Phase 2: answers to the Phase 1 judge feedback (11 workstreams)

[0xdevabir]

Phase 2 improvements in response to the Phase 1 judge feedback. There are 11 workstreams, each with a plan in docs/plans/NN-*.md. All numbers come from runs on synthetic data unless marked otherwise.

#	Workstream	Measured result
01	External validation (make external)	Bootstrap 95% CIs: recall 86.9% [82.7–91.0], FPR 0.52% [0.47–0.58]. Public data: PaySim PR-AUC 0.468 (logistic regression 0.323), BAF 54.8% recall at 5% FPR
02	Bangladesh evidence	docs/BANGLADESH_CONTEXT.md with sources; calibrated simulator profile; survey and interview instruments
03	Business case	docs/BUSINESS_CASE.md and an interactive P&L: ৳88.6 lakh/month kept from scammers at upay scale, ৳5.27 per ৳1
04	Label realism (make label-realism)	Reported-only labels cost 8.5 recall points; mule label propagation and analyst verdicts recover about half
05	Adaptive adversaries (make adversary)	Frozen model 96.7% → 79.6%; weekly retraining 98.0%
06	Horizontal scaling (make worker, make loadtest)	Redis consumer group: 2.8× throughput at 4 workers, p99 1.0 s → 0.29 s, 0 mismatches against offline evaluation; K8s and KEDA manifests (validated with kubeconform, never run on a cluster)
07	Bank ingest and security	Signed ISO 20022 webhook, key rotation, TLS proxy, STRIDE threat model (docs/INGEST.md, docs/SECURITY.md)
08	Customer experience	Bangla/English interventions, customer appeals with SLAs, analyst appeals queue, phone e2e
09	Cross-provider consortium	Privacy-preserving mule listings (OPRF tokens, signed bundles, hash-chained audit): +4–5 mules caught per 100 at 25% SIM sharing, at the same alert rate; 0% control shows zero gain
10	Bangla/Banglish message model (make intel-compare)	External smishing corpus F1 0.66 → 0.97; unseen-script macro-F1 0.81; links a flagged SMS to the payment it triggers
11	Fairness mitigation (make mitigation)	Policy v3: young-sender false-alert ratio 5.90× → 1.14×, recall 81.73 → 81.58%. v2 remains the default until sign-off

Integration notes

• Migrations are chained 0003 → 0004_appeals → 0005 (pending decisions).
• The message-to-payment link is computed in the ordered scoring step and applied in both the live and the deferred (replay) decision paths.
• docs/MODEL_CARD.md §13–§17 cover the new evaluations; Reproduce is now §18.

Verification

• Backend: 364 passed, 0 skipped, including the Postgres and Redis platform tests. Ruff check and format are clean.
• Frontend: tsc, eslint and pnpm build pass.

Known limits (also documented)

• The data is synthetic, and no public dataset contains MFS scams with mule chains: the external results support the method, not the 86.9% figure itself.
• Scaling stops at a few workers because every worker holds the full feature state.
• The message link lives in process memory.
• The consortium serves a simulated state.
• The Bangla texts have not had professional translation review.
• The PDF and PPTX in docs/ are unchanged.