Interactive demo · you act yourself · everything runs in your browser

The claim nobody can reconstruct

Twelve minutes. You are the underwriter. You make the decisions, you run the checks — and at the end you try to cheat us yourself.

Act 1 · Tuesday morning

An application is on your desk

A vendor of automated claims screening wants to insure a performance guarantee. Its AI model is meant to detect manipulated claim photos — the promised detection rate is 94 percent. If it falls below that, the vendor owes its customers a contractual penalty. Your company is asked to carry that penalty.

You turn to the question your questionnaire asks at this point: “How is the model’s performance monitored while the end customer uses it?”

The answer is a PDF. Three pages. Written by the applicant itself, with figures from the past six months. You read: 94.3 % detection rate, measured internally.

Act 2 · Seven months later

The call

The applicant reports a guarantee claim. A major customer has invoked the penalty: for eleven weeks the model waved manipulated photos through. The detection rate allegedly fell to 61 percent. Demand: 2.4 million euros.

You ask the three questions that matter now.

Question 1 Did the rate really fall? → Evidence: an analysis by the applicant. Produced last week. From the same systems that produced the 94.3 %. Question 2 Since when? → Unknown. Log files exist, but their timestamps come from the applicant’s server. They are not signed. Question 3 Was it the model, or the supplier’s base model? → Cannot be determined. Which model version ran when is nowhere recorded in a tamper-evident way.

Your claims handler puts it like this: “We cannot prove the loss happened — and we cannot prove it did not.”

Act 3 · Same situation, one condition changed

This time the system kept a record

Rewind. When you wrote the risk, the applicant accepted one condition: its system issues a receipt about itself on every call. No image content, no prompt — only checksums, chained one to the next. (In the product the receipts are also signed and time-stamped; this demo shows the chain.)

Press the button. The chain is built now, in your browser, with real SHA-256.

No chain yet. Press the button above.

Act 4 · Now you are the attacker

Try to fool us

This is the part that matters. You are not supposed to believe us — you are supposed to try. You have full write access to the chain. So would the applicant.

Delete the drop. Backdate it. Slip in a receipt after the fact. Then check.

Build the chain in Act 3 first.

Act 5 · Back at your desk

What changes for you

without receipts with receipts Is the loss real? a claim recomputable Since when? unknown to the receipt Whose fault? inseparable model version on record Editable afterwards? yes, without trace yes, but visibly Premium calculable? only as a flat rate by behaviour

The receipt does not undo a loss. It makes it attributable — and attribution is exactly what your pricing lacks today.

And it does so without a single image, prompt or customer file leaving your policyholder’s premises. It discloses no trade secrets. It discloses checksums.

What this demo does not show — and what we do not claim:
That a strong chain correlates with low losses. The test procedure for that is built and tested in both directions (it finds an existing correlation, explicitly reports none when there is none, refuses any statement below twenty data pairs — and has a fourth outcome that few metrics allow about themselves: finding against the metric, when the relation is significant but points the wrong way) — but real claims portfolios are missing. That statement is outstanding, and we will make it only once it is proven.

Also open: the qualified timestamp of a trust service provider. Anchoring via Merkle roots is built; as long as the connection is missing, every anchor honestly reports “not qualified” instead of claiming otherwise. And setting the field to “qualified” by hand does not get through: an anchor that claims an attestation without carrying a complete token is rejected on verification. If the device comes online later, the stamp is obtained then — and the late one counts less than an immediate one, because the gap in between remains unattested. That gap is stored in the anchor as a number, not as a reassuring sentence.