Is anyone still checking the AI?
Perpendiculum (Latin): the plumb line. Roman builders used it to check that a wall was true. We use it to check whether your oversight is.
You put AI into the workflow and kept a human in the loop. Somewhere along the way, people may have stopped questioning it, or stopped saying so out loud.
Perpendiculum measures whether human oversight of your AI still works. It is anonymous by design, so people tell the truth.
[ REQUEST_AN_AUDIT ]Five indicators of working oversight.
Candor gap.
The distance between what people say about the AI in the room and what they believe privately. When it grows, the team is performing agreement. The Pretend Era, measured ->
Trust calibration.
Whether people can tell when the AI is likely to be wrong, and still check its work before acting on it.
Decision drift.
Whether the team's decisions have actually improved since the AI arrived.
Repair latency.
How quickly the AI's mistakes get caught and corrected.
Escalation health.
Whether people would raise a disagreement with an AI-assisted decision, and whether anything happens when they do.
One score, five indicators.
A Perpendiculum score from 0 to 100, a score for each indicator, the candor gap, and results by team wherever a team is large enough to stay anonymous.
We read the system, not the person.
People only tell the truth about a tool their boss chose if they cannot be identified. So Perpendiculum is built so that nobody can tell who said what: not the client, and not us.
- No identity is recorded. No names, emails, IP addresses or timestamps. We never receive your staff list.
- One-time, unlinkable links. Each person gets a single-use link through your own email. Blind-signature tokens mean even the link can't be tied to the answer.
- Small teams never stand alone. Any group below the privacy threshold is merged or hidden, so no one can be singled out by elimination.
- Results only after the survey closes. Scores are computed once, with calibrated statistical noise, so they can't be reverse-engineered by watching them change.
- Ratings only, no free text. Nothing anyone writes can give them away.
- Raw answers are deleted. After the retention window (90 days by default), only the published results remain.
Five steps.
- 01Verify. Every engagement opens with an independent cryptographer's review of the privacy layer and a psychologist's check that the questions fit your setting.
- 02Set up. We define reporting groups with you, sized so every group stays anonymous.
- 03Collect. Your people answer ten short statements through one-time links sent from your own email.
- 04Read out. You get the scores, the candor gap, and what we would change first.
- 05Re-measure. We run it again after the changes, so you can see whether oversight actually improved.
If your AI policy says “a human reviews it,” Perpendiculum tells you whether that's still true.
For teams where a person is meant to stay in the loop: claims and underwriting, clinical and care work, legal review, content moderation, customer operations, code review. The readout is evidence your AI governance and risk reviews can use.
Tell us about your deployment.
A few questions so the first conversation is useful. Imran reads and replies to every request personally.
Rather talk first? Book a meeting ->