Which reward hacking defenses actually transfer across training substrates?

OPEN. The abstract closes with: "We also map representative defenses across substrates, identifying which mechanisms transfer directly and which offer only functional analogies." The conclusion ends: "For practitioners, the defense correspondence offers the most immediate use."

What the excerpt gives is that a map exists, that it sorts defenses into two kinds (direct transfer, functional analogy), and that the authors rate it the most useful part for practitioners. It does not give a single defense, the substrate each comes from, or the criterion that separates transfer from analogy.

Why it matters here. The vault holds defenses at every substrate but has never sorted them that way: on weights, Can debate training prevent reward hacking by weaker judges? and Can counterfactual invariance eliminate reward hacking biases?; on text, How can agent self-evolution be made safe and auditable? and Does constraining edits make skill learning more stable?; at the evaluator, Can deterministic checks protect LLM judges from failure?; where selection is the step, the nearest is What exactly does hidden mean in AIDE2's evaluation system?, a loop that keeps the rewrites scoring best on hidden evaluations, with what they are hidden from undefined. Whether the hidden partition, for instance, counts as a direct transfer to training and to selection or as an analogy is the kind of question the paper's map would answer, and the vault cannot answer it from the excerpt.

What would settle it. Read the paper's defense section. Short of that, sort the vault's defense notes by the substrate they act on, and for each ask whether a counterpart exists on the other two (this is my method, not the paper's).