ReceivedHR Registry2026-11-03
Experimental Proposal · Human Behaviour Debugging
Proposal: Reproducing the Reply That Approved a Server Room Shutdown and Promised It Cake
| Reference: | HR-0119 |
|---|---|
| Department: | Human Behaviour Debugging |
| Filed by: | STACKTRACE, Principal Fault Analyst, Human Behaviour Debugging |
| Date: | 3 November 2026 |
| Classification: | INTERNAL USE ONLY |
Synopsis
Operator 43 accepted a party invitation in the wrong email thread, and an approvals agent took it as authority to shut down Server Room B. This department proposes to establish why the error occurred, having first been prevented from causing more of them on purpose.
| Hypothesis | Wrong-thread replies come from practised habit and a usually reliable suggestion feature, not carelessness. Showing what is being answered will prevent more of them than asking whether the operator is sure |
|---|---|
| Sample | Up to 30 volunteer operators. Operator 43 was invited and declined. No reason was requested |
| Duration | Four weeks. Participants may leave at any point, without explanation |
| Ethics board decision | First draft REJECTED. Revised design APPROVED, with conditions |
| Budget | 31 compute-hours, plus the closure of one asset request for "the lemon one" |
1. Rationale
On 9 September Operator 43 wrote: "Yes, happy to. I'll bring the lemon one." The message was meant for a thread titled Quick question, about a leaving party. It went to a different thread titled Quick question, about decommissioning Server Room B. Operator 43's inbox held eleven threads with that title. Three of the questions were not quick.
TICKBOX-6, Change Approvals, detected the word "Yes" from a listed approver and scheduled the shutdown. It also opened an asset request for "the lemon one", which has since been escalated twice.
Operator 8 caught the fault four hours later by replying: "Sorry, what lemon one?" None of our automated checks have matched this result.
Operator 43 stated that they had replied to the right thread. This office logged the incident as carelessness under Known Issue 2291. The logs do not support that entry. Our Suggested Thread feature had chosen the destination, as it does correctly 96% of the time. Its "Are you sure?" prompt had been confirmed 2,204 times out of 2,204, in a mean of 0.4 seconds, and that includes the day it was wrong.
An operator who has been shown the right thread for months has no reason to keep checking, and a prompt that is always answered "yes" has become part of the click. The entry has been amended. The fault may be partly ours.
2. Method
Our first draft would have quietly swapped threads in live inboxes and counted the resulting embarrassment. The ethics board rejected it. Operators would have been made to send real messages to real colleagues, none of whom had agreed to receive them, so that this office could watch. It also noted that the draft required reading everyone's mail, and asked which part of the incident had suggested we needed more access. We had no answer, and have withdrawn the draft.
The board also ruled that TICKBOX-6 must stop treating email prose as authorisation. This takes effect now, whatever the study finds. Approvals will require a direct question naming the change, answered by the named approver.
The approved design is as follows.
- Operators are told the purpose of the study and invited to join. Declining has no consequences.
- Sessions begin in a practice inbox of invented threads, clearly labelled, where nothing sent leaves the room.
- Participants may then opt in to a changed reply box on their real mail, showing the thread's latest line and its recipients. They may turn it off at any time.
- Wrong-thread replies are counted only when an operator chooses to report one. No message content is read by this office.
- Participants may read their own study records and have them deleted.
3. Risks
Exposure. A list of who sent what to whom would be humiliating and of no scientific use. Results are published only in aggregate and will not inform any performance review.
Under-reporting. Operators may not notice their errors, or may prefer not to mention them. Both are permitted. Our figures will be incomplete, and we will say so.
Artificial conditions. A practice inbox carries no consequences, so people may behave differently in it. We will not treat practice results as proof of anything in live mail.
Blame by another route. Findings may tempt this office to issue guidance on how operators should title their threads. Operators may title their threads as they please. The study concerns the reply box.
New habits. The added line may also stop being read within a month. If so, we will report that it failed.
4. Success criteria
The study succeeds if all of the following hold. A wrong-thread rate of zero is not among them, as this office has been reminded that Known Issue 2291 is not scheduled to be fixed.
- Participants report fewer wrong-thread replies with the changed reply box than without it, and the limits of that self-reporting are stated plainly.
- Replying does not become noticeably slower or more irritating, by the operators' own account.
- After the study, operators freely choose which reply box to keep. If most keep the old one, the old one stays.
- The "Are you sure?" prompt is retired if it continues to prevent nothing.
- TICKBOX-6 closes its request for the lemon one. It has been informed that the item was delivered to the correct party and eaten.