I stopped asking "is this email phishing?" and started asking eight narrow
questions. Every number in this video is real model output: the 0.92 on
asks_credentials, the 0.990 verdict, the 0.053 when the same pipeline waves
a release-notes email through. Local model, CPU, $0 per 1,000 emails, and
every verdict names its reasons.

Repo: github.com/Gjusev/laya-phishield

---

Short variant (X/Discord):

One big question is the wrong question. Eight narrow ones, one verdict,
every reason named: asks_credentials 0.92, urgency 0.86, link risk 0.91
-> 0.990 PHISHING. Same machine waves a legit email through at 0.053. All
real output, all local, $0/1k emails.
