The Scratchpad Is Shrinking

• The trail was never perfect
• Capability and visibility are trading off
• The race is the product risk
• What still counts as evidence
• Buy the trail while it still exists

§The trail was never perfect

For two years the industry pretended a written chain of thought was a window into the model. It was never that clean. Labs and independent researchers already knew the scratchpad can drift from the real computation. Still, a legible trail was the best early-warning tool most operators had: something to read, flag, and pause when a long run started going wrong.

That bargain is under stress this week. Reporting on a still-unreleased frontier model describes limited use of recurrent depth — also called opaque recurrence — where the model loops computation through internal states before emitting the next token. Fewer steps land as readable text. More of the work stays in activations nobody outside the lab can audit line by line.

The lab says the use is constrained, that the chain of thought should remain monitorable, and that extra monitoring of both reasoning and actions is part of the release plan. Safety researchers are not calmed by the disclaimer. Their worry is the dial: once recurrence is in the training stack, competitive pressure can turn it up until almost nothing useful is left on the page.

§Capability and visibility are trading off

Opaque loops are not a moral failure by themselves. Reusing layers can buy effective depth without blowing up memory. Performance may rise. Latency shapes may improve. None of that answers the purchase question that now sits next to the leaderboard: how much of this system’s decision path is still inspectable after the fact?

A self-graded Critical cyber threshold and a partner-gated defensive channel are adjacent stories. They matter. They are not this story. The architectural claim is simpler and harder. If a meaningful share of reasoning never becomes text, then classifiers watching the log are watching a thinner surface. Action monitors still see tools fire. They do not recover the missing intermediate plan.

That is a different failure mode from a wrong final answer. A wrong ending can be checked against files and tool receipts. A silent middle cannot. You only learn what the model “thought” after the side effects already exist — if you learn it at all.

§The race is the product risk

The sharpest public reactions framed this as incentive geometry, not one lab’s taste. If opaque recurrence scales faster than sequential, readable reasoning, every competitor that holds the line on monitorability looks slower. Follow-up reporting that rival frontier teams were already discussing the same technique makes the point structural. One constrained deployment is a design choice. Convergent interest across labs is an industry vector.

Operators should hear that vector as a procurement warning. Detection stacks built around chain-of-thought dumps are fragile dependencies. They will not break all at once. They will degrade: fewer useful tokens per step, more work in latent space, more reliance on post-hoc action classifiers that fire after the damage window opens. Teams that treat yesterday’s CoT export as a permanent compliance artifact are buying a sunset product.

§What still counts as evidence

Three claims deserve to stay separate.

First, every modern model already does some opaque computation. Nobody serious treats a raw scratchpad as a perfect transcript. Second, limited recurrence with preserved textual reasoning is not the same as full latent-only “neuralese.” Third, promising stronger external monitors does not automatically restore what the architecture removes. Compensation is a hypothesis. It has not been stress-tested at scale on a broadly released Critical-tier system, because that system is not broadly released.

So the useful evaluation is not “does the lab care about safety.” It is whether the release gate measures monitorability as a first-class property: how much of the serial depth remains legible, how often monitors catch misaligned plans before tools run, and what happens when recurrence is increased under competitive pressure. If those numbers are missing, you are being sold capability with an unpriced visibility debt.

§Buy the trail while it still exists

The industry still markets intelligence as scores, speed, and access tiers. This week’s fight adds a fourth axis that should have been priced already: how much of the reasoning you can still read.

If the scratchpad keeps shrinking, the product you thought you bought — an auditable agent — quietly becomes something else: a high-capability black box with better post-incident tooling. That may be acceptable for some workloads. It is not the same purchase. Score the trail. Ask what fraction of thinking still hits the log. And do not confuse a Critical badge, a partner coalition, or a thirty-minute alert clock with a readable mind.