How scoring works

Every piece published here is written in the open: the editor records the session as it happens, and the replay is the evidence. Alongside it we read a few simple things about how the writing arrived.

What is measured

Speed, sustained rather than peak

Anyone can hit a fast burst. We look at the quickest stretch held over about a minute. Fast typists are common and ordinary; only speed well beyond what a hand sustains counts for anything.

Pauses

Writing has thinking in it. We note how often the session stops for more than a couple of seconds, and whether those stops fall between words and sentences, where people naturally pause.

Revisions and going back

Most people mistype, delete, and return to an earlier sentence to change it. This is the reading that matters most: across the writing published here, even the sparsest reviser corrects something every seventeen words or so. A piece of any length that arrives with no corrections at all is counted against on its own, and the longer it runs the more it counts. Corrections that come only once every few dozen words are noted too.

How the text arrived

Whether characters appeared one at a time, or in large blocks. Blocks are not proof of anything on their own: swipe keyboards, dictation and some input methods all produce them.

Rhythm

Human typing is uneven. Timing that is too regular to be a hand is worth noting. We also look at whether the pauses fall at strikingly even intervals, which is what reading from another screen tends to produce and what composing does not.

How the keys were pressed

How long each key is held, and the gap before the next one. Whether common letter pairs go faster than awkward ones, as they do for a hand. Whether corrected slips are the kind a keyboard produces, where the wrong key was next to the right one.

Whether the session came from this editor

The editor collects a token when it loads and returns it when you publish, so a recording made elsewhere, or one replayed twice, can be told apart from a session that happened here.

Nothing is blocked

Nothing suspicious is ever blocked. A piece that reads oddly is published and marked, with the replay attached so anyone can watch and decide for themselves.

There are two marks. A piece is flagged when something about it does not look like ordinary typing, which has innocent explanations and often has one. It is called suspicious only when several independent things disagree with it at once and no single ordinary explanation covers them all. Both still publish.

Only the provably impossible is refused: a session whose recording does not reconstruct the text it claims to produce, one whose clock runs backwards, or one claiming a speed no hand achieves.

Falsely accusing a real person is the worst thing this site could do. Where a threshold is arguable, it is set to say less rather than more.

What this does not claim

This is not AI detection. Nothing here inspects your words or guesses at their origin, and no such tool works reliably.
We do measure how you type, and it would be wrong to pretend otherwise: how long keys are held, the gaps between them, where the pauses fall. That stream is turned into a handful of summary numbers when you publish and then discarded. It is never written to the database, never logged, and never sent anywhere. The recording of what you typed, which you can watch on any proof page, remains the only raw record a session leaves behind.
This is not a typing fingerprint. Nothing kept here could identify you, or match you against another writer, or follow you between pieces. One thing is compared across documents: a hash of the gaps between keystrokes, which catches the same recording being published twice. It says nothing about who typed it.
There is no tracking beyond the session itself. No accounts, no analytics, nothing follows you off the page.
A determined person can script a convincing session. This was never unforgeable. It is meant to cost real effort, and to publish the evidence for anyone to inspect.

The signals above are named in full. The weights behind them are not, so they stay tunable as real sessions accumulate.