We used to talk about psychological safety as a feeling. Then we started measuring it, and the numbers told us we were wrong about who had it.

Ask a team if they feel psychologically safe and almost everyone says yes — it's a leading question with a socially correct answer. So we stopped asking, and started measuring behaviors that safety actually predicts.

What we measure instead of asking

Three proxies, all pulled from systems we already had: the ratio of "I" statements to passive voice in postmortems, the gap between how quickly senior versus junior engineers post the first comment on an incident channel, and how often someone flags a near-miss that never became an incident at all.

safety_index (computed monthly per team):
  postmortem_first_person_ratio: target >= 0.6
  time_to_first_junior_comment_vs_senior: target <= 1.5x
  near_miss_reports_per_10_incidents: target >= 3
  Composite score below 0.5 triggers a facilitated retro-of-retros,
  not a survey, not a Slack poll.

What the numbers found that the survey missed

Our platform team scored 94% "I feel safe raising concerns" on the quarterly engagement survey. Its near-miss reporting rate was the lowest of any team in the company — 0.4 per 10 incidents, against a company average of 3.1. People said they felt safe. Almost nobody was actually flagging small things before they became big things. The gap between stated and revealed safety was the real finding.

"Of course I'd say I feel safe on a survey my manager can see the aggregate of. Ask me something I can't perform an answer to."

What we did with the platform team

Not a training. A structural change: we made near-miss reports non-attributable in the tracking tool — visible as a count, not tied to a name in any dashboard leadership sees — and gave a small team shout-out (not individual credit) for the team with the most near-misses caught, reframing reporting as a team win instead of a personal admission. Within one quarter their rate rose to 2.6 per 10 incidents, still catching up to the company average but moving fast.

  • Self-reported psychological safety and behaviorally-revealed safety are different things — measure the second one.
  • Near-miss reporting rate is one of the cheapest, most honest safety signals available, and most teams don't track it.
  • Make the metric team-level and non-attributable before you use it to change incentives, or you'll just teach people to hide near-misses too.