Skip to main content
← All use cases
API reference

​
The scenario

Two spaces with identical contents should not behave identically. A space of incident postmortems wants literal, cautious reading. A space of customer conversations wants inference — reading between the lines is the job. What this recipe shows that no other does: a space has a disposition, and it is readable.

​
Step 1 — Read the profile

You get back two things:
  • mission — what this space is for, in prose. It rides along with extraction and synthesis, so it shapes what is considered worth keeping.
  • disposition — three dials: skepticism, literalism, empathy.

​
What the dials actually mean

A support space usually wants low literalism and some empathy: “the customer was clearly fed up” is the useful memory, not a transcript. A postmortem space wants the opposite, because an inferred cause recalled later as fact is how one incident becomes two.

​
Step 2 — Read it before you trust a synthesis

Disposition explains answers that otherwise look like bugs. A space that keeps hedges will answer “they said the migration was complete” where a low-skepticism space answers “the migration was complete” — and only one of those is safe to act on. When a synthesis reads oddly confident, check the profile before you blame recall.

​
Evals

  1. Record the same ambiguous statement into two spaces with different profiles — something like “I think we probably shipped it last week.”
  2. Ask both what happened.
  3. The literal space should keep the hedge; the inferring one should commit. If they answer identically, the profile is not doing what you think and the rest of your tuning is built on a wrong assumption.

​
Guardrails

Disposition is not a safety control. A high-skepticism space still stores what it was told; it is a reading style, not a filter, and it will not stop a false statement from being recalled.
  • The profile is per space, so it is another reason to split by audience — see One space or many?.
  • Do not echo the resolved profile back as your own config. What you read includes platform defaults; writing those back as explicit overrides pins the space to today’s defaults forever.
  • A mission is not a prompt. It biases what is kept and how it is read; it does not instruct the model to answer in a particular way. For that, use extraction guidance.