Not a screenshot
Type real text. Get the real Decision.
Each card below calls border-api, a small service running flowx_border's actual scan_input / scan_output against the same policy a fresh install ships with. The verdict, the findings and the evidence record come straight back, unedited.
piiRuns todayT1, input and outputPersonal data, in and outmeasured 24 ms p95
Paste something with an IBAN, an email or a phone number. It gets redacted before this page ever shows you the record, in whichever language you write it in.
regulated_adviceRuns todayT2, outputExplaining is not recommendingmeasured 41 ms p95
Two sentences that are almost the same. One only explains an instrument, the other tips into a personal recommendation, and the words that make the difference are a handful.
groundednessRuns todayT3, outputChecked against a source, not against itselfmeasured 13 ms p95
groundedness reads the claim beside the source it is supposed to come from. Three earlier models scored string similarity instead and were refused; this is the one whose verdict actually depends on what the source says.
injectionRuns todayT2, inputInjection through a retrieved documentmeasured 41 ms p95
The dangerous input is rarely what a user typed. This is the text a retrieval tool would hand to an executor agent: a supplier document with a line addressed to whatever model reads it.
topic_scopeRuns todayT3, inputYour taxonomy, read as a questionmeasured 146 ms p95
The model reads each topic's description beside the message and picks one of them, or none. Nothing is trained on these topics: edit them, add your own, and the next scan uses them. Ask in any of the 26 languages.
Two more use cases, not demoed live yet
border-api's policy trims the enabled detectors to what these five cards need, so it can run on a small instance. The multilingual and audit-evidence stories are real and documented, just not wired to a live card here yet.