Information density desk
Most of what
you read is slop.
Submit a URL or raw text. SlopShrink reads every paragraph like a ruthless editor — scoring information density, striking the filler, and marking up the facts, numbers, and steps worth keeping.
The house style
three marks · one passScan
Fetch the URL and parse it with Cheerio (or take your raw text), then split it into clean paragraphs.
Score
The editor rates each paragraph's information density and pulls the hard facts, numbers, and steps.
Shrink
Flip X-Ray mode to strike the slop from the page and spotlight only the lines that earn their ink.
Editorial
Provenance isn’t quality.
The internet is flooded with copy that nobody actually checked. Most detectors ask the wrong question — “was this written by AI?” — as if a byline could certify a thought.
But a human can churn out pure filler, and a model can draft a dense, verifiable brief. Authorship is a fact about the writer. It tells you nothing about whether a paragraph earns the minute it takes to read.
So SlopShrink reads for substance instead. It scores every paragraph on information density, finds exactly where the quality breaks, and surfaces the hard facts a skim would miss — then strikes the padding from the page so the signal stands alone.
Not “is this AI?” — but “is this actually useful?”
From the proofs
A specimen, marked up.
In today’s fast-paced and ever-evolving digital landscape, it is more important than ever to truly understand the power of synergy. At the end of the day, leveraging cutting-edge solutions can unlock a world of possibilities.
The global market grew 14% in 2023 to $1.2 trillion, driven largely by enterprise adoption across North America and the EU.
editor’s margin notes
- Market grew 14% in 2023.
- Reached $1.2 trillion in size.
- Growth driven by enterprise adoption.
this is exactly what x-ray mode does to a live article
Standards & practices
Why the scores hold up.
- §Calibrated rubric
- A 0–100 scale with five banded definitions and worked examples, scored on the proportion of signal — not length. Eloquent padding still scores low.
- §Deterministic
- Temperature pinned to zero, so the same paragraph earns the same score every pass. No dice-roll verdicts.
- §Self-consistency
- A reconcile step cross-checks all three signals: real facts force a “keep”; factless generic text is pinned as a cut. The model can’t contradict itself.
- §Index-locked
- Every paragraph is labelled and re-seated by its echoed index, so a verdict can never land on the wrong line.
- §Zero fabrication
- Facts are pulled only from the page in front of it — no outside knowledge, no inference, no invented figures. Nothing there means an empty margin.
- §Graceful fallback
- A dropped paragraph degrades to a neutral score instead of corrupting the whole scan.
the scoring rubric
- 0–20Pure filler — clichés, hype, throat-clearing.
- 21–40Mostly filler with a stray weak detail.
- 41–60Mixed — a real point diluted by padding.
- 61–80Mostly substance with some connective filler.
- 81–100Dense — almost every sentence carries weight.
Dispatches
By the numbers.
0–100
density scale
80
paragraphs / scan
5
llm providers
230
wpm reading base
12
paragraphs / batch
×4
concurrent passes
10s
fetch timeout
2 MB
max page size
The wire
Bring your own desk.
Pick a provider and supply your own key. It lives in this browser tab’s sessionStorage, rides along only as a request header to run your scan, and clears the moment you close the tab. Each provider is locked to its best-value workhorse model for high-volume structured analysis.
| Provider | Locked model | Mode |
|---|---|---|
| OpenAI | gpt-5.4-mini | byok |
| Anthropic | claude-haiku-4-5 | byok |
| Google Gemini | gemini-3.5-flash | byok |
| OpenRouter | openai/gpt-5.4-mini | byok |
| Ollama (local) | llama3.3 | local |
last edition
Send it to the desk.
Paste a URL or a block of text and let the editor mark it up. No account, no setup file — just your key and a page worth shrinking.
mark up a page