Definition
Plain language
A rule that lets a self-improving AI keep a change to itself only if it doesn't break any task it had already gotten right.
As stated in the literature
A no-regression acceptance gate in self-evolving harness systems (HarnessX) that admits a candidate edit only if it improves the target metric without regressing any already-solved task; framed as the defense against catastrophic forgetting, though it can miss slow sub-threshold erosion that only flips tasks all at once.
Also called: seesaw/no-regression constraint
Why it matters: It defends against an agent improving on one front while quietly forgetting how to do things it used to handle, though it can miss slow erosion until many tasks fail at once.
For example, a self-improving system accepts a tweak only after confirming it didn't break any of the tasks the agent had already been solving correctly.
Heard on the show
“So the gate enforces what they call the seesaw constraint.”Episode 147 — Agents Fail at the Body, Not the Brain: A Self-Rewriting Scaffold That Lifts a 9B Model 44 Points