Phronetic adjudication

Instead of collapsing a trade-off into a single number before the question is ever asked, these systems produce, for the particular situation in front of them, an explicit account of which values bear on it and which way each one cuts. Kaleido stops at the list, naming the values, rights and duties in play and whether each supports or opposes the action; chain-of-thought training on value conflicts carries the same material through to a verdict; the automated parliament splits the work across several delegates who each speak for a moral perspective, rewrite one another's answers, and then settle on one together. Because the considerations are set out rather than hidden in a score, they can be read back at scale: sorting hundreds of trolley-problem traces by the kind of reason given shows models reasoning from duties in the chain of thought while explaining themselves by consequences afterward.

Rather than committing to one ethical theory, the system holds several and settles their disagreements with a rule written down in advance: split your confidence across the theories and let those weights shape what a learning agent is rewarded for, or treat the theories as parties bargaining over lotteries and take the bargaining solution. Only the theories' verdicts on the case are inputs; the rule that combines them does not change from case to case. That is the deliberate opposite of open-ended judgment, and the point of it — the answer is predictable and can be argued about before any particular case arises.

These papers argue that what is missing is not a procedure to bolt on at answer time but a trait the model does not have, and that training should aim at the trait itself. One holds that fine-tuning installs surface behavior — refuse this, comply with that — and never the ability to reason when those instructions collide, which is exactly the seam adversarial prompts pry open. The other holds that the harder gap is knowing which approach a problem with no settled method calls for, and sketches how that kind of judgment might be benchmarked, trained and built. They name different traits, and neither has been built, but both make the same move: change what training installs rather than the shape of any one answer.