Governed truth

Anything that would be true for the whole company becomes a proposal that someone with permission approves. What you save about yourself is written directly, because it is yours. That delay is the product, not friction someone forgot to remove. If anyone — or any AI — could change what the company believes to be true, nobody could rely on the contents, and a context layer nobody relies on is worse than no context layer at all.

​

What this looks like in practice

An AI connected to Riktra follows the same rule, through the same door. It can call propose_company_update; it cannot write company context directly. Approval is a separate tool, restricted to administrators and owners, and enforced in the database rather than in the tool description a model reads.

​

Superseding beats overwriting

An established decision’s meaning changes by creating a superseding decision, not by editing the original statement. This matters more than it sounds. If a decision can be edited in place, then the record of what the company believed last March is gone, and with it every explanation of why things were built the way they were. Riktra keeps the chain: what was decided, what replaced it, and when. Every shared-truth write is versioned in the same transaction as the write itself — enforced by a database trigger, so no code path can forget. Restoring an old version creates a new version. History is never rewritten.

​

Why the AI can be trusted with this

Because it is not trusted with it. The guarantee does not rest on model behaviour, a system prompt, or a tool description. It rests on which tools exist and what the database permits the caller to do. That is the difference between a product that says AI will not change your company’s truth and one where it cannot.