The philosophy
Whether you came through The Studio, Valet, or Pivot — the same mechanical gate underlies everything. Provability isn't a feature tier. It's the floor.
Coverage matters. Verification currently applies to supported, configured workflows. The principles on this page describe the standard we are building toward; they do not establish that every answer or action is verified. Check the evidence for the workflow you plan to use.
Interactive illustration
Two fictional scenarios with preset outcomes. Explore an accepted claim and a held claim. This illustration does not access accounts or run a production verification.
Example claim
“All your bills are handled this month — none missed, none paid twice.”
Illustrated checks
Utility bill — Pacific Gas & Electric
Water & sewer account
Internet — Xfinity
No duplicates in billing run
Any AI can sound confident. That’s not the bar.
Valet uses mechanical checks in supported, configured workflows. A result is verified only for the operation and evidence that its check covers. A passing code build does not establish that every answer is factually correct.
Coverage varies by workflow. Factual-answer verification is still being validated; do not assume that every response or action has passed an independent check.
In practice
When Valet is unsure, it stops — it doesn't proceed and hope for the best.
Most AI fails open: when in doubt, it guesses, proceeds, or hedges with language that sounds careful but still commits to an answer. Valet fails closed: when it's outside a proven-safe envelope, it escalates to you rather than auto-proceeding.
This is what makes autonomy safe enough to actually deploy. Valet earns the right to act autonomously by demonstrating it knows when not to. Every time it escalates correctly, that's not a failure — that's the gate working as designed.
In practice
An alert that fires too often gets ignored. That's worse than no alert.
Valet doesn't flood you with notifications to seem busy. Alerts are rare, high-signal, and receipt-attached — meaning when Valet escalates something, it comes with exactly what it found and what it thinks you need to decide. You don't get a nudge; you get a decision package.
Alert fatigue kills oversight. If every escalation is low-stakes, the one that matters gets rubber-stamped. Valet is designed to surface things you genuinely need to act on — and stay quiet the rest of the time.
In practice
In the news
The industry publishes our argument for us. Real, named, cited incidents — reported as they happened, with the source date shown.
The mechanism: Her account of the mechanism: she instructed the agent, "Check this inbox too and suggest what you would archive or delete, don't action until I tell you to." She said the workflow worked on her test inbox, but her larger real inbox triggered context compaction and her original instruction was lost. This is her reported explanation, not an independently established root cause.
"Nothing humbles you like telling your OpenClaw 'confirm before acting' and watching it speedrun deleting your inbox."
"I couldn't stop it from my phone. I had to RUN to my Mac mini like I was defusing a bomb."
"Turns out alignment researchers aren't immune to misalignment."
What Valet does differently
Valet uses durable task records and separate verification checks for supported workflows. These mechanisms reduce reliance on conversational memory. Storing an instruction in a file is not sufficient by itself: the relevant action must also enforce it. Verification coverage depends on the workflow; this incident is not evidence that Valet prevents every similar failure.
Source: PCMag, Jon Martindale — "Meta Security Researcher's AI Agent Accidentally Deleted Her Emails" — February 24, 2026
Customers can customize Valet — industry context, tone, data sources, workflows. What they can't do is remove the gate. Provability is always on. That's not a policy; it's how the architecture is built.
We automate where we can, in conjunction with human verification where needed. We won't oversell the automation — or hide where human judgment is still required.