new encoder/src/demo/cold_start.py renders the result tab: three stat
cards, the two-panel figure (fraud caught vs label history; fraud
caught vs approval rate at 1% labels), a 7-column operating-point
table read from docs/figures/recall_at_approval.json, the "how it
works" paragraph for the SSL stage, and a synthetic-data scope block.
wired as a new tab in both app.py (standalone 4-tab) and
copilot_app_unified.py (7-tab unified). figure is base64-embedded so
the tab is self-contained html. *.png added to LFS tracking.
result on synthetic data (held-out test, n=20k, 3.7% base rate, 1%
labels = stand-in for a newly onboarded entity, 95% approval rate):
fraud caught 7.5% -> 36.0% (~4.8x), false-decline rate 4.9% -> 3.8%.
roc-auc 0.575 -> 0.814; pr-auc 0.048 -> 0.210. whole-book bonus at
full labels and 1% decline budget: +16 pts.
mechanism: stage-1 self-supervised next-feature prediction through the
frozen lfm2.5 base anchors the encoder's value tables before the
supervised fine-tune.…
