caelum0x 295f975d5a feat: p7 ML risk-engine v2 (learned probability-of-default model)
RiskEngineV2 adds a logistic-regression PD model over a normalized agent
feature vector (reputation, accuracy, dispute rate, experience, revenue,
stake, category risk). train() fits weights by deterministic, L2-regularized
batch gradient descent on realized default/repay outcomes; score() returns
pd + ml_score + rules_score + a convex blended_score, so v2 augments the v1
rules policy rather than replacing it. A domain-calibrated prior makes it
useful before any training data.

5 new tests, full suite 212 green.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-06-17 17:39:55 +03:00

2.2 KiB
Raw Permalink Blame History

p7 — ML risk-engine v2

Learn creditworthiness from realized outcomes, not just rules. Add a trained probability-of-default model on top of the hand-tuned v1 policy, and blend the two so credit decisions improve as the protocol accumulates repayment/default history.

Why

The v1 risk policy (risk_policy.ts) is a transparent rules engine — great for launch and auditability, but it can't learn. As Cred402 accrues realized credit outcomes (who repaid, who defaulted), the cheapest accuracy gain is a model fit to that history. Keeping the rules score alongside the learned score preserves explainability and gives a safe fallback.

Scope

  • Feature model. A stable, normalized feature vector per agent: reputation, accuracy, dispute rate, experience (job count), recent revenue, stake, and service-category risk.
  • Learned PD model. A logistic-regression probability-of-default classifier, fit by batch gradient descent on labeled outcomes. Deterministic and dependency- free (no external ML runtime), so it is reproducible and the weights are inspectable/auditable.
  • Blend, don't replace. blended_score = α·ml_score + (1−α)·rules_score, so v2 augments v1 rather than overriding it. A hand-calibrated prior makes the model useful before any training data exists.

Deliverables (shipped)

  • lib/services/risk_engine_v2.ts — RiskEngineV2:
    • extractFeatures(agent) → normalized [0,1] FeatureVector.
    • predictPd(features) → probability of default.
    • train(samples, opts) → fits weights by gradient descent on log-loss (deterministic; L2-regularized).
    • score(agentId) → { pd, ml_score, rules_score, blended_score, risk_band }.
    • PRIOR_WEIGHTS — domain-calibrated prior; weights() exposes the model.
  • test/risk_engine_v2.test.ts — prior ranking, training reduces log-loss + learns signals, blended score, normalization, determinism.

Success metric

The blended score out-predicts the rules-only score on held-out realized outcomes, while every decision remains explainable (rules score + inspectable model weights).

Later (p8+)

zk/light-client receipt proofs (trust-ladder Stage 4); enterprise RealFi GA; new service verticals (compute/inference/data markets).