Users trust a model that tells them how sure it is.
We design the product around the model's failure modes as well as its answers.
Layer 1 · the bare answer. The model's answer, with nothing around it. It reads as certain, whether it is or not. The user can only accept it or reject it.
Layer 2 · how sure. 84 out of 100: likely, not certain. The score is checked against past answers that users verified. 90 and up shown as the answer; 70 to 89 shown as likely, source one tap away; Below 70 the model says it doesn't know
Layer 3 · where from. Clause 14.2, on page 6. One tap opens the sentence. Clause 9 can end the agreement early, so it's flagged next to the renewal date.
Layer 4 · I don't know. The notice period is on page 7, which is a poor scan. The model says so instead of guessing, and points to the page the user can check.
Layer 5 · for this user. The same facts, at the depth each user needs. A buyer gets the date and what to do. A lawyer gets the clause and what the model couldn't read.
Designed around what the model does well, and where it can be wrong.
Product · Mentally Copilot. A legal and tax research assistant that answers with cited sources. Each source opens to the matched passage, the full document and the official link, so the professional can check it before relying on it.
Product · Mentally Copilot. The standard financial figures come from fixed, visible formulas. A language model never computes them. Models work where judgment or language is needed, with a professional reviewing the output.
Internal · design record. A prepackaged report first, then the user asks open questions where they see a gap. Open answers can be wrong, so a professional reviews them. The cost of that choice is on the record.