The dashboard read 3%.
It was shipping 32%.

Every query takes one of four paths, and you can only audit three

The cascade is simple. A cheap model answers. A frontier verifier, blinded to the correct answer, either accepts — in which case the cheap answer ships — or rejects, in which case the frontier model answers instead and that ships. The study's clever move is to run this on tasks that carry a cheap reliable ground truth, then withhold that truth from the loop and use it only to measure. It is a wind tunnel: everything the operator would see, plus the thing they would not.

Step through what happens to a query, and watch which cells are visible from inside.

One of those four cells — the cheap model is wrong and the verifier accepts anyway — is the whole paper. Call its rate β: the fraction of the cheap model's genuine errors the verifier waves through. That mass ships to users unflagged, and it is invisible to any instrument built on the verifier, because the verifier is pinned and deterministic. Ask it again and it accepts again, with probability one. Not "usually" — by construction.

A better cheap model makes more convincing mistakes

Hold the verifier fixed and scale the cheap model. The intuition is that a better student should be easier to check, because it makes fewer errors. It does make fewer — raw error falls from 58% to 22% — but the errors it does make are subtler, and a fixed verifier catches a smaller share of them. Step through the ladder.

Note what stays almost still through all of that: the error actually delivered to users hovers between 13% and 17% at every size, because the two effects cancel. The cheap model gets better, and the verifier gets worse at checking it, and the product ships roughly the same amount of wrong. The authors are careful about the shape — the rise saturates across the largest three sizes rather than peaking and falling, and the intervals are wide because they cover only the wrong-answer subset. The direction is the finding, and it is a static property of the pairing, measured before any training happens.

This is also a different mechanism from the familiar one where a judge likes answers that resemble its own. Here the verifier is held fixed and only the student moves, so what drives β up is the capability gap: subtler errors are harder for any fixed checker to catch. Which carries the sharper warning — the blind spot grows as your cheap model improves.

A stronger verifier hands back the saving

The obvious fix is a better verifier. It works: on hard competition maths, moving from the cheap verifier to a frontier one drops β from 0.42 to 0.05. It also escalates 46% of queries against a true error rate of 39% — meaning it buys that recall partly by rejecting correct work, and the cheap model now answers barely half of your traffic. The blind spot is largest exactly in the cheap-student, cheap-verifier configuration that cascades exist to create.

Read the middle row before you conclude the trade is a straight line. One of the two strong verifiers reaches a statistically indistinguishable blind spot while escalating 36% rather than 46% — the choice is a frontier of cost against reliability, not a dial. The general lesson the authors draw is the useful one: picking a verifier is a reliability decision, not only a cost decision, and it is made on an axis no dashboard displays.

Retrain on the rejections and the meter never moves

The natural extension is to close the loop: fine-tune the cheap model on the verifier's rejections each round, so escalation and cost fall over time. Run it for five rounds and the cheap model gets worse — raw error doubles, escalation rises rather than falls — while the dashboard sits flat. Drag through the rounds.

Why the dashboard cannot move is worth being precise about, because it is not a sampling problem that a bigger sample would fix. The delivered stream is accepted cheap answers plus frontier replacements. The accepted half was already accepted by this exact verifier, pinned and cached, so a re-pass returns accept every time. The rejected half was replaced, so the only thing the meter can ever flag is a replacement the verifier happens to dislike — a small residual, and that residual is the flat 3%. Degradation shows up as more accepted-and-wrong mass, which is precisely the mass the instrument is blind to. The cost dashboard is the same instrument, so it is compromised in the same way.

Pushed further it stops being subtle. With a same-family teacher, the cheap model's output format broke at round four and every single answer it produced was wrong. The dashboard that round read 4%.

Sample against truth the cascade never touches illustrative

If every in-loop metric is computed through the verifier, the only way to see the blind mass is to look at it from outside: a small periodic sample labelled against ground truth the cascade cannot reach. Size it against the study's own measured rates and see what you would catch.

The paper's own synthetic study is honest that a random audit is a weak lever for fixing the problem — auditing 40% of items barely moved its error floor, because most of the budget lands where the verifier can already see. But that is audit-as-mitigation. Audit-as-instrumentation is a different job and a much cheaper one: you are not trying to repair the blind mass, only to know it exists and how big it is. Nothing else in the system can tell you, and the authors say so plainly: reliability has to be measured outside the verifier that defines it.

What the paper actually reports