AI but make it real

Three things you can break in your browser right now

Amelia Saxton

Anyone can show you a screenshot of a demo that worked once. These three run in your tab, with no server, no account, no API key, and no network call at all. Open the developer tools and watch nothing leave.

Each one is built around the same idea: the interesting engineering in AI is not getting the output, it is proving the output. So each demo hands you the controls and invites you to make it fail.

The demos

Roughly two minutes each. Every one of them has a button that shows you the thing going wrong, because a system that has never been shown failing has not been shown at all.

DEMO 01

A policy gate that fails closed

Agent governance

Multi-agent systems get approved on what they can do and killed on what they cannot prove they refused. Compose a message between agents, watch the governor evaluate it before it lands, and read the refusal with its rule name and its receipt.

Eight preset attacks, including an agent that does not exist and a policy file that failed to load.
An append-only decision ledger, because "we blocked that" needs an artifact behind it.
The last preset is a one character typo that shipped in my own code, inverted the only guarantee the component made, and passed review.

Open demo 01 → DEMO 02

Trace every number back to the source

Verification and hallucination control

The expensive failure is not a refusal. It is a fluent paragraph with one number in it that came from nowhere, pasted into a board deck by someone who had no way to check. This traces every figure in a generated summary back to the data it was written from.

Deterministic. No model checks the model, because that just moves the problem.
Catches invented figures, and the rounding drift that survives a skim because the sentence reads perfectly.
Has a panel titled "here is what breaks the checker", with three buttons that make the tool itself return the wrong answer. Every other button on these pages shows the subject failing. Those three show my own work failing, which is the only failure that would cost you anything.

Open demo 02 → DEMO 03

Count the defects before you sell the fix

Data repair at catalog scale

A messy product catalog, audited product by product, bucketed by named defect, reconciled to the row count, with the worst single case quoted in full. Then the scriptable half is repaired in front of you and the CSV comes back down.

Paste your own CSV. It never leaves your browser.
Rows in equals rows out plus named drops, each with a reason. No silent truncation.
Refuses to write product descriptions, and counts them as open work instead. Inventing content about a food product is the one thing a tool like this must never do.

Open demo 03 →

How to read these

All three are single HTML files. No build step, no dependencies, no framework, no analytics, no cookies, no fonts pulled from anyone's CDN. Save one to your desktop, turn off your wifi, and it still works. That is a deliberate constraint and it is the same one I apply to client work whenever the job allows it, because a deliverable that breaks when it leaves the folder it was built in is not finished.

All data on these pages is synthetic and labelled as synthetic. No client data, no employer material, no personal information. The catalog in demo 03 is invented, because a prospective client's product data belongs to them and does not belong in my portfolio.

Each demo states its own limits on the page. What it does not catch, what it cannot honestly do in a browser, and where a real deployment would need a file system or a pipeline behind it.

That invitation has already been taken up once, and it is the most useful thing that happened to these pages. Before publishing I handed the files to an adversarial reviewer with no context at all, no explanation of what they were for, and an instruction to refute rather than to check. It found nine defects, including a checker that passed a negative revenue figure as correct and a quote box that cut a sentence in half at a decimal point, in public, on page load. Six are fixed, each one commented in the source next to the case that produced it. Three were not defects but the real edge of the tool, so they are now buttons on demo 02.

My own end to end tests passed all of it. They tested whether the tools catch a fabrication and never tested whether a wrong answer could get through, and I wrote the expectations myself off the same assumptions that produced the code. So the invitation stands and it is genuine: if you find a tenth, tell me and I will put it on the page.