Signals overview
A Signal is a pattern the AI has found in your feedback and can point at evidence for — “burnout risk is showing up in this team’s answers”, “people keep raising the same blocker”. You don’t ask for signals; they accumulate on their own as feedback arrives.
The feature has two halves, and they answer different questions:
- Signal Catalog (
/signals) — what are we watching for? A library of signal rules you switch on and off. - Signal Intelligence (
/signals/active) — what has actually been found? The list of detections, each with its evidence.
Rules and events are different things
Section titled “Rules and events are different things”This distinction runs through every page in this section.
A signal rule is a standing instruction: a name, a description, a detection prompt describing what to look for, and a minimum confidence. Fifteen rules are built in, and paid plans can add custom rules.
A Signal Event is one detection produced by one rule. It carries a title, severity, confidence score, and the IDs of the feedback entries it drew from. A rule that never matches produces no events; a rule that keeps matching accumulates hits on a single event rather than spawning a new one each time.
What gets read
Section titled “What gets read”Each evaluation pulls a rolling 90-day window of a team’s feedback, capped at 500 entries, from three places:
- feedback messages sent by people in the workspace
- text answers to survey questions
- follow-up answers within survey responses
Feedback that isn’t attached to a team is evaluated at workspace level instead, so a workspace with no teams still produces signals.
Connected sources and integrations contribute enrichment context — Git, calendar, project-management, and network metrics — alongside the feedback text. That context is best-effort: if a fetch fails, evaluation carries on without it.
Why detections are conservative
Section titled “Why detections are conservative”Signals are generated by an LLM, which makes over-claiming the obvious failure mode. Three mechanisms exist specifically to suppress it.
Triage first. One cheap pass sees all the feedback and only the rule names and descriptions, and returns which rules are plausibly relevant. It’s deliberately inclusive — false positives here cost a little money, false negatives cost a finding — but it stops fifteen full evaluations running against feedback that clearly has nothing to do with twelve of them.
Detect, then critique. A rule that survives triage runs a detection pass. If detection claims a signal, a second call re-examines the claim against the source text alone — not against the detection’s own reasoning — and returns which sources genuinely support it and which don’t. Detection alone is never enough.
Then the threshold. A signal is only recorded if the critic verified it and the verified confidence clears the rule’s minimum. Built-in minimums sit between 0.5 and 0.65 depending on the rule.
The upshot: a signal event you can see has already survived a pass designed to throw it away. That is what makes the evidence list worth reading rather than a formality.
Severity is derived, not assigned
Section titled “Severity is derived, not assigned”Nobody sets a signal’s severity. It falls out of the verified confidence score:
| Confidence | Severity |
|---|---|
| 0.9 and above | Critical |
| 0.7 – 0.89 | High |
| 0.5 – 0.69 | Medium |
| Below 0.5 | Low |
So severity answers “how sure are we this pattern is real”, not “how bad is it”. A Critical morale signal means the evidence is unambiguous, not that morale is at its worst possible point. It’s a useful thing to keep straight when you’re triaging a list.
Signals age out
Section titled “Signals age out”Evaluation runs hourly, and skips teams with no new feedback and no new rules since the last pass — so an unchanged team costs nothing.
When a rule matches again, the existing event is updated rather than duplicated: its hit count goes up, its confidence takes the higher of the two, its severity ratchets upward and never down, and the new sources join the existing list. That accumulated hit count is what the interface calls Signal Strength.
A signal that hasn’t been detected for seven days is archived. Active signals are therefore a picture of what is happening now, not an accumulating pile.
Signal, Conflict Signal, Analysis Confidence
Section titled “Signal, Conflict Signal, Analysis Confidence”Three similar-sounding terms, three different things:
- A Signal is what this section is about: a workspace-wide pattern detected by a signal rule.
- A Conflict Signal is a survey-analysis result — meaningful groups of Participant Response Transcripts disagreeing about the same topic. It belongs to one survey run, not to the signal catalog.
- Analysis Confidence is the product’s confidence in a survey’s analysis given the transcripts available for that run. A signal event’s confidence score is a separate number about a separate claim.
What to do with a signal
Section titled “What to do with a signal”A signal is an observation, not an instruction. Two things follow from one:
- Actions — lightweight items you attach to a signal event to record what you’re doing about it. See Signal events and evidence.
- Automations — rules that pair a condition on a signal with an action. See Automations.
Where to go next
Section titled “Where to go next”- Signal catalog — the fifteen built-in rules and what each detects
- Signal events and evidence — reading a detection and tracing it back
- Custom signals — writing your own detection rules
- Automations — turning conditions into actions