Overview: Open in-dashboard doc Reference doc: memory/reference_link_ranker_c2.md (chat context) Plan: plans/ml-tinygrad-knowledge-base.md Code: /var/lib/jarvis/knowledge/ml/integration/c2-link-ranker/ Cron: link-ranker-score (hourly, sentinel) Mode: shadow · routing untouched
Scored total: Scored last 7d: Ground truth: AUC baseline: 0.863 Shadow validation: acc 76.5% / recall 93.6% / precision 51.8%
Score Distribution — last 7 days
Loading...
Calibration — predicted p vs fraction approved
Loading...
Throughput — items scored per hour (last 24h)
Loading...
Drift — rolling 24h vs baseline
Loading...

Items to Review — approve / reject

What this does: Approve / Reject are ground-truth labels for the ranker — nothing else. Approve = "I'd want to see/act on this content" (positive label). Reject = "I'd skip this" (negative label). The content itself is unchanged — no routing change, no Telegram alert, no follow-up triggered. These labels feed the calibration / drift panels above and become training data for the next retrain.
p Label Title / Description Status Action
Loading…