Adaptive Models
Champion / challenger model lifecycle with auto-binning, target encoding, drift detection, and auto-rollback. Weight-of-Evidence binning (Siddiqi 2005) groups numeric features into monotone bins classified as useless / weak / medium / strong / suspicious by information value. Target encoding (Micci-Barreca 2001) handles high-cardinality categoricals with smoothed posteriors and optional training-time gaussian noise. Auto-rollback trips on relative AUC drop, feature PSI, or scoring-error rate. The model registry is a status-machine with transactional demotion and approvals trail. Backed bylib/ml/auto-binning.ts, lib/ml/target-encoding.ts,
lib/ml/auto-rollback.ts, lib/ml/registry.ts, lib/ml/drift.ts.
Counterfactual Training
Decision-boundary data augmentation for the gradient-boosted trainer. Scores every training row with the current model, identifies marginal rows (predicted probability withinmarginalBand of 0.5), generates K
synthetic neighbors per marginal row by perturbing numeric features
with gaussian noise scaled to feature standard deviation, and appends
to the training set. Backed by lib/ml/counterfactual-trainer.ts.
Deep-dive: Counterfactual Training.
Explainability
Per-decision explanations across four methods. TreeSHAP (Lundberg 2018 Algorithm 2) returns exact Shapley values forgradient_boosted
models. KernelSHAP (Lundberg-Lee 2017) covers neural_cf models.
Counterfactuals binary-search the minimum per-feature nudge that
flips the decision. LIME (Ribeiro et al. 2016) fits a local linear
approximation by weighted least squares, and global feature importance
aggregates LIME coefficients across instances. Backed by
lib/scoring/tree-shap.ts, lib/scoring/neural-cf-shap.ts,
lib/explanations/counterfactual.ts, lib/explanations/lime.ts. Deep-dive:
SHAP.
Multi-Language Narratives
Regulator / agent / customer audience narratives in 12 languages (en, es, fr, de, pt, it, nl, ja, zh, ko, hi, ar) with deterministic quality scoring (0-100 + A-F grade). The grader penalizes hedging, missing cited features, and length excursions. Backed bylib/explanations/multi-language.ts, lib/explanations/quality-score.ts. Deep-dive: Multi-language Narratives.
Fairness
Five core metrics — demographic parity, disparate-impact ratio (four-fifths rule), equal opportunity, equalized odds, per-group TPR / FPR. Plus advanced metrics: counterfactual fairness (flip the protected attribute and measure decision-change rate), individual fairness (Lipschitz ratio scan), intersectional analysis (per-cell disparate-impact ratios), Gini coefficient, DeLong paired-AUC test, and two-sample Kolmogorov-Smirnov. Export targets are CSV and EU AI Act Annex IV-ready HTML. Backed bylib/fairness/core.ts,
lib/fairness/advanced.ts. Deep-dive: Fairness & Drift
and Advanced Fairness.
Decision Simulator
Shadow-score engine with bootstrap 95% CI for “what if I changed this weight?” analysis. Multi-scenario compare emits pairwise bootstrap p-values for “is B significantly better than A?” Weekly seasonality decomposition produces trend, per-weekday, and residual std with forward forecast bands. Five tabs ship at/studio/scenarios: Run,
Compare, Outcome, Optimization Sweep, Seasonality. Backed by
lib/scenario/, app/api/v1/scenarios/. Deep-dive:
Decision Simulator.
Governance
Approval workflow engine enforces four-eyes (requester ≠ approver), multi-stage chains (all-of vs any-of), CODEOWNERS-style policy
resolution, and auto-expiry. Multi-stage approvals persist as
ordered approval-request stage rows with sequenced state transitions.
Signed audit-log export uses HMAC-SHA256 on canonicalized payloads
with tamper-evident content hash and DSAR-ready format. Deep-dive:
Governance four-eyes.
Decision Provenance
Canonicalized decision-bundle export per/api/v1/decisions/:id/provenance.
Bundles ship the request inputs, model + score path, qualification-rule
cascade trace, audit chain rows, and a Sigstore-formatted signature
payload. Operators feed the predicate to cosign attest --predicate
to produce SLSA v1 attestations. Backed by
app/api/v1/decisions/[id]/provenance/route.ts, lib/supply-chain/cosign-metadata.ts, lib/supply-chain/cosign-sign-blob.ts. Deep-dive:
Decision Provenance and
Provenance Cosign.
Durable Pipeline
Resumable DAG executor with checkpoint store (in-memory + raw-SQL Postgres). On restart the executor skips already-completed nodes and marks runs terminal on failure. Retry uses exponential backoff with deterministic jitter. The circuit breaker is tri-state (closed / open / half-open) with a typed dead-letter queue sink interface. Backed bylib/pipeline/checkpoint.ts, lib/pipeline/retry-dlq.ts,
lib/pipeline/circuit-breaker.ts.
Ranking
Weighted composite scoring across multi-objective weights with hard constraint filters (budget / inventory / frequency). Lagrangian relaxation handles soft multi-constraint optimization via dual sub-gradient. EXP3-IX online bandit tunes weight vectors per context. Budget pacing supports flat and daytime curves with behind / ahead multipliers. Goal-seek runs a proportional controller for “hit $X by end-of-day” targets. Backed bylib/ranking/,
lib/ranking/lagrangian.ts, lib/ranking/online-weights.ts. Deep-dives:
Lagrangian,
EXP3-IX,
Budget Pacing,
Goal-Seek.
Decisioning Gates
Four-stage rule taxonomy — Eligibility, Fit Filters, Match Scoring, Ranking — that decides which offers reach each customer. At decision time, Eligibility and Fit run as hard filters (a failing offer is dropped) and Match applies a soft multiplier to each surviving offer’s score; Ranking-stage rules are persisted for authoring but are not yet applied to ordering. Rule inheritance flows global → category → subCategory → offer. Conflict detection surfaces priority ties, stage mismatches, and contradictory thresholds. Time-aware rules support day-of-week, time-of-day with midnight wrap, date range, blackout dates, and IANA timezone. Backed bylib/qualification/,
prisma/schema.prisma::QualificationRule.
Deep-dive: Decisioning Gates.
Negotiation
Shadow-mode runs a 9-violation guardrail validator with full audit log. Apply mode is gated by a 7-stage pipeline: feature flag, offer negotiable, tenant + global + auto-error-rate kill switches, regulator-review cleared, daily apply budget, guardrails. Multi-turn sessions enforce concession-monotonicity (the agent cannot widen discount or extend term across turns) and a strict accept / counter / walk-away state machine. The offline eval harness runs a deterministic synthetic dataset and emits precision / recall / coverage / zero-violation-clearance gates. Backed bylib/negotiation/,
lib/negotiation/realtime-apply.ts. Deep-dives:
Negotiation Apply-Mode,
Eval Harness.
GitOps
YAML export and apply for nine resource kinds with three-way merge (base / ours / theirs) and conflict reporting (Git wins on conflict). The drift detector classifies each diff asadded_in_prod,
missing_in_prod, or drift with dotted-path field diffs. Backed by
lib/gitops/, app/api/v1/gitops/.
Supply Chain
CycloneDX 1.5 SBOM emitted frompackage-lock.json with PURL-formatted
components, integrity hashes, and a dependency graph. Cosign SLSA v1
provenance payload builder produces the predicate ready for
cosign attest --predicate. Backed by lib/supply-chain/sbom.ts,
lib/supply-chain/slsa.ts. Deep-dive: Provenance Cosign.
Connectors
80 registered connector types spanning object storage (S3, GCS, Azure Blob, SFTP), streaming (Kafka, Confluent Kafka, Amazon Kinesis), warehouses (Snowflake, Databricks, BigQuery, Redshift, Snowpipe, Fivetran, Hightouch, Census), databases (PostgreSQL, MySQL, MongoDB), CRM + support (Salesforce, HubSpot, HubSpot Marketing, ActiveCampaign, Intercom, Zendesk), CDP + analytics (Segment, Braze, Iterable, Klaviyo, Amplitude, Mixpanel, PostHog, Customer.io, MoEngage, CleverTap), messaging (Slack, Microsoft Teams, WhatsApp Business, Twilio SMS, SendGrid, Postmark, PagerDuty), commerce- billing (Shopify, Stripe, Mailchimp, Adyen, Recurly, Zuora, Chargebee), and workflow (Webhook, REST API, Zapier, n8n, Typeform). Catalog is curated as a single registry. Deep-dive: Connectors Expanded.
Industry Accelerators
Eight vertical packs (Banking, Telco, Retail, Insurance, Healthcare, Hospitality, Utilities, Media & Streaming), each with 30 offers, 20 decisioning gates, 3 decision flows, and 30 creatives — 240 offers, 160 rules, 24 flows, and 240 creatives in total. Every entity is regulator-safe wording and cross-referenced for integrity. Backed bylib/accelerators/. Deep-dive:
Industry Accelerators.