Changelog

Version 0.0.14

New

• Projects and segments now have their own addresses. Refreshing the page, following a bookmark, or using the browser's Back and Forward buttons returns you to the same project or passage.

Fixed

• Refreshing inside a project now loads the complete application, including its styling and controls, rather than bare HTML.
• Translations and grammar analyses stay saved. A completed result no longer appears locally and then vanishes on reload, and an older request cannot overwrite a newer result or a manual edit.
• Large-image text extraction gets enough time to finish. OCR can continue in the background for up to 50 minutes, and its result is saved to the passage that started it.

Version 0.0.13

Fixed

• Chat no longer asks the model the same question several times. The Send button collected an extra handler every time you opened a segment, so after reading three passages one press sent three questions and spent three lots of tokens. The same fault filed an empty chat for every press of New Chat.
• An answer that never arrived is no longer written as if the model said it. When a reply came back empty the app composed “I'm sorry, I didn't receive a proper response” in the model's own voice and put it in your transcript. It now says plainly, in its own name, that nothing came back — and the case that produced it could never actually be detected before, so the spinner simply vanished instead.
• A failed question comes back to the box instead of disappearing from both the box and the conversation.
• Chat stopped announcing itself as “coming soon”. One dropped request and a feature you had been using all morning declared it did not exist yet, and disabled itself until you reloaded.
• Annotations highlight the words you actually selected. The highlighter searched the passage for matching text, so annotating the second dhamma in a line marked the first; it highlighted nothing at all if your selection crossed a bold or a line break; and annotating a word that also appeared inside markup — span, say — corrupted the surrounding element.
• Inserting a passage at a position now inserts it. Two passages could end up claiming the same place in the text, and the field treated 0 and 1 as the same answer.
• Every page loaded your account, your projects and your requests twice over, from six competing start-up routines.
• Dialogs — add a segment, share a project, manage collaborators — had no styling at all, and appeared as a bare form below the footer.

Improvements

• Every action that spends tokens says what it will cost before you press it, priced on the passage in front of you with the same rule the server bills by. The header says how many tokens you have left rather than an unlabelled number, and it keeps up on its own — the “Refresh Token Count” button is gone.
• Dialogs behave like dialogs: Escape closes them, clicking outside closes them, Tab stays inside them, and focus returns to whatever you opened them from.
• Your chats are named after what you asked rather than all being called “New Conversation” — or, in the list, a raw database identifier. The list only appears when there is more than one, and New Chat now says that the chat you are in will still be there.
• The sign-in page fits on a screen. The full instructions and privacy policy were reproduced on it in full, and were also two separate pages saying the same thing.
• The blinking marker beside every heading stopped blinking, headings no longer grow when the pointer crosses them, and the whole app now honours your system's “reduce motion” setting.
• The passage list's sort button says what it sorts by — position in the text — instead of calling it “Oldest First”.
• Failures say whether waiting will help: a busy server, an ended session and a deleted project no longer read identically.
• Admin: the accounts table lists whoever needs something first — paused, then out of tokens, then nearly out — with the counts stated above it, and granting tokens is a form that shows the current balance and the resulting total rather than a bare pop-up box.
• Warhog: a long run no longer drags you back to the experiment it belongs to if you have moved on to another one; publishing hands you a link to the published report.

Version 0.0.12

New

• Every panel saves itself. The original text, the grammar analysis, the translation and your notes no longer have Save buttons. They save as you edit them, and the line under each one says where its text stands — Saved · 14:02, Saving…, or, if something went wrong, Not saved with a way to try again. Closing the tab with unsaved text now asks first.
• Walk through the text. Previous and Next arrows beside the segment title move to the neighbouring passage in reading order, with your place shown as 3 of 12. Reading a text no longer means going back to the project list between passages.
• Segments can start empty. If the text is in a photograph, leave the text field blank when adding the segment; it opens ready for the image. You no longer have to invent a placeholder to get past the form, and text read from an image is saved for you instead of waiting on a button.
• Anything waiting for your answer is announced at the top of your projects page, and nothing is shown there when nothing is waiting.
• Creating something opens it. A new project opens on its own page; a new segment opens ready to work on.

Fixed

• All 115 languages can now actually be chosen. Only nine could be stored: picking Pali, Tibetan, Ancient Greek, Classical Chinese, Japanese, Hindi or most others silently saved no language at all. Pali, Tibetan, Ancient Greek and Classical Chinese were also missing from the reader's list altogether.
• Collaborators invited as editors can edit again. Adding a segment to a shared project was refused for every editor, no matter what the project page showed them.
• The share form and the collaborator list work. Both had been written but were unreachable, leaving sharing to a pair of stacked pop-up boxes; and roles now use one pair of words that say what they mean — an Annotator can read everything and add annotations, an Editor can change the text.
• Annotation highlighting no longer writes its own colour markup into your text each time a segment is saved, and existing text is cleaned as it loads.
• Analyse Grammar, Translate and Extract from Image no longer render as one indistinguishable block of colour.

Improvements

• No more pop-up boxes. All 64 of them are gone. Confirmations now appear briefly in the corner and say what happened in useful terms — “Translated by Claude Fable 5 · saved to this segment” — while problems stay on screen until you dismiss them. Changing the model no longer raises a box telling you what you just chose.
• Deleting is off the surfaces you scan: deleting a project lives in a fold at the foot of its page and counts what would go with it, and segments are deleted from inside the segment rather than from a button on the card you click to read.
• Model settings are folded away with your choice and its cost on one line.
• Running out of tokens explains that the budget is this app's own, that it does not refill on its own, that retrying will not help, and that everything already saved is safe.
• Empty sections say what fills them instead of only reporting that they are empty.

Version 0.0.11

New

• Presentations are now written, not extracted — the deck for a report is composed by the evaluator model itself, at the moment it finishes judging. It is a separate piece of work from the report: 10–15 slides with their own argument, not the report's text reflowed onto phone screens.
• Each deck chooses its own look. The model picks the accent colours and visual register to suit the passage under study, and can emphasise a phrase inside any slide. A deck about a Pali sutta no longer has to look like a deck about a legal contract.
• Slides come in kinds: a thesis cover, a single big statement, one enormous number, a quoted phrase as evidence, two renderings set head to head, a distilled finding, the leaderboard, a chart, and a closing.
• A signature chart in every deck — one chart per presentation is composed from scratch by the model, in a form invented for that experiment's particular question, rather than chosen from the standard set.
• Regenerate the deck on its own (Warhog lab) — rebuild just the presentation, optionally with direction (“cooler palette”, “lead with the latency gap”) and optionally with a different model. The evaluation, scores and findings are left untouched.

Improvements

• Reader Two has moved to readertwo.com. The old address still works and redirects here, so existing links and bookmarks keep working.
• Charts are legible at phone width: plot areas no longer collapse to a sliver behind their own axis labels.
• Larger image uploads now reach the app instead of being rejected by the web server.
• Reports published before this release have no deck; their presentation link is hidden until one is generated.

Version 0.0.10

New

• Mobile presentations — published Warhog reports can now open as a full-screen, 9:16 slide deck with horizontal swipe navigation, recording-friendly typography, and one legible chart per slide.
• Presentation slides carry a subtle readertwo.com watermark and close with a dedicated Reader Two conclusion card.

Improvements

• The application is now Reader Two, with its new identity applied throughout every public page.

Version 0.0.9

New

• Research reports — a new public Research page (linked in the footer, no login needed) with published model-comparison reports: side-by-side outputs, an independent evaluator's verdict with point-by-point findings, and interactive charts (criterion radar, craft-vs-soul slope, score-distribution violins).
• Models are now judged on eight criteria, including artistic taste and empathy, plus a literary characterization of each model's voice and language.
• Blind judging — experiments can be evaluated blind: the outputs are anonymized and shuffled, so the evaluator doesn't know which model produced which. Blind verdicts carry a "judged blind" badge on the report.
• Each report carries its own constellation emblem, generated from the studied passage itself.
• Available languages now include classical languages such as Pali, Tibetan, Ancient Greek, and Classical Chinese.
• Text extraction from images is available when setting up experiments — upload or paste a photo of the passage.

Improvements

• The AI model list now updates itself automatically — new models appear in the selector without an app update.
• Long AI operations no longer time out on large texts; progress is shown while they run.

Bug Fixes

• Fixed "JSON.parse" errors when grammar analysis, translation, or text extraction ran long on dense passages.
• Fixed grammar analyses being cut off partway through with the most thorough models, on long or dense passages.
• Fixed overlapping admin navigation buttons.

Version 0.0.8

Improvements

• You can now paste a screenshot straight into the Original Text field (Ctrl/Cmd+V) to extract its text automatically — no need to save the image and upload it separately.
• Faster, more reliable text extraction from images.

Bug Fixes

• Fixed text extraction failing or timing out on larger, text-heavy screenshots.
• Fixed uploads of PNG and other non-JPEG images not being recognized correctly.

Version 0.0.7

Improvements

• Updated the supported models to the latest generation: GPT-5.5, Claude Opus 4.8, and Claude Sonnet 4.6
• Added Claude Fable 5 to the supported models

Version 0.0.6

Improvements

• Added Claude Sonnet 4.5 to the supported models

Bug Fixes

• Fixed an issue with the image recognition not working with the updated OpenAI API

Version 0.0.5

Improvements

• Expanded system prompt.

Version 0.0.4

Improvements

• Switched the supported models to GPT-5 and Claude Opus 4.1

Version 0.0.3

Bug Fixes

• Fixed an issue where when editing a project does not update the language dropdowns properly with the stored information

Improvements

• Chat and grammatical analysis now are by default executed in the target language
• Project description is now passed to grammatical analysis, translation and chat prompts for a more detailed and optimized execution
• Added an instructions section

Version 0.0.2

Bug Fixes

• Fixed an issue where editing a project would create a duplicate project instead of just updating the existing one

Improvements

• Optimized prompt and prompting mechanisms for better AI responses
• Added separate pages for Privacy Policy and Changelog
• Added footer to all pages for easier navigation

Version 0.0.1

Features

• Initial application release
• Project creation and management
• Text segment organization
• Translation support with AI models
• Grammar analysis
• OCR functionality for image text extraction
• Collaborative features for shared projects
• Friend system for easier collaboration
• Token system for API usage management