The page for somebody who will never read the README - #57
Merged
Conversation
Closes #43. Everything this project has published is written for people who already know what a false-positive rate is: the README opens with badges, CALIBRATION.md opens with a confidence interval. Both are honest and both are unreadable to the person we most need to reach. It opens by answering the only question that matters, in the size of the question: **can software tell whether a machine wrote this? No.** Not this one, not the paid ones. Every competitor implies otherwise, and saying it first is the product. Then the three things it does, laid out so the argument is visible before a word is read: the score spans the row as an opinion about prose, and the two facts — characters nobody types, a bibliography against itself — sit paired underneath. A percentage is argued about for half an hour; the same identifier on two references is settled with "can you send me the paper?". Then the number nobody publishes, drawn rather than tabulated. Ninety circles, none of them filled, and the interval drawn around the zero so the reader sees why we do not claim zero. Below it, one bar per language, because handing the pooled figure to everyone would be convenient and false — 5.6% over 65 English texts, 13.3% over 25 Spanish ones, with the axis marked at 0, 5, 10 and 15 so the drawing can be checked against the number rather than trusted. The first draft of that chart could not be: it drew both bars on a 0-10% axis and labelled one of them 13.3%, so the Spanish bar showed 9.3% under a number that said otherwise. On the page whose entire argument is that we do not lie with numbers. Then what it does *not* tell you — how much machine writing it misses, never measured and deliberately so; that the ninety are published articles and not student essays; and that the signal firing most often on human writing is uniform sentence rhythm, at 27.8%. Then why trust this one: because on 10 August 2026 it was accusing people falsely and we published that. A tidy bibliography listing further reading was announced as "1 source contradiction" while the code held a field that knew the difference and a comment explaining it. Ask any paid detector when it last accused someone falsely. Then Monday morning: straight into Docs/Teaching, with the Spanish sheet carrying the Spanish figure. **It phones nowhere.** No CDN, no web font, no analytics, no image host, and no script at all — the two languages switch on a radio input and :has(). A page arguing that your students' work never leaves your machine cannot itself call somebody else's server, and the defect we shipped in the report this month was exactly that shape. 31 KB, opens from a downloads folder. Not linked from inside the app yet, on purpose: the interface is shared with the desktop host, where this file does not exist, so an in-app link would be broken there. Placement is a decision, not an oversight. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes #43. One self-contained page at
/why.html, shipped alongside the demo, bilingual, 31 KB, no script and no external request of any kind.What it says, in the order the issue asked for
1. Can software tell whether a machine wrote this? → No. In the size of the question, before anything else. Not this one, not the paid ones. Every competitor implies otherwise and the honesty is the product.
2. So what does it do? Three things, laid out so the argument is visible before a word is read: the score spans the row as an opinion about prose; the two facts — characters nobody types, a bibliography against itself — sit paired underneath. A percentage is argued about for half an hour in a meeting; the same identifier on two references is settled with "can you send me the paper?"
3. How often is it wrong? Drawn, not tabulated. Ninety circles, none filled, and the interval drawn around the zero so the reader sees why we do not claim zero. Then one bar per language, because handing the pooled figure to everyone would be convenient and false: 5.6% over 65 English texts, 13.3% over 25 Spanish ones — with the axis marked at 0, 5, 10 and 15 so the drawing can be checked against the number rather than trusted.
4. What do I do on Monday? Straight into
Docs/Teaching/, with the Spanish sheet carrying the Spanish figure.Plus two sections the issue did not ask for and I think earn their place: what it does not tell you (how much machine writing it misses — never measured, deliberately; that the ninety are published articles and not student essays; that the noisiest signal on human writing is uniform sentence rhythm at 27.8%), and why trust this one — because on 10 August it was accusing people falsely and we published that.
A chart that could not be published as first drawn
The first version drew both language bars on a 0–10% axis and labelled one of them 13.3%. The Spanish bar was showing 9.3% under a number that said otherwise — on the page whose entire argument is that we do not lie with numbers. Caught by rendering it and reading the picture against the figure, not by re-reading the markup. Rescaled to 0–15% with a visible axis.
Verified
published-calibration.json, not from memory: 90 texts, threshold 25, 0 flagged, interval 0–4.1%, measured 2026-08-05 on engine 0.4.0.src,url(and@import; the only matches are GitHub links the reader chooses to click.file://, which is also the claim: it opens from a downloads folder with no server.:has().One thing deliberately left out
It is not linked from inside the app. The interface is shared with the desktop host, where this file does not exist, so an in-app link would be broken there. The README links it; where it belongs in the product is a decision, not an oversight.
🤖 Generated with Claude Code