Glean Compass prototype: a Drift flags view with orphaned and stalled work cards, and a Metrics and trust dashboard showing flag precision, coverage, guardrails, and the path to autonomy.

Compass by Glean

Acme Platform TeamWeek of Jul 6
CT
Chloe Tan
Team lead

This week’s drift

2 flags need review · high-precision only
⚠
Orphaned work
Billing migration
●●●○
Confidence
34 tickets · 4 engineers · active 3 weeks — links to no current Q3 priority.
Signal34 Jira tickets + 6 PRs + 1 design doc, all tagged “billing-v2”.
Best match“Reduce checkout latency” — semantic score 0.28 (below 0.60 bar).
Explicit linkNone — no OKR / initiative reference in any artifact.
Why flagged: real, sustained work with no priority behind it. Verify in seconds, then decide.
Dismiss reason (trains Compass for your team):
◷
Stalled priority
Launch SSO for enterprise
●●●●
Confidence
0 active work streams in 18 days. Owner: unassigned.
SignalStated Q3 priority with no linked tickets, PRs, docs or threads in 18 days.
OwnerUnassigned since the goal was imported.
RiskEveryone assumes someone owns it; nobody does.
Why flagged: a priority the company said matters, with no one moving it. Assign an owner or consciously de-scope.
Assign an owner to “Launch SSO for enterprise”:
Dismiss reason (trains Compass for your team):

Linked & healthy — no action needed

Checkout latency fixes→Reduce checkout latencysemantic 0.83
OTel logging rollout→Migrate logging to OTelexplicit link
API rate-limit hardening→Improve API uptime to 99.95%semantic 0.71

Q3 priorities · Platform

Reduce checkout latency · J. Rivera
On track
Migrate logging to OTel · K. Shah
On track
Improve API uptime to 99.95% · M. Lee
On track
Launch SSO for enterprise · unassigned
Stalled

How Compass decides

Precision first.High-confidence, explainable flags only. Five you agree with beat fifty you don’t.
Flags work,never people. Every card is framed for a decision, not a verdict.
Shows its work.Each flag ships with its evidence, so you verify in seconds.
Prototype · PM exercise — Not a Glean product

Metrics & trust

Compass lives or dies on one thing: do people trust the flags? So the scorecard leads with precision, not volume — and pairs every outcome with a guardrail against harm.

Outcomes — is it working?
★ North star
7
Confirmed flags that led to a decision, this team, this week.
re-link · assign · accept
◎ Flag precision
78%
Flags you confirmed as real. Pilot bar ≥ 70%.
✓ above the ≥70% bar
⏱ Time-to-verify
41s
Median seconds to act on a card — an evidence-quality proxy.
✓ under 60s target
◈ Coverage
86%
12 of 14 active work streams linked. This is recall.
tracked, never chased
Precision — the metric Compass lives or dies on

Flag precision over the pilot

Rises as the per-team classifier learns your dismiss/accept signal
Flag precision Pilot bar (≥70%) Shadow mode — flags visible only to the lead
Guardrails — proof we watch for harm
⚖ Dismiss rate
Rising dismissals mean we’re becoming noise. An auto-throttle raises Compass’s own bar before that happens.
18% of flags
0%healthy — well below throttle
🔕 Snooze / mute rate
How often the weekly digest gets muted — an early signal the cadence or content is off.
9% of digests
0%low — digest is landing
🛡 “Felt surveilled” signal
A qualitative check that ICs don’t experience Compass as monitoring. If it ever reads as surveillance, the product fails.
✓None flagged this pilot
Compass flags work streams, never scores people.
Trust → autonomy — earning the right to act

The autonomy ladder

Each rung earns the next · gated on precision
  • Observe silently — shadow mode, 2 weeks done
  • Surface to the lead — weekly digest, high-precision flags you are here
  • Surface to the team — ICs self-link their own work
  • Act on the low-risk, reversible slice — auto-link, auto-nudge
  • Higher-stakes actions stay human-approved — always

When Compass may act

The rule that governs “suggest → just do it”: confidence × reversibility.
Reversible
link · nudge
Irreversible
reassign · close
High conf.
Auto ✓
Suggest
Low conf.
Suggest
Suggest
Auto only in the green cell — and an auto-action there is one click to undo. By construction, nothing permanent happens automatically.
Precision over coverage. A few trusted flags beat a noisy dashboard — so recall is tracked, never chased.
Prototype · PM exercise — Not a Glean product