A technical specification for how Arbiter explains its verdicts. Decision rationale, contributing sub-agents, signal-level evidence, and counterfactual on every output.
A risk decision is only useful if you can defend it. Examiners ask why. Customers dispute the answer. Compliance officers need to explain the verdict to the analyst who has to act on it. A risk score with no rationale is technically a number; operationally it is a liability.
Arbiter is built around the principle that explainability is not a separate report you generate after the fact. It is part of the decision itself. Every Arbiter verdict returns the rationale, the contributing sub-agents, the signal-level evidence, and the counterfactual that would have changed the outcome.
Every Arbiter decision returns a structured response with five sections: verdict (allow / deny / challenge / escalate), contributing sub-agents (ranked by contribution weight), signal-level evidence (the specific signals that fired within each sub-agent), counterfactual (what would have changed the verdict), and examination documentation (auto-generated to OCC SR 11-7 format).
The top-line outcome plus verdict_confidence and verdict_id for traceability.
Each sub-agent reports its contribution_weight and sub_score. Behavioral analytics, transaction risk, adverse media, sanctions, and PEP all feed into the synthesized verdict with calibrated weights.
The specific signals that fired within each sub-agent. This is the most granular explainability layer and is what examiners ask for. Each signal records its value, baseline, z-score where applicable, and the timestamp it fired.
What would have changed the verdict. The counterfactual is computed by perturbing each input and re-running the decision. The response reports the minimum changes required to flip the verdict.
Auto-generated documentation aligned to OCC SR 11-7 model risk management format. Includes model_id, model_release_date, validation_status, fairness_test_status, and a doc_export_url for examination submission.
Layer 1: SKILL.md training (static). Each of the six sub-agents (Sanctions, PEP, Adverse Media, Re-verification, Transaction Risk, Behavioral Analytics) ships with a SKILL.md file defining what signals it monitors, how it scores them, what verdict thresholds it uses, and what regulatory frameworks it operates under.
Layer 2: MCP tool orchestration (live). At runtime, Arbiter calls each relevant sub-agent as an MCP tool. The sub-agent returns its sub-score and the signals that fired. Arbiter synthesizes the sub-scores using a calibrated ensemble.
Sub-agent parallel scoring runs in under 100ms. Verdict synthesis under 10ms. Signal-level evidence assembly under 15ms. Counterfactual computation under 20ms. Examination doc generation under 5ms. Total p99: under 150ms. The counterfactual is precomputed for the most common perturbations.
The auto-generated examination doc covers every requirement in OCC SR 11-7: model identification, conceptual soundness, ongoing monitoring, outcomes analysis, fairness testing, process verification, and counterfactual. Exportable in PDF for examination submission, CSV for internal review, and JSON for programmatic integration with model governance platforms (Workiva, Archer, MetricStream).
For regulated deployments, the explainability layer integrates with the customer's model risk management practice. Arbiter ships with validation reports for each model release, disparate impact analysis aligned to OCC SR 11-7 and the EU AI Act, performance monitoring dashboards, drift detection alerts, and escalation paths for model degradation. Documentation is delivered as a quarterly model risk management package signed by the deepidv chief risk officer.
Why does explainability matter for risk scoring? Risk decisions need to be defensible to examiners, customers, and regulators. A score with no rationale is a liability. Explainability lets the customer-facing analyst understand why the verdict came back the way it did.
How is the counterfactual computed? Arbiter perturbs each input signal and re-runs the decision. The counterfactual reports the minimum set of input changes that would have changed the verdict. Most counterfactuals precompute in under 20ms because the perturbations are bounded.
Can I export Arbiter decisions for my model risk management committee? Yes. Every decision exports as a PDF, CSV, or JSON. Quarterly model risk management packages include validation reports, fairness testing, performance monitoring, and decision-level audit trails.
What about adversarial examples? Arbiter's counterfactual layer surfaces decisions that change easily under small perturbations. Decisions that change with minimal input changes are flagged for human review.
How does Arbiter handle disagreement among sub-agents? The contribution weights come from the calibrated ensemble. When sub-agents disagree, the weighting reflects the calibration on historical decisions. Disagreement between sub-agents at the upper bounds of confidence triggers a 'challenge' verdict by default.
Can I customize the explainability layer for my regulator? Yes. Customers in jurisdictions with specific examination requirements (FFIEC US, FCA UK, BaFin Germany, MAS Singapore, AUSTRAC Australia) can configure the examination doc format to match local requirements.
Go live in minutes. No sandbox required, no hidden fees.
Building vs. buying identity verification infrastructure is one of the most consequential technical decisions a growing company makes. Here is the framework for getting it right.
Evaluating identity verification providers? This comprehensive guide covers every criterion that matters — from technical capabilities to pricing models to vendor stability.
Monolithic KYC bundles force you to pay for checks you do not need. Modular identity verification lets you compose workflows that match your exact requirements — and nothing more.
