of ‘Times’, which the reference writes as an apostrophe. The first run found one fault in this translator. Inside single quotation marks, the mark after a plural possessive was taken for the closing mark: ‘to pay three months’ rent and to lose the produce of your garden.’ It is now an apostrophe when the word ends in s, a word follows, and a mark that follows punctuation closes the quotation before another opens. The rulebook says of this that it "cannot be wholly automated with complete accuracy" (7.6.13); ‘the girls’ school’, with nothing before the last mark to show that it closes, is still read the old way. `verify-pages.js` section 19 holds eight sentences of that kind and the figures the study and `/checks/` give. `build-checks.js` publishes the new file: eighteen files in `/checks/verify/` (nineteen with the four languages' file, added later the same day). ## Running text in four more languages (11 October 2026) `verify-languages.js` section 13 translates one whole book in each language with this site and with liblouis 3.39.0 and compares them word by word. Both are given the same pieces, each paragraph cut at spaces to at most 300 characters. Every word that differs is a row of `expected-running-text-differences.tsv`: language, kind, the words around, the word, both spellings, the reason. A piece the two do not split into the same number of words is a row of its own. | | text in `data/` | words | differ | |---|---|---|---| | Spanish | `quijote-body.txt` — Don Quijote, Project Gutenberg #2000, from the first part to the end of the second | 379,584 | 353 | | Portuguese | `casmurro-body.txt` — Dom Casmurro, Project Gutenberg #55752, in the spelling of 1899 | 65,119 | 132 | | Arabic | `kalila-body.txt` — Kalila wa-Dimna, the 19 chapter pages of Arabic Wikisource | 45,749 | 58 | | Hindi | `sevasadan-body.txt` — Sevasadan (Premchand, 1919), scan pages 6 to 299 of the Hindi Wikisource index, once each | 89,098 | 161 | What it found: - **Portuguese.** The translator left out a vowel with a grave accent that has no cell (flòr, què) and named it. The Grafia Braille para a Língua Portuguesa (3rd edition, 54.2) puts a diacritic before the letter — 2-6 grave, 4 circumflex, 4-5 diaeresis, 5 tilde. `engine-pt.js` now does, for è ì ò ù î û ä ë ï ö ñ, and reads them back; the 2-6 of the grave and of the question mark are told apart by what follows. The reference writes ò with the cell of ó. The other 99 differences are the underscores the edition marks italics with. - **Arabic.** 57 of the 58 are one thing: a shadda that follows its letter's vowel mark. That is the order Unicode normalisation (NFC) produces — the shadda's combining class is 33, the fatha's 30. The reference table moves a shadda before its letter only when the two are adjacent (`ar-ar-g1-core.uti`, the `always بّ 6-12` rules), so here it leaves it after the vowel. The 5,000-word list has no word typed in that order, which is why it never showed. It is on the defects page as a dated card after the five Arabic rows, and `verify-defects.js` holds it from both ends (ten defects now). No fix is offered: none has been tried on a copy of the table. - **Spanish.** Nothing but the edition's two apostrophes for an inner quotation (346) and seven words with ï, à or ù. - **Hindi.** Eleven kinds, most of them slips in the transcription that the two recover from differently: a comma with no space after it, a colon typed for the visarga, a mark typed twice. One is the reference's two-letter rule again (खम्भे, the fault of भर्ती). The Hindi text needed care. The chapter pages of the Wikisource edition overlap one another from chapter 21 on and repeat passages up to six times; the body is built from the scan pages themselves, rendered once each in order (`ws_range.py`, `hi_body.py` in the working notes). Regenerate the expected file with the generator kept beside these notes; it assigns each difference its kind and stops at one it cannot place. ### The lost sections `verify-pages.js` section 18 and `verify-languages.js` section 12 were missing from two releases this day. The helper that inserted a section replaced everything from that section's start to the end of the file's checks, so putting section 17 or 11 back removed 18 and 12. Nothing failed, because a missing check cannot fail. They are restored; sections are now put in one at a time, and `verify-pages.js` section 19 fails if the numbering of either suite has a gap. Section 16, which had not been lost, caught two cells of the name page that the Portuguese change had made wrong, on the first full run after it. ## The last pages read (11 October 2026) `verify-languages.js` section 10 now also holds the two alphabet hubs: how many letters are one of the ten base shapes with dots added (31 of 33 in Spanish, all but í and ó; 36 of 39 in Portuguese, all but ã, í and ó), the dots of all thirteen Portuguese letters in the answer that lists them, the 15.8% the study gives, and that the pages say the 2018 Grafia's table has no ü (read on page 23 of the 3rd Brazilian edition; 54.2 writes Führer with the diaeresis diacritic) while the engine, like liblouis 3.39.0, gives it dots 1-2-5-6. Section 14 recounts the tables of names and phrases on the Arabic and Hindi pages (35 rows: letters, cells, and the cells of the Latin spelling in English braille), tries what the matra and nukta pages say (कइ is क, dot 1, इ; both spellings of क़ are dot 5 and then क; liblouis 3.29.0 wrote ऩ, ऱ and ऴ with the cells of the anusvara, ड़ and ल), counts the two spellings of a nukta letter in the 5,000-word list (85 and 228), and holds the Arabic study to the number of differences it lists, which is six since the shadda after a vowel mark was added. ## Changing every number by one (11 October 2026) Reading a page finds what is wrong on it. It does not find what no check would notice going wrong later. So every number of two or more digits in the prose of the 233 English pages was changed by one, a page at a time, and the six suites that read English pages were run. A change counted as noticed when a failure named its page. Of 2,252 runs of digits, 407 were not numbers a check could own (the 39 of an apostrophe's character reference, "CC BY 4.0", "letter 3 of 26"). Of the 1,845 that were, the suites noticed 464. About 830 of the rest stood in sections that a generator writes: the checks had compared the figures they were written for and not the sentence around them. `verify-pages.js` section 20 now runs the six generators with the writing replaced by a comparison, and fails if any of the 192 files is not what its generator writes today. Where the generators are absent (they are not published) it says so. The remaining 553 are not this site's to compute, or are computed by nothing kept: a clause number, a figure read in a standard on a stated day, a count from the 5,000-word list given once in a sentence. They are held instead. `build-number-ledger.js` writes `page-numbers.tsv`, every number on every page in order, and section 21 reads each page again and fails on the first number that differs. **What the ledger is not.** A number in it is not thereby true. It says the number is the one that was read, and that it has not changed since without the ledger changing with it. node verify/build-number-ledger.js # what differs node verify/build-number-ledger.js --write # after a change that was meant; read the difference first Run again with the ledger, every one of the 2,079 runs of digits that are numbers was noticed: 881 named by a suite's report and 1,198 by the ledger alone. (The report names twelve pages at a time and the test changes every page at once, so the second figure was measured on the ledger directly.) ### What the ledger did not hold The ledger was then tried by itself on every run of digits on all 444 pages — 85,025 of them, in the text, the title and descriptions, the structured data and the attributes a reader is given. It held all 37,104 in the text it was written to read and none of the rest: | where | runs of digits | held by the first ledger | |---|---|---| | the label of a drawn cell or a run of braille (`aria-label`) | 37,205 | 0 | | text of `
`, titles, descriptions, text of the structured data | 37,102 | all | | navigation and footer | 1,880 | 2 | | term code in the structured data | 1,627 | 0 | | image text (`alt`) and its caption in the structured data | 2,590 | 0 | | dot numbers set in ``, `
`, a `