It makes what you already have leaner, safer and clearer, and shows its work for every change.
What is the Schwarz Diamond? A shape whose channels follow the pattern of the atoms in a diamond: each point joined to its neighbours, the least material holding up the most surface. The tools work the same way: what you give goes in, is turned once, and comes out ready to be the next step's input, chained like a diamond's atoms. After seven steps the whole becomes one new piece, so nothing grows out of control. No AI involved.
Dependencies with published flaws, keys and tokens left in the project's files (masked in the report), scripts loaded from other sites, and dependencies never imported.
Checked by
Standard checks, including the public advisory database.
Where input, transform or output takes over: patterns that can run away on a long input, loops with no way out, data accepted with no size limit, and failures swallowed in silence.
Checked by
Each runaway pattern is proved by timing it on longer and longer input, in a separate thread with a time limit.
Stray spaces, doubled words, repeated paragraphs, page clutter, filler, and sound-alike words where the words next to them settle the spelling (its going, their is, your welcome); and a summary made of whole sentences copied word for word.
Checked by
Every number, name, quotation and hedge word is counted before and after, and must match.
What makes AI-written text read machine-made: stock phrases, empty lead-ins, stiff wording, no contractions. It rewrites them the way a person would, and a meter shows how AI-like the text reads before and after.
Checked by
The same fact count as Text: a rewrite that would change a number, name or quotation is held back.
Never
It does not judge who wrote anything. AI-written text is welcome here; the point is to make it read naturally.
Answers to your questions in your documents' own sentences, each with its file and line; a partial answer names what is missing; sources that disagree are both shown. A follow-up with "he", "they" or "it" is read against the last answer; if that could mean more than one, it asks you to rephrase.
Checked by
A question word found in none of your documents blocks an answer.
The article inside a web page, without the menus, advertising and clutter; its reading level and layout problems (walls of text, overlong sentences); and, across several articles, the same thing given different numbers.
Checked by
The text fixes' fact check; reading level by the standard Flesch-Kincaid formula; contradictions shown with both sentences, neither chosen as right.
Never
It changes nothing on the original page or its server, and publishes nothing: the readable copy opens in a new tab on your own computer, for readability and clarity.
In your bank statement (the PDF your bank gives you, or a CSV): merchants charged many times, subscriptions billed more than once, money in and out of advance and pay-later services, payments returned unpaid, fees, and where the money went. Moves between your own accounts are kept apart.
Checked by
What it reads must add up to the statement's own totals, to the cent, or it says so.
Every figure below was measured on real projects and files, and every change was checked before it counted.
Area
Measured
How it was checked
3D models
2.0 MB proven removable across 98 of a finished browser game's 147 models, which were already compressed
Each model drawn before and after from 6 angles at two sizes: at most 0.31% of pixels moved, 47 dB or better. Animations played in both.
Source code
3 projects tested unused imports and code nothing calls, removed from copies of each
Each project's own test suite, before and after: 31, 28 and 10 tests passing, unchanged. No new errors when the pages load.
Pictures
Hidden details removed location, camera, software and editing history
Both versions decoded by Chrome and compared: every pixel identical.
Sound
Tags and padding removed loop points and cue markers kept
Both versions decoded: every sample in every channel identical.
News coverage
4 articles, 0 factions on one college football game; shared facts found in all of them
Loaded language 0.3 to 0.7 words per 100, against 4 to 7 in deliberately slanted test articles.
Your documents
Abstains when unsupported asked whether the conference fined the school, it answered "your documents do not say"
Answers are copied sentences with file and line; a key word absent from every document blocks an answer.
Honest scale: on the game's models the free saving was about 3%, because an earlier compression pass had already done the heavy work. The hardener reports what is there, including when there is little to gain.
What it is for
The problems it solves
Most of what slows software down, leaks private details or misleads a reader is not missing. It is extra, hidden or lopsided. The hardener finds it, fixes only what it can prove is safe, and hands the rest to a person.
Weight that does nothing
Code that nothing uses, model detail that never reaches the screen, filler words. Removed only when a check proves nothing visible or working changed.
Exposure you cannot see
Where a photo was taken, which software made a file, keys left in source code, dependencies with known flaws.
Fragile points
The one part everything rests on, and changes with no way back. Pointed out for a decision, never changed behind your back.
Slanted or machine-sounding text
Articles about the same event that split into sides, and what every side still agrees on. Writing that reads machine-made, rewritten so it reads like a person wrote it.
Answers you can check
Questions answered from your own documents, in their own words, with the file and line. When they do not say, it says so.
Check it yourself
The same input gives the same answer every time. Open your browser's network panel while you use a tool: the tool's code arrives the first time, and nothing of yours ever leaves.
How it works
Built level by level
Every tool runs in three steps: input, transform, output. You do not need to know what the transform does; this is the shape it follows, the seven steps of building anything. Each level rests on the one before it, and nothing is added that the level below cannot hold up.
0
Void
Nothing is assumed. Before you give it anything there is no answer and no guess, and where your input says nothing, the result says so.
1
Point
Your input is taken apart into single units, each with its own identity: a sentence, a word, a claim, a block of a picture, a piece of code.
2
Line
Each unit is related to its immediate neighbours: the sentence it leans on, the word beside it, the code that uses it. A unit knows only its neighbours, and what is learned passes from neighbour to neighbour.
3
Plane
One purpose, one table: every unit laid out for a single question. What is extra? What is exposed? What does this sentence rest on?
4
Volume
Tables are stacked into levels: several files, several articles, several kinds of weakness, each measured apart, so a saving in one place never hides a cost in another.
5
Depth
Down through the levels: what rests on what, from the surface to the bottom, so the cause beneath a problem is found, not only the problem on top.
6
Wholeness
It comes together as one result you can use: the clean copy, the answer with its source, the verdict. That result is a new point. Carry it into the next tool and the build starts again.
The Schwarz Diamond is the shape behind it: the least material for the most surface, and every point held up by the lines beneath it.
Without AI
Measure, do not predict
A language model
predicts the likeliest words
can be confidently wrong
can answer differently each time
usually sends your files to a server
The hardener
counts, compares and checks
makes a change only when a check proves it safe; otherwise it holds back and says why
gives the same answer every time
runs on your computer; nothing leaves it
Limits
What it does not do
It does not write. It cannot compose new sentences or new code; it removes, rearranges and quotes.
It matches what is written. Two articles saying the same thing in different words can be missed, and so can an answer phrased differently from the question.
It measures agreement, not truth. A consensus is what the sources share; they can all be wrong together.
The humanizer's meter is a guide, not a verdict. It shows which habits make text read machine-made; it says nothing about who wrote it, and nothing here puts down text because AI helped write it.
It leaves judgment to people. Anything it cannot prove safe is reported, not changed.
Schwarz Diamond