An overclaim caught by the thing it overclaimed
I built a small tool that checks your work before you say it out loud, then made a video to explain it. Before posting the video I held it to the tool's own standard. It failed, in exactly the way the tool exists to catch.
A check measures the distance between claim and source and stops. A validator that repairs its own findings is not an audit; it is an accomplice with a timestamp. Tao of ARCS, 273
The tool is called Check it before you say it, and it lives at honest.arcs.care. It walks a piece of work through five steps on the way from an idea to a message: find the facts, check the facts, poke holes, make it short, write the note. At each step a set of fixed rules looks at what you made and gives it a stamp, go or stop.
The rules are plain on purpose. Every fact you mark has to name a source. Big claims have to name two. Every fact has to carry your verdict. At least one review has to disagree, because a test nothing could fail is not a test. The short version cannot contain a marked fact the earlier steps never saw. The note has to still say draft, because a person sends it, not the tool. No model decides a stamp. The checks run in your browser, and nothing you paste is uploaded.
That is the whole product. It does not know whether anything you wrote is true. It knows whether you did the steps that would let you find out.
Elaboration is not evidence. The more beautiful the monument to a conclusion, the harder it needs the check; the pleasure is not proof, it is the thing proof has to survive. Tao of ARCS, 269
Then I made a sixty-five second explainer, with calm captions, a stamp slamming down on a sneaked-in claim, and a walkthrough of the page. It looked right. It read right. I was one click from calling it done.
The house rule for anything outward-facing is that it passes a check before it leaves, so I pointed a claims check at the video. Every caption and every line of on-screen text went through two independent readers. One compared the words to the code. The other read them the way a stranger scrolling past on a phone would, and wrote down what that stranger would come away believing, then checked the belief against the code.
Of one hundred eleven lines, eighty-four were flagged by at least one reader. An editor pass rewrote fifty of them.
The explainer for a tool that stops claims nobody checked was full of claims nobody checked.
The failures were not scattered. They were one failure, repeated. The video kept saying the tool does the work. The code only checks work a person did.
DRAFT. It never writes, and it never sends.
ClaimedA five-step check for anything you are about to tell someone.
The codeReads facts you have marked, or sentences with numbers pulled from a notebook. Plain prose gets no check at all.
ClaimedA meter with a limit line at eighty percent.
The codeThe real rule counts agreeing and disagreeing words, and reads the work as too agreeable when agreements outnumber the disagreements plus one by more than four to one. The eighty percent line was invented for the picture.
None of these were lies I meant to tell. Each one was the natural sentence, the one that makes a product sound helpful. That is exactly why they got through: a sentence that sounds right is the kind nobody stops to check.
an unmeasurable claim is noise; one specific believable claim outweighs an impressive vague one. Tao of ARCS, 143
The same drift had crept into the product page, in a quieter form. A rule label said every fact says where it came from. The rule only checks the facts you marked. Another said someone actually disagreed, when the rule counts reviews. They became every marked fact names a source, and at least one review disagrees. Small changes, but the video had been quoting the page, so the page was where the drift started.
A separate panel of reviewers went over the design at the same time. It measured the old captions against how people actually watch on a phone with the sound off: fourteen of seventeen captions ran faster than three words a second, and most of the text would have shrunk to a few pixels in a feed. The story needed to be truer. It also needed to be readable.
show the repair seam. a thing mended and shown honest outranks a thing that pretends it never broke. Tao of ARCS, 184
The rewrite came down to one sentence: you do it, rules check it. Every caption now says what you do and what the step checks, and the pictures follow the same split. Only the checks move on screen. Your work arrives whole, as grey bars, because it is yours and the tool never shows it anywhere.
The best idea in the re-cut came from asking what belief updating should look like. When the sneaked-in claim gets stopped, the tool does not delete it. You strike it through, and the struck line stays on the record, dimmed, while the stamp beside it turns from stop to go. The person changes their mind. The rules only stamp. The break stays visible instead of quietly disappearing, which is the point of keeping a record at all.
The strike stays on the record, and so do the fifty lines I had to rewrite. A tool that checks claims should show its own breaks first.
never consume a claim that has lost a coordinate. trust it only to the fidelity it can still prove. Tao of ARCS, 138
A day later, proposing new sections for the product page, I wrote a heading that said why you can trust it. It was the same overclaim in a new costume. Trust is not a property of a tool. It depends on what you lean on it for, and this one gives opposite answers depending on the job.
Lean on it to tell you whether every marked fact names a source, and you can lean all the way: the same input always gets the same stamp. Lean on it to tell you whether a fact is true, and it will let you fall, because it never looks. Deterministic is not the same as objective, either. The rules are mechanical, but a person chose them. The line that calls agreement too thick past four to one is a threshold I picked.
So the promise is not that you can trust it. The promise is that you do not have to. Every check is visible, it runs in your browser, and it shows its work, so you can check the checker. And the trust that matters was never yours in the tool anyway. It is your reader's in you. The tool does not make you trustworthy. It makes your claims traceable, so a reader can see where each one came from.
Not a tool you trust. A tool you can check, so the people you write to can check you.
a named gap is honest; a hidden one is a lie. when you cannot check the box, state the residual risk out loud. Tao of ARCS, 137
The tool is live. The rules are fixed and tested: every one is proven to fail on a bad example and pass on a good one, and the browser version is checked against the reference version on every case. What is not proven is judgment. Each step also has a short scorecard for a person to fill in, with pass floors that I set and nobody has yet calibrated against real outcomes. The re-cut video is posted on this page; whether it explains the tool to anyone who was not in the room is unmeasured. Those parts stay marked unproven until something outside my own head says otherwise.
an account is earned, not demanded. ask after the tool has shown its worth, and the asking answers itself. Tao of ARCS, 195
One more claim, marked the way this piece asks every claim to be marked. The engineering here is where I have the most reps. Front-end design and visual craft are where I am still putting them in: the video above was built frame by frame from a web page and a render script, not in a studio, and it shows its seams.
If you make product marketing with Runway, Stable Diffusion, or whatever arrives next, and you want to help tools like this one look as good as they work, I would like to build with you. zachharris.xyz is the way in.