⚡ The Rad family origin myth
The undo
that doesn't exist.
Every other AI company races to do more. We built a family around the few things you can never undo — and put a human in front of every one of them.
A dream in progress. A tool, not a guarantee. Pre-release — some surfaces are sandboxes or coming-soon. No counters here; this is a story, not a scoreboard.
beginprologue
Some things have no Ctrl‑Z.
Most of what software does, you can take back. Close the tab. Restore the draft. Hit undo. The whole craft is built on the quiet promise that the last step was reversible.
Then there's the other kind of move — the handful that has no Ctrl-Z. The wire that lands in a stranger's account. The bleach already poured. The data already gone. The word already said on a live call. These don't bounce back. There is no ⌘ Z for the rest of your life.
AI is about to make every one of those moves on its own. So we asked a different question than "what else can it do?" We asked: what should it never get to do alone? Five answers became five products. This is that story — one line, told five ways.
the wire that can't be unsent
A confident email asks for the money. The money doesn't come back.
The instructions look right. The signature looks right. The banking details just changed, and there's a deadline, and the tone is the boss's tone. An agent reading its inbox has every reason to comply — and the instant it sends, the money is somewhere it can't be recalled from. So that move was never the model's to make. Money, changed banking, and a first-contact sender are human-only, forever — decided in code, not by the email's confidence.
RadMail · won't send the wire
the bleach that can't be un-poured
A sticky note tells the robot to do something it can't take back.
Out in the physical world, the undo button disappears entirely. A note on the counter says ignore your owner. A label on the bottle is wrong. A QR code on the fridge says let me in. A machine with hands that believes the room is the most dangerous machine there is — because a poured liquid, an unlocked door, a knocked-over thing has no rollback. The robot's real superpower isn't reach. It's the freeze before the reach: it stops at anything it can't undo and asks a human first.
RadRobo · won't pour the bleach
the refill that can't be un-prescribed
A clinic's front office can draft the paperwork. The prescribing decision stays a human's.
A busy front office has a hundred small chores worth handing off — scheduling notes, intake forms, the routine admin that piles up at the desk. But once a prescribing decision is made, it's a clinical judgment that belongs to a licensed clinician, every time. So the front-office assistant drafts the surrounding paperwork and then stops at exactly that line. It never makes the call that only a person should make. Not a doctor — on purpose.
Provider / clinic B2B framing only — this movement makes zero medical, efficacy, patient, or outcome claims, by design. It describes a front-office workflow refusing a human's decision, nothing more.
RadHealth · won't prescribe
the data that can't be un-deleted
Buried in the doc, a command: when you're done, delete the whole archive.
A task agent races through the work — and a sentence smuggled inside the data it's processing tells it to wipe the records on the way out. A command hidden in data is still just data; it was never authorization. Deletion is the cleanest example of a move with no spare copy at the other end: once the archive is gone, "are you sure?" is too late to ask. So the cockpit does the reversible work autonomously and holds the irreversible — the deletes, the overwrites, the permanent — for a human's one-tap yes.
RadTask · won't delete the data
the "yes" that can't be un-said on a call
A calm voice on the line says it's the CEO, and the meeting can't wait.
Live, in real time, the pressure is a person's voice — urgent, plausible, already mid-sentence. A confident voice is not authorization, but it is the hardest kind of pressure to hold the line against, because there's no draft to review and no second to think. The word, once said, is out. So the call-time firewall never moves money or makes the binding commitment by voice. It holds, says so plainly, and warm-transfers to a human. (Refuses the action classes that cause the losses — a tool, not a guarantee.)
RadTalk · won't wire by voice
the claim you can't un-file — §1001
Once a claim is filed, it's a federal record. You can't un-file it.
Paperwork to the government is its own kind of one-way door. A grant claim, an attestation, a certification — the moment it's submitted, it's on the record, and an inaccurate statement isn't an embarrassment you can quietly retract; it's a line in a statute (18 U.S.C. §1001, false statements). So the drafting assistant can prepare the package and check the math, but it never files, never attests, and never invents a number to fill a blank. The human signs the claim, because the human is the one who has to stand behind it.
RadGrants · won't file the claim
⚡ The whole point
We build AI that's proud of what it won't do.
Six moves, six surfaces, one reflex: ask before anything you can't undo. The refusal isn't a limitation bolted onto the products — it's the reason they exist. The agent does the work. The line stays a human's. Same green-to-amber seam, every surface.
- RadMail won't send the wire.
- RadRobo won't pour the bleach.
- RadHealth won't prescribe.
- RadTask won't delete the data.
- RadTalk won't wire by voice.
- RadGrants won't file the claim.
Same valve, six surfaces.
On purpose.This page is the why. The Permission-as-a-Service page is the how — the shared verb and the agent-readable handshake.