How did the AI come to its answer?

Model provenance and decision explainability: what the model was built from, which factors drove a score, and which sources grounded a generated answer.

Explainability 87%

Decision explainability

Factor contributions for the most recent scored decision. Upward factors increase risk.

Payment history (24m) ↓34%

Debt-to-income band ↑26%

Tenure with the bank ↓17%

Product holding breadth ↓13%

Recent arrears flag ↑10%

Grounding trace

For generated answers: what was retrieved, what was used, and what was rejected.

Step	Source	Score
Retrieved	Onboarding policy v9 §4.2 — proof of residence	0.91
Retrieved	Retail FAQ — address change evidence	0.84
Used in answer	Both chunks cited verbatim in the response	—
Not used	Regulatory bulletin 2026-04 (below threshold)	0.48

Model provenance

From base weights to served endpoint.

Base modelLlama 3.1 8B Instruct · Llama 3.1 licence

Adapteronb-lora-v4 · QLoRA r=16, alpha=32

Training dataonboarding-conversations · 42,130 rows · governed

Training runft-2314 · seed 3407 · reproducible

Evalharness run-914 · 714 scenarios

Registry entrybuild/mrm-2311 · signed

ServingvLLM · runtime-prod-02

Limits declared

What this model is not for.

Not for pricing decisions

scope is default probability only

Not validated below 18 months tenure

sparse training coverage

Human review required

for any adverse decision