ADR 0003 — Deterministic scoring before AI

• Status: Accepted
• Date: 2026-07-13

Context

The readiness score is the core output. In a regulated domain, users and future bank partners must be able to trust and audit it. An LLM-generated score would be opaque, non-reproducible, and hard to defend.

Decision

Implement v0 scoring as a pure, deterministic engine: seven weighted categories summing to 100, each applying explicit point deductions that produce a flag with a severity and a plain-language fix. Risk bands: 80–100 low, 55–79 medium, 0–54 high.

Consequences

• Positive: fully explainable and unit-testable; identical inputs always give identical outputs; every flag maps to a concrete fix — explainability is the product.
• Negative: rules are hand-authored and must be maintained. Mitigated by docs-as-code generation and unit tests.
• AI is reserved for later, additive layers (document OCR, narrative drafting) — never for the final score in v0.