How this thing works

What is deslopify, exactly?

deslopify is a diagnostic. You paste a piece of writing and get back a Slop Score from 0 (reads human) to 100 (industrial-grade slop). It is not a binary AI detector — those have a track record of false-flagging non-native writers and getting their authors sued. It is a register meter, decomposed into three independent signals so you can argue with the headline number.

How does the Slop Score work?

Three engines run on every paste. Each returns its own 0–100 sub-score plus a list of highlight spans. The composite is a fixed weighted average:

slop_score = round(0.15 × lexical + 0.30 × structural + 0.55 × epistemic)

The epistemic weight is heaviest because it captures the most AI-specific failure mode (hedged stance, generic specifics, balanced non-conclusions). The composite then runs through a small curve that leaves human writing alone and reads borderline slop a touch heavier — the curve sits on the /calibration page if you want the exact shape.

What are the three engines doing?

Lexical — deterministic, no LLM. A frequency-weighted dictionary of words and phrases that are fine in isolation but cluster together in AI text (delve, navigate, in today's landscape, moreover). Normalized per 500 words. Catches surface vocabulary tells.

Structural— deterministic + light NLP. Tricolon density (“X, Y, and Z” lists), the "it's not just X, it's Y" pivot, paragraph length uniformity, em-dash density, sentence-variance flattening. Weighted sum of signals. Catches rhythm tells.

Epistemic — the only engine that calls an LLM (Claude Sonnet 4.6). Reads the text the way a critical editor would and returns a structured verdict: is the stance hedged, are claims specific, does the piece commit or balance, are the examples real or named-only? Weighted heaviest because this is where AI text actually leaks. Hard 12-second timeout; on failure the API returns the deterministic result with a partial flag rather than failing the request.

What do the verdict bands mean?

Score-to-label mapping. Higher = sloppier (intentional — “my score is 12” should be the brag).

• 0–20: Reads human. Rare. Mostly seen in personal essays, lived-experience writing, opinionated criticism.
• 21–40: Mostly clean. The default for engaged human writing that uses some shared idiom but commits to specifics.
• 41–60: Sloppy in places. A few engines are firing. Read the per-engine bars to find out which.
• 61–80: Heavy slop. Multiple engines hot. Either heavy AI assistance or a register that mimics it (corporate marketing copy, glossy newsletters).
• 81–100: Industrial-grade slop. All three engines lit. Either pure AI generation or boilerplate so generic it overlaps with it.

What's the voice profile, and what's it for?

Once you're signed in, /account lets you build a voice profile in two pieces:

An archetype — pick one of six calibration anchors per locale as your target voice (plain-spoken essay, feuilleton, literary-personal, etc.). The rewriter aims at this voice when it reshapes text.

A fingerprint — paste 1–3 samples of your own writing. We score them, take the four numbers (lexical / structural / epistemic averages plus the modal verdict band), and throw the prose away. The numbers are what stays. Per locale, so your German register is profiled separately from your English one.

Both feed the rewriter prompt. They also feed the score banner: if an engine reads > 20 points off your usual baseline, we surface a note saying so, because deviating from your own typical is more useful information than a generic verdict.

What does the rewriter do?

Below every score sits an inline rewriter preview. For anonymous users it streams Sonnet's attempt to reshape your text toward a generic plain register. For signed-in users with a voice profile set, the prompt names your archetype and your fingerprint numbers explicitly — the rewrite aims for your voice, not a homogenized one.

The preview is a sample, not the paid product. The full rewriter (rewrite-the-whole-piece-with-edits) is on the waitlist.

Is my pasted text stored anywhere?

Pasted text is sent to Anthropic for the epistemic call and is not persisted anywhere on our side — score, hash, char count, highlights yes; raw text no. Sample texts used to build your fingerprint are scored in memory and discarded. IPs are hashed before storage. The full version with retention, vendors, and your rights lives on /privacy.

Which languages are supported?

English, German, Spanish ship today. Each locale has its own lexical dictionary, structural regex bank, calibration anchors, Sonnet judge prompt, and (per signed-in user) fingerprint. If you paste English on the German page we detect the mismatch, score the text with the right engine, and offer to switch the site locale.

What does onboarding ask for, and why?

Sign in with a magic link or Google. The first screen asks for writer roles (newsletter author, ghostwriter, marketing copywriter, blogger, student post-AI, tech writer, essayist, other) — multi-select; at least one required. Roles tailor the rewriter targets and (later) which research pieces we surface to you.

The same screen offers an optional voice samples block. Paste 1–3 short samples and we build your fingerprint right then. Skip it and complete the same step later from /account when you have your own writing handy.

Where to go next

Curious about the misfires? /limitations catalogs them — short text, L2 writers, code-heavy posts, the usual suspects. /calibration shows the six anchor texts per locale and the score bands they hit. The seven /research pieces dig into the underlying claims about detectors, blind reader preference, and AI prevalence.