Docs · LLMs and developers
Version3.13.0How memory apps mark each value, what counts toward a score, and what happens when a source or evidence is unavailable.
Evidence and inferred conclusions
Every value in an app carries its status and the evidence it came from.
Value status
| Status | Meaning |
|---|---|
| Recorded | A fact from one of your connected sources |
| Inferred | A conclusion Achiral inferred, with confidence at or above your team's setting |
| Approved | An inference an Owner or Admin approved |
| Missing | No evidence was found |
| Unavailable | The source that would supply it cannot be reached |
A missing value is shown as missing. It is never counted as zero. A score with a missing factor has no value and is shown as missing.
Opening the evidence
A value lists the events and memories it was worked out from. The evidence drawer shows each item's source, date, and a summary.
Evidence is read only from the app's team, its environment, and its bound sources (the sources the app reads). If an item has since been removed from memory, it is shown as removed, and the result is marked incomplete.
When several sources carry the same field, the app uses your source authority order and shows which source supplied the value.
The confidence setting
The confidence setting decides how confident an inference must be before it counts toward a score or stage. Every team uses Balanced. Only the team's Owner can change the setting; this is not yet available in the app.
| Setting | Minimum confidence |
|---|---|
| Strict | 0.9 |
| Balanced (default) | 0.8 |
| Permissive | 0.7 |
No setting goes below 0.7. An app can raise the bar for a specific factor but cannot lower it below the team's setting. Each change to the setting is recorded.
Inferences below the bar do not count. They appear in the review queue.
The review queue
The review queue lists what needs a person's attention, such as low-confidence inferences and sources that disagree. It is shown to the steward, Owners, and Admins.
Owners and Admins can act on an item:
- Correct a fact. The correction is saved to team memory.
- Record a decision. The decision, approval, or rejection is saved to team memory in the app's environment.
Approving a low-confidence inference makes it count as approved from the next run. Each action records the member, the app, and the version it came from. Repeating the same request does not create a second action.
Members who are not Owners or Admins can open evidence but cannot correct or approve.
Component states
| State | Meaning |
|---|---|
| Available | Loaded normally |
| Empty | Nothing matches |
| Stale | A bound source is paused or failing |
| Partial | Some bound sources are disconnected, or the results reached a run limit |
| Unavailable | All bound sources for this part are disconnected |
| Denied | You do not have access to this part |
| Quota limited | A narrative could not be generated, because generation is stopped or the page reached its limit of new narratives for one load |
Parts that depend on an unavailable source say which source. Reconnecting the source restores the numbers without republishing.
When a source disconnects
When a source an app reads is disconnected, the steward, the Owner, and the Admins get one in-app notice. Owners and Admins are pointed to Integrations to reconnect it.
Narratives
A narrative is text generated from an app's evidence, with citations to that evidence. Generation goes through your team's model providers.
A narrative is stored and reused while its evidence is unchanged. An editor can regenerate it. When generation is stopped, the narrative shows as quota limited and the rest of the page still loads.