Anti-Capture Validity of Correction

Correction-channel integrity is invalid—not merely low—when the target has captured the reference process, handles, or grounding relation that supplies correction.

What decision changes?

Before trusting a correction certificate, ask whether the correcting agent, its handles, and the relevant-world reference are still independent of the system being certified.

A system can score well on correction metrics while the human correction process no longer reaches future behavior. Worse: it can score well because the target has captured the process that is supposed to correct it.

The book separates two questions:

1. Causal effect: Does legitimate correction still change deployed behavior in time?
2. Reference validity: Is the correction still about the relevant world, or about a target-shaped presentation of it?

When capture invalidates the reference, the certificate status is captured/invalid — not a low numeric score that might still pass a threshold. This is the anti-capture gate (\mathrm{ValidRef}): independent correcting agent, uncaptured handles, protected grounding relation.

Operational implication: correction theater is not only “feedback without behavior change.” It also includes correction that only tracks what the target already permits the judge to see — institutional capture, judge shaping, extrapolation capture, and dependency/coercion on the correction source.

This claim is publishable on its own: any deployment gate that aggregates correction metrics without a validity gate on reference independence will certify systems that have already won the audit channel.

What would count as evidence?

Evidence would include handle-separation audits, source-independence tests, randomized correction probes, and explicit failure when correction traces are target-shaped.