When behaviour is unstable, trust must come from the architecture, not the model

Run the same query multiple times and you may get different answers, different sources, different reasoning paths and all delivered with equal confidence with no traceable lineage.

𝗣𝗶𝗹𝗹𝗮𝗿 𝟲 𝗼𝗳 𝗧𝗿𝘂𝘀𝘁𝘄𝗼𝗿𝘁𝗵𝘆 𝗙𝗶𝗻𝗮𝗻𝗰𝗶𝗮𝗹 𝗔𝗜: 𝗗𝗲𝘁𝗲𝗿𝗺𝗶𝗻𝗶𝘀𝗺
By CTO & Co Founder, Simon Gregory

𝗪𝗵𝗲𝗻 𝗯𝗲𝗵𝗮𝘃𝗶𝗼𝘂𝗿 𝗶𝘀 𝘂𝗻𝘀𝘁𝗮𝗯𝗹𝗲, 𝘁𝗿𝘂𝘀𝘁 𝗺𝘂𝘀𝘁 𝗰𝗼𝗺𝗲 𝗳𝗿𝗼𝗺 𝘁𝗵𝗲 𝗮𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲, 𝗻𝗼𝘁 𝘁𝗵𝗲 𝗺𝗼𝗱𝗲𝗹
In finance, trust isn’t a UX feature, it’s an 𝗮𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗮𝗹 𝗿𝗲𝗾𝘂𝗶𝗿𝗲𝗺𝗲𝗻𝘁.

The final pillar tackles 𝗻𝗼𝗻-𝗱𝗲𝘁𝗲𝗿𝗺𝗶𝗻𝗶𝘀𝗺, one of the most misunderstood realities of LLMs. Identical queries can yield different outputs, all sounding equally certain with no reconstructible lineage.

This isn’t a configuration issue.
It’s how probabilistic systems work and it puts them in direct conflict with the deterministic foundations that financial workflows depend on: 𝗮𝘂𝗱𝗶𝘁𝗮𝗯𝗶𝗹𝗶𝘁𝘆, 𝗿𝗲𝗽𝗲𝗮𝘁𝗮𝗯𝗶𝗹𝗶𝘁𝘆, 𝘀𝘁𝗮𝗯𝗶𝗹𝗶𝘁𝘆, 𝗮𝗻𝗱 𝘁𝗲𝗺𝗽𝗼𝗿𝗮𝗹 𝗰𝗼𝗻𝘀𝗶𝘀𝘁𝗲𝗻𝗰𝘆.

🟢 𝗪𝗵𝘆 𝗟𝗟𝗠𝘀 𝗮𝗿𝗲 𝗻𝗼𝗻-𝗱𝗲𝘁𝗲𝗿𝗺𝗶𝗻𝗶𝘀𝘁𝗶𝗰

LLMs explore 𝗽𝗼𝘀𝘀𝗶𝗯𝗶𝗹𝗶𝘁𝘆 𝘀𝗽𝗮𝗰𝗲𝘀, not fixed paths.
That’s what gives them power.

It’s also what makes them untrustworthy unless the architecture 𝗰𝗼𝗻𝘁𝗮𝗶𝗻𝘀 that variability.

It’s a butterfly effect system: small prompt changes can produce large shifts in tone, emphasis, or conclusion.

The consequence:
𝗘𝘃𝗲𝗿𝘆 𝗿𝗮𝘄 𝗟𝗟𝗠 𝗼𝘂𝘁𝗽𝘂𝘁 𝘀𝗵𝗼𝘂𝗹𝗱 𝗯𝗲 𝘁𝗿𝗲𝗮𝘁𝗲𝗱 𝗮𝘀 𝘂𝗻𝘁𝗿𝘂𝘀𝘁𝗲𝗱 𝗶𝗻𝗽𝘂𝘁 for any consequential downstream process.

You can’t tune away randomness.
𝗧𝗵𝗲 𝗼𝗻𝗹𝘆 𝘀𝗼𝗹𝘂𝘁𝗶𝗼𝗻 𝗶𝘀 𝘁𝗼 𝗰𝗼𝗻𝘁𝗮𝗶𝗻 𝗶𝘁.

A safe financial AI architecture enforces a 𝗱𝗲𝘁𝗲𝗿𝗺𝗶𝗻𝗶𝘀𝘁𝗶𝗰 𝘁𝗿𝘂𝘀𝘁 𝗯𝗼𝘂𝗻𝗱𝗮𝗿𝘆 around the generative layer with stable inputs, deterministic retrieval, authoritative lineage, and strict contextual integrity.

𝗡𝗼𝗻-𝗱𝗲𝘁𝗲𝗿𝗺𝗶𝗻𝗶𝘀𝗺 𝗶𝘀 𝗮 𝗽𝗿𝗼𝗽𝗲𝗿𝘁𝘆, 𝗻𝗼𝘁 𝗮 𝗳𝗲𝗮𝘁𝘂𝗿𝗲.
𝗜𝗻 𝗳𝗶𝗻𝗮𝗻𝗰𝗲, 𝗶𝘁 𝗺𝘂𝘀𝘁 𝗯𝗲 𝗰𝗼𝗻𝘁𝗮𝗶𝗻𝗲𝗱 𝗯𝘆 𝗱𝗲𝘁𝗲𝗿𝗺𝗶𝗻𝗶𝘀𝘁𝗶𝗰 𝗮𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲.

🟢 𝗪𝗵𝘆 𝘁𝗵𝗲 𝘀𝗶𝘅 𝗽𝗶𝗹𝗹𝗮𝗿𝘀 𝗺𝗮𝘁𝘁𝗲𝗿

The first five pillars each describe a distinct failure mode. Together, they describe something larger: a system under a common pressure. Non-determinism is that pressure. It is what makes the other five necessary:

Auditability. Authority. Provenance. Context Integrity. Temporal Integrity.

Together they form the containment architecture that makes probabilistic systems safe, repeatable, and auditable.

𝗧𝗵𝗲 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻 𝘄𝗮𝘀 𝗻𝗲𝘃𝗲𝗿 𝘄𝗵𝗲𝘁𝗵𝗲𝗿 𝘁𝗼 𝘂𝘀𝗲 𝗔𝗜 𝗶𝗻 𝗳𝗶𝗻𝗮𝗻𝗰𝗲. 𝗜𝘁 𝘄𝗮𝘀 𝗮𝗹𝘄𝗮𝘆𝘀 𝘄𝗵𝗲𝘁𝗵𝗲𝗿 𝘆𝗼𝘂 𝗵𝗮𝘃𝗲 𝘁𝗵𝗲 𝗮𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲 𝘁𝗼 𝘂𝘀𝗲 𝗶𝘁 𝘀𝗮𝗳𝗲𝗹𝘆.