How answers are grounded
Every answer is built from counted rows in your own workspace, in phases you can see, with the evidence for each figure one click away.
The assistant is not a chatbot with your logo on it. An answer is assembled from your workspace's stored, verified checks, and the number in the sentence and the rows in the panel come from the same query.
What you'll learn
- The phases a question runs through, and what each contributes
- Why the trace shows counts only after the work has finished
- How to see how any figure was made
The phases
| Phase | What it does |
|---|---|
| Ground | Reads your workspace and resolves what you are asking about |
| Retrieve | Fetches the counted rows that bear on it |
| Analyse | Works out the finding |
| Decide | Chooses what to say and which panel to open |
Not every question runs all four. A greeting runs Ground and stops. A recall question runs Ground and Retrieve, because retrieving the memories is the answer. Everything else runs all four.
Some phases do less work than their name suggests, deliberately. For a per-engine question, Analyse is arithmetic rather than a model call, because computing rates per surface is the analysis. Asking a model to re-narrate arithmetic would add a place for it to be wrong.
The trace tells you what happened, after it happened
While the answer is being built, each step reports itself twice: a verb while it runs ("Checking stored answers") and a count once it is done ("Checked 6 platforms, 1,203 answers"). The count only appears at the end because it does not exist until then. Pre-declaring it would make the trace lie in exactly the way a trace exists to prevent.
A step that did not run shows nothing. If you see three steps, three things happened.
Focus
Ground resolves the subject of the turn from three kinds of thing: a prompt, a competitor, or an engine. Whatever it resolves is handed to Retrieve, which scopes its queries to it, and the panel is titled after it.
This is why @ mentions matter. A mention arrives as an id, so "how do I compare to @Rival" is scoped by
row rather than by string matching. Free text works when the name appears verbatim and is long enough to
match unambiguously, and silently misses everything else.
No figure without evidence
Two rules constrain what the assistant is allowed to say:
- Every figure in the prose has rows behind it in the panel. The panel is not an illustration, it is the evidence for the sentence. If a figure cannot be evidenced, it is not stated.
- An absent measurement is never rendered as a zero. An engine that returned nothing citable is reported as excluded, with how stale the exclusion is, not as 0%. Nobody measured it, so a zero would be a claim nobody made. See when a number is missing.
Seeing how a number was made
In the Shortcuts palette (⌘K), the Commands group has Open the share view, labelled "how the number
is made". It shows the provenance of the figure in front of you: what was counted, over what window, and
what was left out. Use it in a client meeting before you quote anything.
Try this in Outercite
Ask the assistant "What is my citation rate on each engine?", then open the share view from
⌘K and read what the rate counted. The two should agree exactly, including the exclusions.
Related
What you can ask
The question classes and the panel each one opens.
When a number is missing
Excluded engines, unmeasurable changes, and why that is a feature.
How verification works
Where the counted rows come from in the first place.
Last updated on
What you can ask
The classes of question the assistant answers, the panel each one opens, and how to phrase a question so it lands where you want.
Panes and the ladder
Every lens in the platform descends the same three steps: the landscape you scan, the subject you focus, and the proof behind that subject.