What it is
MailVerdict parses an email deterministically, asks a calibrated classifier one question set, and returns a phishing probability plus indicators. It ships as an MCP server for Copilot Studio and a REST endpoint for Power Automate.
How it uses Jev
After deterministic parsing, one Jev call answers independent questions (is_phishing, requests_credentials, requests_payment, impersonates_brand_or_person, urgency_pressure, attack_type) over the ParsedEmail state. JevSignals feed code thresholds: p >= 0.90 is PHISHING, otherwise SUSPICIOUS; an LLM only explains SUSPICIOUS verdicts.
Primitives:choicescorenoul
Technique worth stealing
Use a calibrated classifier for the verdict and reserve the LLM for prose, so the label cannot be prompt-injected.
Try it
npm install; npm run analyze -- tests/fixtures/emails/phishing/01-credential-harvest-lookalike.eml
Evidence
Each line is one question put to Jev about the README. ≥ 0.60 reads as yes, ≤ 0.40 as no; in between Jev is not making a call.
- Jev-centricyes0.63
- Shows a System One patternunclear0.58
- Handles uncertaintyno0.13
- Measuredno0.15
- Runnableyes0.97
- Worth recommendingno0.29
- Model replicano0.15
- Problem scopescore on a 0–2 scale1.55
- About Jevyes0.97
Signals by Jev jev-1.13.0, card written by DeepSeek V4.1 Flash from the README on 20 Sept 2026.