Hallucination Checker
A private, offline browser tool that identifies claims requiring verification and carries the user through a disciplined human-review workflow.
AI mistakes often arrive disguised as good writing.
A polished answer creates the feeling that the thinking is finished. The dangerous claims are not always outrageous. They are often a plausible number, a source that sounds real, a date that looks exact, or an absolute statement hiding inside a smooth paragraph.
The obvious solution was dishonest
A universal "truth detector" cannot reliably judge arbitrary text. Building one would recreate the overconfidence the tool is meant to prevent.
The real need was operational
Users needed help noticing what to check, challenging the AI, recording the review, and remembering that unchecked text remains unchecked.
You shouldn't have to trust one AI to check another.
No account. No upload. No hidden model. No claim that the software knows what is true.
Every feature reinforces human accountability.
Rule-based highlighting
Sources, research language, numbers, absolutes, names, and dates are surfaced because they commonly carry factual risk.
Checklist, not verdict
The tool says "check this," never "this is false." Silence is not presented as approval.
Challenge prompts
Users receive precise language for asking the AI to provide evidence, isolate corrections, and avoid rewriting unrelated material.
Review record
Printing and copied checklists turn a moment of caution into a repeatable workplace control.
What the first round of real use changed.
Working example added
The tool needed to teach itself before the user brought a document.
Example embedded directly
A preview environment did not reliably run the JavaScript, so the example was made visible in the HTML before scripting enhanced it.
Mobile contrast failure caught
A tester discovered that desktop text colors were being placed on a dark mobile background. The checklist background, type, tap targets, and spacing were corrected.
Factual claim corrected
The tool's own legal example was tightened so that its credibility matched the standard it asks of users.
This is stage three of three.
Catching a fabricated claim is the last resort. The better work happens before it is ever written.
Stage one: the prompt
Shape the request so the model has fewer openings to invent. Most of what a checker finds should never have been generated in the first place.
Stage two: the voice
Keep a draft sounding like the person who will put their name on it, and catch the changes a rewrite makes that nobody asked for.
Stage three: verification
This tool. It catches what survives the first two, and it is the only stage that assumes something already went wrong.
Stages one and two are designed, not built. The Hallucination Checker was built first because it is the stage where a mistake has already reached a real document, and because a control nobody can inspect is worth less than one they can.
This is not just a coding sample.
The artifact demonstrates problem definition, product judgment, honest limitation-setting, privacy-by-design, UX writing, workflow design, rapid iteration, and the ability to turn operational experience into a usable control.
Free version
Identifies what needs checking and teaches the behavior privately, in the browser.
Organizational version
Could be customized around a company's risks, evidence rules, workflows, and training without turning the free tool into a crippled demo.