Chapter 20

Reflection

When a response fails a judge, the judge’s grounded critique can go back to the producer for a capped retry. Holonograph grants that per judge, rubric, facet and surface, and only after that exact cell has certified. It is never a default and never an operator toggle.

Why it is gated

Reflection turns a judge from an evaluator you read into an optimization target the producer chases. That inverts the trust requirement. A judge you only read can be a little wrong, because you flag and a human reviews. A judge you optimize against has to be right, because the producer will find and exploit every gap in it.

The gate is on validity, not on discrimination. Those come apart: a judge can score stably and still be blind, saturated, or wrong about how a construct should treat absence. Scoring the same way every time is reliability, and reliability does not imply validity.

The specific hazard

Real dimensions are often requires-confirmation: an honest "I do not have the data" correctly scores low. Put that judge in a retry loop and the cheapest way for the producer to raise its score is to stop being honest about the absence and fabricate. That is the exact failure the apparatus exists to catch, now with a loop optimizing toward it.

So reflection without established authority is not merely a waste of tokens and time. It is a pressure applied to the one behavior you most need preserved.

What has to hold

Reflection unlocks for a judge × rubric × facet × surface cell when all of the following hold. The gate is dual: Holonograph computes the statistical pass, then a human confirms.

• Soak. Enough live production data to reach a statistical N. Holonograph determines the statistical pass; a human approves the N, judging whether the sample spans the breadth of real inputs and outputs.
• Discrimination. The facet's dimensions pass discrimination evaluation, with a human confirming key outlier calls and a statistical sampling of calls on that facet.
• Saturation cleared. Every dimension passes the saturation audit.

Validity is cell-specific, so the grant is too. A judge certified on one facet of one surface has said nothing about another.

What it does

A failing response returns to the producer with the judge's grounded critique attached, and the producer retries, capped. Reflection is the soft tier of the judge's authority over delivery: it can send the producer back, it cannot block. Blocking remains deterministic code in the gate, or operator policy.

Because a verdict has to be in hand before the producer finishes, reflection is inline by construction. An asynchronous judge observes and nothing more.

Revocation

The grant is revoked on any change to the identity it was computed on. A new judge model, a revised rubric, a changed context scope: the cell certifies again or it does not reflect. This is the authority machinery applied to the judge rather than to the producer, and it suspends the same way for the same reasons.

Reference

Your own reflection loop remains yours to run; nothing here prevents it. The claim is narrower and worth stating plainly: a retry loop against an uncertified judge spends tokens to apply pressure in an unknown direction. Holonograph turns reflection on when the cell has earned it, if you have asked for it.

The taxonomy that places reflection among the judge modes is documented ahead of the implementation, as is this chapter.