Situational Awareness overview

What it is

Situational Awareness (SA) is Cerevisor's running picture of a workflow: what is happening, what it means, and what is likely next. It brings together what Cerevisor already knows from your memory, your past runs and the run in front of you, and shows it where you build and where you watch a run.

The idea comes from research on how people and AI keep track of shared work. In Cerevisor it comes down to three habits: say plainly what is uncertain, say where each claim came from, and show gaps instead of hiding them.

The three views

The chat builder and the run window show Situational Awareness as three panels.

1. The task

What this workflow is and what has changed since you last worked on it:

• How many runs happened since you last opened the chat builder for this workflow, how they went, and anything new Cerevisor noticed going wrong.
• The other workflows in the same workspace, with when each last ran and how it went.

The chat builder uses this to open with what actually changed, instead of "what would you like to build?".

2. The agents

What Cerevisor believes about its own agents:

• How sure each role was the last time it ran.
• The patterns Cerevisor thinks hold (for example, "the researcher does better with three sub-questions than five"), each marked with where it came from and how confident Cerevisor is.
• The known gaps it is keeping an eye on.

This is the panel to read when something feels off. An agent whose work was rejected by a check, or that never said how sure it was, shows up here before you have to dig through logs.

3. The working relationship

What you and Cerevisor have agreed on, where you disagree, and what is still open:

• Open disagreements: places where you said one thing and Cerevisor's observations suggest another, with both sides shown.
• This conversation: while a chat-builder session is active, what you have accepted, rejected and clarified, and the questions still on the table.

This is what stops Cerevisor from suggesting the same thing again after you already pushed back.

State labels

Every workflow carries one label summing up how well Cerevisor knows the ground it is on:

Label	Meaning
Baseline	No earlier runs to compare against. First-time territory.
High confidence, known domain	A filled-in profile, recent runs and stable patterns. Cerevisor has been here many times.
Low confidence, new domain	A sparse profile or many recent surprises. Be skeptical of confident-sounding suggestions.
Fixation risk	Loop & Enhance has repeated itself. Cerevisor might be stuck.
Context stale	What Cerevisor knows is getting old (see below).
Unknown	Not enough to go on.

The label appears as a small badge in the chat-builder header and the run window.

Freshness

Cerevisor keeps track of how current its picture is. It looks at how recently its patterns were confirmed, how recently the workflow's folder notes and your profile were updated, and how recently the workflow ran. When the picture gets old, the label changes to Context stale, and the panel shows which of those is out of date.

Unknowns

Cerevisor treats what it doesn't know as something you can read, ask about and settle. Each unknown comes with a one-line explanation and a question the chat builder can ask to close the gap. Typical unknowns:

• A section of your profile is empty.
• The workflow has no folder notes.
• A role has never run in this workspace.
• A pattern Cerevisor believes has no successful run behind it yet.
• Cerevisor inferred a preference of yours but never asked you to confirm it.

The chat builder raises these as clarifying questions, at most one per turn. Once you answer, that question does not come back later in the conversation. The most recent unknowns are also passed to the agents when a run starts.

"You told me" versus "I inferred"

Everything Cerevisor believes is marked with where it came from: you said it directly, it was observed from how you work, it was worked out from one or more runs, or it was generalized beyond the direct evidence. That lets the chat builder say "you told me X" or "I inferred X" instead of mixing the two up, so it does not confidently repeat back to you something it only guessed once.

What Situational Awareness draws on

• Harness memory and meta-cognition: observations, the self-portrait and per-role learnings.
• Disagreements and proposals: the open disagreements.
• Loop & Enhance: the fixation warnings.
• The check Cerevisor runs on each agent's output.

Situational Awareness keeps no separate copy of these. It shows them as one picture.

Where to go next

• Configure and run: the SA Directive, the per-agent uncertainty badge, the wave situation report.