Banking-Grade AI Governance: How Claris Solves the BFSI Explainability Wall

What Is the 'Explainability Wall' That Blocks AI Adoption in Banking and Finance?

Where the Wall Appears in the Model Approval Process:

In a typical Indian bank, a new AI model must pass through a multi-tier approval gate before production deployment:

1. Model Development — Data Science team builds and validates the model
2. Independent Validation — Model Risk Management team independently validates performance
3. Risk Committee Approval — Senior risk management signs off on the model's risk classification
4. Legal & Compliance Sign-Off — The legal team verifies that the model complies with all applicable regulations
5. Board IT Risk Committee — Final approval for high-risk models

The Explainability Wall typically appears at steps 3, 4, or 5 — when non-technical senior stakeholders ask:

• "Can you explain to me in plain language why this model rejects an application?"
• "If a customer asks why they were rejected, what do we tell them?"
• "If the RBI asks us to prove this model doesn't discriminate against gig workers, what's our answer?"
• "How do we know this model will behave the same way in 12 months as it does today?"

For models built on gradient boosted trees, neural networks, or ensemble methods — the dominant architectures in modern BFSI AI — these questions often cannot be answered satisfactorily with standard MLOps tooling.

The Legal Liability of Opacity:

In India, Section 21 of the Credit Information Companies (Regulation) Act creates obligations around credit decisions. RBI's Master Direction on Credit Information implicitly requires that credit decisions be explainable on request. A model that cannot produce explanations is not just a governance risk — it is a potential basis for regulatory enforcement action.

The EU AI Act goes further: Article 86 gives individuals the right to obtain an explanation of a decision made by a high-risk AI system. For EU-operating Indian banks, this creates a hard legal obligation — not a best practice recommendation.

How Does Claris's XAI Consensus Engine Provide 'Mathematics-First' Explanations for BFSI?

Why Single-Algorithm Explanations Are Insufficient for Regulators:

A regulator examining a credit rejection explanation sees a SHAP waterfall chart showing that "income_stability = -0.38" was the top negative factor. The regulator's question is: "Is this explanation reliable? Could you produce a different explanation tomorrow and claim that was also correct?"

With single-algorithm XAI, the honest answer is: "Possibly, yes." SHAP values can shift between model versions, between sampling approaches, and between background datasets. A single algorithm's output is an approximation — a useful one, but an approximation nonetheless.

The Consensus Approach:

The XAI Consensus Engine generates four independent explanations for every inference:

1. SHAP — Global feature attribution using Shapley game theory
2. LIME — Local approximation using neighbourhood sampling
3. Integrated Gradients — Path-based attribution for smooth decision surfaces
4. Counterfactual Analysis — Actionable "what would change the outcome" guidance

Each method produces a feature importance vector. The Consensus Engine computes:

• Direction Agreement — Do all methods agree on whether Feature X increases or decreases risk? If yes, the direction is reliable.
• Magnitude Correlation — How closely do the importance magnitudes correlate across methods?
• Consensus Score — A [0,1] score summarising overall explanation reliability

For BFSI Regulatory Purposes:

A Consensus Score ≥ 0.85 means: four independent mathematical methods agree on the explanation. This is the evidentiary standard Claris targets for regulatory submissions.

Fraud Model Explainability — The Unique Challenge:

Fraud models present a specific explainability challenge: the features that indicate fraud are also features that indicate legitimate unusual behaviour. A large international wire transfer can be normal (for a business) or fraudulent (for a retail account). The explanation must be contextualised to the customer segment.

Claris's BFSI-specific Fraud-Logic Visibility module provides segment-aware explanations that are calibrated against the customer's historical behaviour — not against population averages. This produces explanations that make sense to the customer and satisfy regulators, rather than generic feature attributions.

What Does a Cryptographically Signed 'Logic State' Integrity Proof Look Like?

The Audit Problem That Integrity Proofs Solve:

Consider a common regulatory audit scenario: An RBI examiner identifies a credit decision made on March 15, 2025, that appears to have discriminatory outcomes. They ask the bank to demonstrate that the model used on March 15, 2025, was the validated version — not a shadow model, not a partially updated version, and not a version that was modified after the decision was made.

Without Logic State Integrity Proofs, the bank can only provide:

• The current version of the model (which may have been updated since March 2025)
• Historical Git commits (which show code changes but not the deployed model state)
• Validation reports (which cover the model at validation time, not at inference time)

None of these directly prove that the March 15, 2025 decision was made by the validated model.

How Logic State Integrity Proofs Work:

At every inference, Claris captures a "Logic State Fingerprint" — a JSON document containing:

{
  "model_id": "credit_scorecard_v4.2.1",
  "model_hash": "sha256:a3f1c8...",
  "feature_importance_snapshot": {
    "income_stability": 0.312,
    "credit_utilisation": 0.287,
    "employment_tenure": 0.201,
    ...
  },
  "decision_boundary_params": { ... },
  "inference_input_hash": "sha256:b7e2d4...",
  "timestamp": "2025-03-15T14:32:17.823Z",
  "jurisdiction": "IN",
  "model_validation_ref": "MRM-2025-CR-004"
}

This Fingerprint is hashed using SHA-256, then signed with Claris's HSM-stored private key, creating an Integrity Proof that cannot be forged or modified.

Regulatory Submission:

When the examiner asks about the March 15, 2025 decision, the bank:

1. Retrieves the Integrity Proof for that inference from the Evidence Vault
2. Verifies the cryptographic signature (confirming Claris generated it at the stated time)
3. Compares the model_hash against the registered hash of the validated model version
4. Shows the examiner that the hashes match — proving the validated model was running at inference time

The Black Swan Visibility Feature:

During market stress events (COVID-19, global interest rate spikes, banking crises), model behaviour can shift dramatically in ways that no static validation can anticipate. Claris's Fraud-Logic Visibility module specifically tracks how "Black Swan" events influence model logic — monitoring for sudden shifts in feature importance that could indicate the model is behaving unexpectedly under stress conditions.

Banks with Fraud-Logic Visibility deployed detected unusual model behaviour within 48 hours of the March 2023 US banking stress event — allowing them to implement human review overlays before regulatory scrutiny arrived.