What was tested
The run used a fresh database built from every schema change the live site uses, with its triggers, access rules and grants, a real Ed25519 signing key held in a vault and the site's own server code. A client saved a decision, attached two evidence files, recorded the outcome and asked for a seal. A reviewer re-read the files and approved. North Star's owner sealed the record.
The public side was checked the way a stranger would check it: the receipt lookup, the verify page in Google Chrome on phone and desktop screens in English, Spanish and Portuguese, the receipt badge, the public register and the agent tool that checks receipts.
What it found and fixed
Before the fixes, 24 of 29 flow checks passed. Some verified decisions could never be sealed, because the receipt kept shorter text limits than the decision form. A seal could be refused after review, because the server and the database rounded measured values differently. A badge for a made-up receipt ID read as sealed. The verify page accepted a receipt signed by a key North Star never published.
Each problem now has a fix and a test that fails without it. Long decisions and a blank next action seal with their own words, measured values round the way the database rounds them, a badge for an unknown ID no longer says sealed, and the verify page requires the signing key to be on the published key list.
The result
After the fixes, 59 of 59 flow checks, 18 of 18 page loads in Chrome and 5 of 5 agent checks passed. A seal changed after signing reads as broken on the page and through the agent tool. A receipt signed with an unpublished key reads as broken, with its own explanation. The fixes went live on northstarintel.ai on November 17, 2026.
Check it yourself
Open the verify page. The production key signs a test record on the server, and your own browser checks that signature while you watch. A receipt ID entered on the same page is checked the same way: fingerprint, record, signature and the published key list.
What a test can and cannot show
This is a test of the software, not a client result. No client receipt had been published when it ran. A seal shows that a record has not changed since North Star signed it. It does not make the measured result true or prove what caused it.
