How this system is itself governed

Trust and audit

Where the models run, what may leave the bank, the rules every answer passes, and a tamper-evident log of everything anyone did.

Deployment

On-premise console and gateway. No inbound internet access.

Local only

Stays in the bank

Raw logs, configurations, identities, customer records, the evidence store and the audit log.

Local models

gpt-oss-120b and qwen3-32b on the bank's own GPUs, served by vLLM behind Mesh.

Cloud models

Only questions about the frameworks themselves, enforced by Mesh data classification.

Off

Answering with the local grounded engine: retrieval and templated prose over the same evidence, with zero network calls.

Each compliance task is pinned to a route. In a bank the local model is the default; the cloud chain is optional.

Task	Local route	Cloud route and fallbacks	Why
Readiness questions	gpt-oss-120b · vLLM	openai/gpt-oss-120b → anthropic/claude-sonnet-4.5	Reasoning over controls and evidence, answered with citations.
Evidence pack narratives	gpt-oss-120b · vLLM	openai/gpt-oss-120b → anthropic/claude-sonnet-4.5	Examiner-facing prose, graded on precision.
Change impact	qwen3-32b · vLLM	qwen/qwen3-32b → openai/gpt-oss-120b	High volume, runs on every configuration change.
Policy versus configuration	Rules comparator	typesafe/jev-1.13 → openai/gpt-oss-120b	Near-instant comparison of written policy with live access and configuration.
Claim verification	Deterministic verifier	—	No model: citations are checked against the evidence store.

Enforced in code, not in a policy document.

1
Rules set status
A deterministic rule over the evidence decides every control's status. Models only explain it.
2
Every claim cited
Each factual sentence ends with the control and evidence it rests on.
3
Verifier after every generation
Citations that do not resolve to evidence on that control are stripped; uncited claims are flagged.
4
No compliance conclusions
Sentences that declare compliance or predict an exam outcome are removed. The examiner concludes.
5
Refuse, don't guess
No matching evidence means "insufficient evidence", never an estimate.
6

10 golden questions, one with no answer in the evidence. Re-run on every model, prompt or route change.

Append-only. Each entry's hash covers the previous entry's hash, so any edit breaks the chain.

#	When	Who	Action	Detail	Hash

No entries yet

Nothing yet. Ask a question, draft a narrative or run a collection and it appears here.