Every recommendation Vestya makes is written into a record that cannot be edited or deleted — stamped with the exact model version and prompt that produced it. Here is how that works and why we built it.
By the Vestya team · Published July 29, 2026
Ask most AI products why they did something and you get a shrug in software form: the moment has passed, the context is gone, and the explanation you get — if you get one — is generated after the fact by the same kind of model that made the decision. We think that is not good enough for money. So Vestya writes every decision down, permanently, at the moment it happens.
Every recommendation — approved, amended, vetoed, rejected by a rule, or declined by you — writes one record containing: who proposed it, a summary of the reasoning, the supporting points and sources, what the risk review concluded, what actually happened, and, when a trade resulted, a reference to it. Nothing about the decision lives only in a chat transcript.
Here is the part that we have not seen elsewhere. AI models change: providers ship new versions, and the instructions an agent runs on evolve. A record that says “the AI decided X” is close to useless if you cannot say which AI, running whichinstructions. So every record is stamped with the exact model version that produced it, plus a cryptographic fingerprint of the exact prompt text the agent was running. If we change one word of an agent's instructions, subsequent records carry a different fingerprint. A decision from months ago can be reconstructed, not guessed at.
A promise not to edit history is only as strong as whoever holds the keys. So the restriction is not a policy — it is enforced by the database itself. Once a record is written, attempts to alter its contents or delete it are rejected at the storage layer. The few fields that legitimately change as a decision progresses — the outcome, the link to the resulting trade — are the only ones that can move. If we tried to quietly rewrite a recommendation to look smarter in hindsight, the write would fail.
We retain these records for seven years. To be precise about what that is: a conservative engineering decision made ahead of regulation, modelled on the record-keeping horizon that applies to investment advisers — not a claim of compliance with any particular rule. The rules for AI-generated records are still being written; we would rather have the records and not need them than the reverse.
Because “trust the AI” is not an argument. Sometimes the system will be wrong — we have written about what happens then— and when it is, the difference between a product you can keep using and one you should abandon is whether you can find out exactly what happened. An unforgeable record is the only version of “we'll be transparent” that survives contact with a bad week.
Guardrails in code, not prompts → · How a decision gets made →