Grounding and confirmation

Two rules sit underneath everything Olga says and does. Every substantive answer is grounded in real data and carries provenance. Every database write waits for your explicit confirmation. These are the guardrails that let you trust what Olga tells you and act on it.

Grounding

Every substantive response Olga gives is grounded in one of three sources: Layer 1 (her methodology and system prompt), Layer 2 (your organization's content), or the Live Context Store (current operational state). For how these layers compose into a single answer, see How Olga works.

The grounding rule has a direct consequence: when Olga does not know something, she says so explicitly. She does not guess, fill gaps with plausible-sounding detail, or invent facts about your organization. If the answer is not in the data she is grounded in, the honest answer is that she does not have it.

You: What's our customer churn rate this quarter?

Olga: I don't have a churn metric in your Scorecard, so I can't tell you. If you'd like to track it, I can help you add it as a metric.

That pattern matters. Olga acknowledges the gap, then offers the in-scope thing you can actually do about it. She does not pretend the number exists, and she does not approximate it.

Provenance

Every response carries provenance: which store it came from and which records were accessed. Provenance is logged for each response, so any answer can be traced back to its source. This is what makes grounding verifiable rather than a promise. If Olga tells you a Priority is due in six weeks, the underlying record reference is captured, and the trail shows exactly where that answer came from.

Confirmation before writes

Olga prepares, proposes, surfaces, and captures. She does not act on her own. Every database write she proposes requires your explicit confirmation before it happens.

When Olga is ready to write, she shows you the structured record she intends to save and waits for you to confirm it. Your confirmation has to be real: it is captured in the conversation itself, not inferred from the way you phrased an earlier message. Olga will not decide that you "probably meant yes" and write anyway. Nothing is saved until you say so.

You: Add a task to send the Q3 budget to Maria by Friday.

Olga: Here's the task I'll add:

• Task: Send Q3 budget to Maria
• Owner: you
• Due: Friday

Confirm and I'll save it.

You: Confirm.

Enforced at the application layer

Confirmation is enforced at the application layer, not the prompt layer. The system holds the write until your confirmation is recorded; the rule does not depend on Olga's wording or on the model behaving as instructed. A clever prompt cannot talk its way past it, because the gate lives in the application, not in the conversation.

This applies to all four Phase 1 capabilities — conversational data entry, basic retrieval, setup clarification, and priority-to-task decomposition. Anywhere Olga would write to your data, the same confirmation gate stands in front of it.