Articles on Tetrak OCR

Which way up is the page?

Fri, 16 Oct 2026 00:00:00 +0000

The Armenian Institute’s Empire on Trial project is transcribing the papers of Rafael v. Verelst, the suits that Armenian merchants of Bengal brought in London against Harry Verelst, the East India Company’s Governor of Bengal, in the 1770s. The papers are in the India Office Records at the British Library, and the project works from the Library’s page images: six documents, some four hundred and fifty frames, microfilmed in 1995 and scanned from the film. We built the project a reading site, each frame beside its transcript, and the first build looked wrong. Most of one document, the brief and proofs assembled for Verelst’s defence, was lying on its side.

Transliterate for display, fold for finding

Mon, 12 Oct 2026 00:00:00 +0000

Hovhannes Tumanyan is about as famous as an Armenian writer gets. He is Armenia’s national poet, and his fairy tales and verse for children are read to almost every Armenian child, in Armenia and across the diaspora. His name is on a great many book covers. Look through the Armenian Institute’s catalogue and you find it spelled at least four ways:

Tumanyan
Tumanean
Toumanyan
Tumanian

None of these is a mistake. Each is an honest romanisation of one of the two ways Armenian spells the name: Թումանյան in the reformed spelling used in Armenia, and Թումանեան in the classical spelling the diaspora kept.

What changed in ISO 9984 and ISO 9985 in 2026

Mon, 12 Oct 2026 00:00:00 +0000

ISO publishes one standard for writing Georgian in Latin letters, ISO 9984, and one for Armenian, ISO 9985. Both first appeared in 1996, and both were replaced in January 2026 by second editions. We found out by accident, while checking the references for an article about search: the citation we had for ISO 9985:1996 led to a page marking it withdrawn.

We haven’t found anyone describing what changed. ISO’s own pages list the new editions, standards bodies’ shops sell them, and that’s all. At the time of writing, Wikipedia’s page on romanising Georgian still describes the 1996 edition as current. So this article describes the changes, from the published standards, for anyone who keeps romanised Georgian or Armenian in a catalogue, a database, or a gazetteer.

Telling classical from reformed Armenian by counting one letter

Sun, 11 Oct 2026 00:00:00 +0000

For most of its history Armenian had one spelling. The alphabet Mesrop Mashtots devised at the start of the fifth century was made for the classical language, Grabar, and its spelling stayed put while the spoken language moved on. By the nineteenth century there were two modern literary standards, both a long way from Grabar: Eastern Armenian, centred on Tiflis in the Russian Empire and spoken across Persia, and Western Armenian, centred on Constantinople in the Ottoman Empire. They differed in grammar and pronunciation, but both were written in the same classical spelling, including letters that had stopped being pronounced.

When your training set has read your test set

Tue, 06 Oct 2026 00:00:00 +0000

Before anything trained on a work, we reserved pages of it. We wrote about that with some satisfaction: the whole of volume 2 of the Armenian Soviet Encyclopedia and a fixed set of pages from each of seven other works, registered in a module whose only job is to refuse. The harvester checks every harvest’s manifest against it, so a copied or renamed directory cannot get past it. The crop cutter checks it. The synthetic sampler checks it. The word-list builder checks it. The pages the evaluator downloads are, by construction, exactly the pages the training code will not touch.

Beating the baselines, starting with our own ruler

Fri, 02 Oct 2026 00:00:00 +0000

At the end of September, the comparison with every external engine left us level on word recall and well behind on character similarity. Calfa’s Tesseract model led the order-sensitive metric on almost every register. We wrote that what survived reading order was a recognition-sharpness gap, and that it was the next thing to train against.

That turned out to be mostly wrong, and the way it was wrong is the more useful half of this article.

The engine that made it up

Thu, 24 Sep 2026 00:00:00 +0000

We have been looking at a new class of OCR engine: compact vision-language models, around a billion parameters, built specifically for reading documents and small enough to run on a laptop. Two were worth measuring. One of them, PaddleOCR-VL, roughly doubled the best local character similarity we have ever recorded on a dense page of 1920s newsprint, and we are building a backend for it.

This article is about the other one, and about the twenty minute experiment that turned it from a strong candidate into something we will not ship.

The engine we had left out

Thu, 24 Sep 2026 00:00:00 +0000

Note added 2 October 2026. The tables in this article are rendered from the current comparison, so they now show v6 under a corrected metric, while the prose describes 24 September. Most of the character-similarity gap it discusses turned out to be the metric itself. See Beating the baselines, starting with our own ruler.

The cross-engine comparison we published on 4 September carried an admission in its third paragraph. Three external engines had read every page of every register, but a fourth had not: Calfa’s hye-paddle, which had held the best encyclopedia recall until v5 overtook it, had only ever been run on the encyclopedia pages, because the model directory it needed had since gone. We wrote that its row was absent rather than guessed and left it at that.

Built to be forked: how to get involved in the Armenian recogniser

Fri, 04 Sep 2026 00:00:00 +0000

Most of what we have written this week has been about numbers — which engine reads which register, and by how much. This article is about the two repositories those numbers come from, and it is written for a different reader: not someone deciding whether to use the model, but someone wondering whether to get their hands on it.

That might mean sending us a fix. It might mean adding a source we have not harvested, or a typeface we have not rendered, or a held-out set for material we do not have. It might mean none of those and instead taking a fork to train a recogniser for a script that is not Armenian at all, which the pipeline is shaped to allow. All of those are welcome, and the repositories have been arranged so that each is possible without asking us first. Here is how they are arranged, and why.

The comparison we had been putting off

Fri, 04 Sep 2026 00:00:00 +0000

Note added 2 October 2026. The character similarity figures in this article were computed with a metric we have since found to be wrong: it used difflib’s autojunk, which on a page of text ignores most of the alphabet. Word recall was unaffected by that, though three smaller scoring corrections made at the same time moved it slightly. See Beating the baselines, starting with our own ruler for what changed and the corrected comparison.

The Armenian recogniser learns both dialects

Thu, 03 Sep 2026 00:00:00 +0000

Note added 2 October 2026. The character similarity figures in this article were computed with a metric we have since found to be wrong: it used difflib’s autojunk, which on a page of text ignores most of the alphabet. Word recall was unaffected by that, though three smaller scoring corrections made at the same time moved it slightly. See Beating the baselines, starting with our own ruler for what changed and the corrected comparison.

Shipping the Armenian recogniser: what the language taught the model, and us

Tue, 01 Sep 2026 00:00:00 +0000

Note added 2 October 2026. The character similarity figures in this article were computed with a metric we have since found to be wrong: it used difflib’s autojunk, which on a page of text ignores most of the alphabet. Word recall was unaffected by that, though three smaller scoring corrections made at the same time moved it slightly. See Beating the baselines, starting with our own ruler for what changed and the corrected comparison.

Armenian OCR: the engines, the numbers, and our first training run

Sun, 30 Aug 2026 00:00:00 +0000

Note added 2 October 2026. The character similarity figures in this article were computed with a metric we have since found to be wrong: it used difflib’s autojunk, which on a page of text ignores most of the alphabet. Word recall was unaffected by that, though three smaller scoring corrections made at the same time moved it slightly. See Beating the baselines, starting with our own ruler for what changed and the corrected comparison.

What we will be writing about

Sat, 29 Aug 2026 00:00:00 +0000

The documentation on this site answers the questions a person has once they have decided to try Tetrak. What it installs, what the flags do, what the engines are, what the benchmark measured. It is deliberately static: the research pages are regenerated from the harness, so the figures on them cannot drift away from the data, and the reference pages describe a surface that only changes when the code does.

None of that is a good home for an argument, or for a note about something that did not work. This section is.