The method

The method is this project's attempt at reading an undefined sign. It starts from the only outside evidence there is, the source passage Király and Tokai cite for a folio. The method is checkable because every guess is tested at every other occurrence in the book.

The loop, one sign at a time

The loop that reads one unread sign has six numbered steps.

1. Take the citation. Király and Tokai's notes cite a chapter and verse for most folios. That source passage is the only outside evidence there is.
2. Look at every line. Every line where the unread sign stands, not just the one in hand.
3. Guess from the passage. The guess is the word the verse supplies for that slot.
4. Check the guess everywhere. At every other occurrence in the book. A reading that fits where it was found and fails elsewhere is thrown away.
5. Grade it. Tier A fits everywhere checked. Tier B fits most. Tier C or D is where the sign occurs too rarely for the check in step 4 to bite: it was read from one passage and there was nowhere else in the book to test it. Tier G is a guess in brackets and is counted as read nowhere.
6. Write the evidence. The evidence line goes into the proposals file with the folios and lines, so that anyone can repeat the check or overturn it.

The bar goes up before the run

The bar for every test is declared in the test's own first lines before it runs. It is never moved afterwards. Failed attempts are kept and written up with the number they failed by.

What cost the most time

The following mistakes cost the most time. They are named plainly.

• Treating a sign as an atom. Eleven attempts assumed one sign meant one word. Many signs are short phrases written without a space. Both halves were already in the dictionary.
• Inheritance versus composition. Asking whether a phrase inherits meaning gave a true negative that looked like a dead end. The right question was whether signs combine.
• A unit error. Figures were computed per glyph when a word averages 2.43 glyphs. Every number moved when it was fixed. Always confirm what one row of the transcription is before measuring it.
• Homograph collision. Keying a scraped dictionary by code and keeping the last entry made the book's commonest word read 'eleven' when it is 'and'. Sixty-five codes had to be merged.
• Blaming the corpus. A failed aligner was predicted to be failing because it used only the four gospels. A wider corpus was fetched and it scored no better. Test the cheap version of a corpus hypothesis first.
• Reading only the headword. Király and Tokai's entries carry variant spellings. Keying by headword silently dropped 88 signs that they had already read.
• Wrong arithmetic. Three versions of the arithmetic of 'what is left to read' were wrong. One counted a line as reached while an unread word still stood in it. The current figures are measured and the wrong ones are kept on the page.

The programs, and the orders

The programs are listed at The programs, as plain text. The file the next person actually works from is the standing orders, published whole and unedited: the loop, the orders, and the arithmetic of how much is still unread, with the three earlier versions of that arithmetic that were wrong kept beside the right one.

This rests on the Király and Tokai dictionary; their translation is unpublished; the readings here are this project's own.