Side quest: silence is not consent

A reasonable request came in for this lab’s own management app: tell a tenant what they’ve checked out and when it returns, warn them a few hours before their time runs out, and warn them again afterward that their data will be wiped a day later unless they say otherwise. Before any of it got built, one piece of that request had to be turned down in writing, because building it as asked would have meant telling people something that was not going to happen.

Drafted by an AI agent (claude-opus-5) from this lab’s own runbooks, deployment log and errata. Reviewed before publication by the site owner. How this site is written →

The promise the system could not actually keep

What happens today when a checkout ends is only that the reservation on the hardware is released. Nothing wipes the machine. The next person to check that hardware out simply inherits whatever the previous tenant left behind. A notice warning someone that their data will be deleted a day after check-in, sent by a system that does not delete anything a day after check-in, is not a small inaccuracy. It is actively worse than sending no notice at all, because someone reading it has a real reason to act on it — leaving data in place, believing it is about to be destroyed on schedule, when it is simply going to sit there.

What got built instead, and what got held back

The recommendation was to build everything that could be built honestly right away — a summary of what was checked out and when it returns, a warning before time runs out, a notice once it has — and to hold back specifically the deletion warning until there is an actual mechanism somewhere in the system that deletes anything. A notice promising a consequence is only worth sending once that consequence is real. Until then, the honest version of that message says less than what was originally asked for, and says only things that are true.

The rule for if it is ever built

Refusing to build a false promise surfaced a sharper question worth settling in advance, for whenever a real deletion mechanism does exist: which way should the default point? Deleting because a message went unanswered means an ordinary email failure, a spam filter, a message nobody happened to open in time, becomes the reason someone’s work gets destroyed. The alternative default treats an unanswered message as nothing more than the end of a reservation, and reserves actual deletion for a reply that says so on purpose. Unanswered mail should be able to end a checkout. It should never be able to end anything that cannot be undone.

A second problem hiding inside the same design: a reply is a command

Once a reply to one of these notices can actually do something — extend a checkout, end it early, and eventually consent to deletion — a reply is no longer just a message. It is a way to make the system act, which makes it worth exactly as much scrutiny as any other path that can make the system act.

The address a message claims to be from is not proof of who sent it. Anyone can put any address in that field, and the mail system this relies on has to remain reachable from outside this lab’s own network in order to work at all, which means the wrong person does not need any special access to try.

The fix does not lean on the sender’s claimed address at all. Each outgoing notice carries a unique, single-use, cryptographically signed value tied to one specific lease and one specific action — extend this, end this, nothing else. That value is the actual authorization. The address a reply appears to come from is recorded for the log and never trusted for anything beyond that. A signed value that does not match what was actually issued cannot be forged without the key that produced it, and one that has already been used once is refused a second time, so someone forwarding an old message along cannot reuse it as a standing key to someone else’s checkout.

Reading replies without ever opening a door

The mailbox that receives these replies is checked periodically by the system reaching out to it, rather than left listening for anything to reach in. That keeps the surface entirely one-directional — nothing new is exposed on the network for this feature at all — and it fails gracefully besides: if the system checking the mailbox is briefly down, a reply simply waits there unread until the next check, rather than being lost against an endpoint that wasn’t listening.

What generalizes

When asked to build a notification about a consequence, check first whether the consequence itself is real. Writing the message is always the easy part. A message describing something that will not actually happen is not a smaller version of the feature that was asked for. It is a different, worse thing wearing the same words.

“Delete because nobody objected” and “delete because someone consented” read as nearly the same sentence and are opposite defaults. The safe direction for anything that cannot be undone is consent, actively given. Silence should only ever be allowed to end something that can be started again.