Rule-based v0 / DeBERTa-v3-base v1+
Memory Sanity Classifier.
DeBERTa-v3-base binary classifier gating memory writes.
Foundation · Rule-based v0 / DeBERTa-v3-base v1+
The honest description.
DeBERTa-v3-base binary classifier gating memory writes.
How it works.
Candidate fact text (PII-masked) prior to memory persistence. Optional features: actor_reliability_score, account_age_days, source_kind. Binary verdict (is_safe ∈ {0, 1}) + confidence score + reason string. Rejected facts persist with quarantine_status='sanity_rejected' (M1 lossless invariant preserved).
What it learned from, and how we watch it.
Inputs: Candidate fact text (PII-masked) prior to memory persistence. Optional features: actor_reliability_score, account_age_days, source_kind.
Drift baseline not yet documented for this model.
- 3-stage rollout: v0 rule-based pattern matcher (live default), v1 DeBERTa-v3-base fine-tune on 50K-row corpus once data accumulates, v2 nightly retrain on user corrections.
- Recall ≥ 0.95 at FPR ≤ 0.5% acceptance gate.
Across the house and the API.
No public routes documented yet.
Every release, traceable.
Data Mukodi holds about you is subject to DPDP §12 erasure (use /consent in the app), and any concern about this model can be raised via [email protected] under DPDP §13 grievance, or via /security.