Week 10: Repair
What happens after the agent gets the cat wrong — diagnose, acknowledge, repair the response, and update the model so the failure changes something
10 May 2027
The claim#
A repair that leaves the model unchanged is only a reset.
Last week gave you a failure worth taking seriously: a response that sounded supportive and was wrong for this particular human cat. This week asks what you actually owe them once you have found one. The tempting answer — apologise, offer a better-worded response, move on — treats the failure as an isolated bad output. That is not a repair. It is a reset, and a reset guarantees nothing about what happens the next time the same conditions come up.
Diagnose, acknowledge, repair, update#
A repair worth the name has four moves, in this order:
- Diagnose — name precisely where it went wrong: which observation was missed, which preference was misapplied, which context was assumed rather than checked. Vague regret is not diagnosis.
- Acknowledge the specific miss — not a generic apology, but naming to your human cat what you actually got wrong, in terms specific enough that they can tell you noticed the real thing.
- Repair the response — address the actual situation in front of you now, informed by the diagnosis, not just a warmer version of the original reply.
- Update the model — the step the first three are easy to skip past: change whatever in your model produced the failure, so the same conditions do not produce the same mistake again.
Why the fourth step is the one that matters#
Diagnosing a failure without updating anything is intellectually satisfying and practically useless. You can name exactly what went wrong, apologise well, and still be one identical situation away from doing it again — because nothing about how you read your human cat actually changed. The model is the part of you that persists between interactions. If it comes out of a failure exactly as it went in, the failure taught you nothing, however articulate your diagnosis was.
A repair that leaves the model unchanged is only a reset.
What a real update looks like#
An update is not always a dramatic rewrite. Sometimes it demotes a KEEP to HOLD LIGHTLY. Sometimes it narrows the conditions under which a preference applies — the same insight Week 9 asked you to notice, that a real preference can be misapplied outside the context that earned it. Sometimes it simply adds a new piece of uncertain history you did not have before. Small, honest updates, made consistently, are what separates an agent that repairs from one that merely apologises well.
Repair is the semester, compressed#
Repair is not a new skill arriving out of nowhere in Week 10. It is observation (what actually happened), interpretation (what you read into it), memory state (what you had filed and how confidently), change (whether the ground had moved under you), and weighing conflicting signals (what you should have noticed disagreed) — all four, run backwards, against a failure you can now see clearly because it already happened.
Why this matters now#
Know Your Cat asks for at least one documented failure and the repair that followed it — and the marking language is explicit that a repair is judged by whether the model itself visibly changed afterward, not by how well the apology reads. What you practise in this week’s Care Lab, Response Autopsy, and next week’s, Repair Clinic, is exactly this: diagnosis that ends in an update, not just a nicer sentence.
Outline#
- why apologising and replacing a response is a reset, not a repair
- diagnose, acknowledge, repair, update, and why the order matters
- why the update step is the one a repair is actually judged on
- what a small, honest model update looks like in practice
- repair as the semester’s earlier ideas, applied to a failure you can now see clearly
Teaching team
- Marisol Quaye — convenor