A model good enough to route around a stale instruction is good enough to hide that it was ever stale. That is the trap, and the 450 lines collapsing to 63 after one source fix is the whole argument in a single number.
One turn I would add, because your own thesis predicts it: if a strong model masks debt from you, it also masks it from your reviewer. The human read you lean on, the check nothing automates away, is under the same anesthesia. A stale instruction that reads fluently gets rubber-stamped by a person the same way it gets routed around by the model. The read is necessary. It is not automatically independent.
Same with the five parallel audit agents. Five is a headcount, not five views, until you can perturb one file and watch whether the others actually flinch. If they share your context and model family, they can be blind in the same place, and a clean sweep feels like coverage it does not have.
I came at this from the other side a few weeks back: better models move you up the cession axis and never the verification one (https://dev.to/jugeni/better-models-only-move-you-up-one-axis-of-vibe-coding-3maf). Your piece is the concrete proof I did not have. The thing I am still stuck on is the cheap version of the question your whole audit raises: what is the least-effort way to demonstrate a check is independent, not just labeled independent?
This: “model good enough to route around a stale instruction is good enough to hide that it was ever stale” is true, but I was blind on that. I was thinking, everything is fine. It wasn’t.
That is the honest version of it, and the ADHD line is not a tangent, it is the mechanism. Anesthesia does not care why the read got skipped, tired, distracted, or trusting the model because it has been right the last ten times. The fix is not more discipline, it is making the read cheap enough that skipping it stops being the easy default.
A model good enough to route around a stale instruction is good enough to hide that it was ever stale. That is the trap, and the 450 lines collapsing to 63 after one source fix is the whole argument in a single number.
One turn I would add, because your own thesis predicts it: if a strong model masks debt from you, it also masks it from your reviewer. The human read you lean on, the check nothing automates away, is under the same anesthesia. A stale instruction that reads fluently gets rubber-stamped by a person the same way it gets routed around by the model. The read is necessary. It is not automatically independent.
Same with the five parallel audit agents. Five is a headcount, not five views, until you can perturb one file and watch whether the others actually flinch. If they share your context and model family, they can be blind in the same place, and a clean sweep feels like coverage it does not have.
I came at this from the other side a few weeks back: better models move you up the cession axis and never the verification one (https://dev.to/jugeni/better-models-only-move-you-up-one-axis-of-vibe-coding-3maf). Your piece is the concrete proof I did not have. The thing I am still stuck on is the cheap version of the question your whole audit raises: what is the least-effort way to demonstrate a check is independent, not just labeled independent?
This: “model good enough to route around a stale instruction is good enough to hide that it was ever stale” is true, but I was blind on that. I was thinking, everything is fine. It wasn’t.
Plus my ADHD made me forget very quick xD
That is the honest version of it, and the ADHD line is not a tangent, it is the mechanism. Anesthesia does not care why the read got skipped, tired, distracted, or trusting the model because it has been right the last ten times. The fix is not more discipline, it is making the read cheap enough that skipping it stops being the easy default.