{
  "claim": "Every per-email row I could read off the recording, transcribed by hand. Twenty of the 100 decisions are legible at some point in the 12.8 seconds: the Jev confidence, the route the gate chose, the final verdict, and whether the app scored it against the DIFrauD label. Five took the direct route and read 97-100%; fifteen were escalated and read 10-94%. All three legible misses are escalated rows that Kimi K3 reviewed and still got wrong.",
  "captured": "2026-09-18",
  "method": "frames extracted at 4 fps from the 2920x2160 variant of the post's video (ffmpeg -vf fps=4), rows read visually from the sharp frames; the subject column is truncated by the app itself",
  "source": "https://x.com/nutlope/status/2100614659690713543",
  "columns": [
    { "key": "id", "label": "email" },
    { "key": "conf", "label": "jev confidence", "align": "right" },
    { "key": "route", "label": "route" },
    { "key": "verdict", "label": "verdict" },
    { "key": "scored", "label": "vs label" }
  ],
  "rows": [
    { "id": "MSG-002", "conf": "10%", "route": "Kimi K3", "verdict": "Legitimate", "scored": "correct" },
    { "id": "MSG-012", "conf": "78%", "route": "Kimi K3", "verdict": "Legitimate", "scored": "MISS" },
    { "id": "MSG-039", "conf": "75%", "route": "Kimi K3", "verdict": "Legitimate", "scored": "MISS" },
    { "id": "MSG-052", "conf": "87%", "route": "Kimi K3", "verdict": "Fraud", "scored": "correct" },
    { "id": "MSG-053", "conf": "89%", "route": "Kimi K3", "verdict": "Fraud", "scored": "correct" },
    { "id": "MSG-057", "conf": "93%", "route": "Kimi K3", "verdict": "Legitimate", "scored": "correct" },
    { "id": "MSG-061", "conf": "94%", "route": "Kimi K3", "verdict": "Legitimate", "scored": "correct" },
    { "id": "MSG-068", "conf": "87%", "route": "Kimi K3", "verdict": "Legitimate", "scored": "correct" },
    { "id": "MSG-069", "conf": "31%", "route": "Kimi K3", "verdict": "Fraud", "scored": "correct" },
    { "id": "MSG-076", "conf": "80%", "route": "Kimi K3", "verdict": "Legitimate", "scored": "correct" },
    { "id": "MSG-083", "conf": "13%", "route": "Kimi K3", "verdict": "Fraud", "scored": "correct" },
    { "id": "MSG-086", "conf": "75%", "route": "Kimi K3", "verdict": "Legitimate", "scored": "correct" },
    { "id": "MSG-088", "conf": "89%", "route": "Kimi K3", "verdict": "Fraud", "scored": "correct" },
    { "id": "MSG-091", "conf": "80%", "route": "Kimi K3", "verdict": "Legitimate", "scored": "MISS" },
    { "id": "MSG-049", "conf": "78%", "route": "Kimi K3", "verdict": "Fraud", "scored": "not legible" },
    { "id": "MSG-081", "conf": "99%", "route": "Jev direct", "verdict": "Legitimate", "scored": "correct" },
    { "id": "MSG-082", "conf": "99%", "route": "Jev direct", "verdict": "Legitimate", "scored": "correct" },
    { "id": "MSG-098", "conf": "99%", "route": "Jev direct", "verdict": "Legitimate", "scored": "correct" },
    { "id": "MSG-099", "conf": "97%", "route": "Jev direct", "verdict": "Fraud", "scored": "correct" },
    { "id": "MSG-100", "conf": "100%", "route": "Jev direct", "verdict": "Fraud", "scored": "correct" }
  ],
  "note": "Twenty rows is a convenience sample, not a random one: it is whatever happened to be on screen while the page was scrolled. It cannot be used to estimate the confidence distribution over all 100 emails. What it does establish is the range on each side of the gate, and that at least three of the run's four residual errors sat in the escalated bucket rather than the accepted one."
}