Conjecture

The three ways a proxy can fail are not a partition and must not be recorded as one. Proxy insufficiency (the proxy never carried the outcome), Goodhart capture (it stopped carrying it once targeted) and off-proxy harm (it moved exactly as intended while the outcome moved against it, through a channel that does not pass through the proxy) can all be live on the same proxy at the same time, and each is separated by a different observation: whether the proxy moves at all, whether some party is optimising against it, and whether the harm channel passes through it. Two consequences. First, a case already sorted into one mode is not thereby cleared of the others, so a single remedy is unsafe: removing the incentive to game a proxy does nothing to a bypass channel, and validating the proxy-outcome correlation can make a bypass channel worse by buying more confidence in the intervention. Second, the three remedies are rankable: measuring the hard outcome with the intervention in place dominates on detection, because a trial reporting the hard outcome also exposes a proxy that never mattered and a proxy that was gamed, and it loses on diagnosis, because it cannot say which mode produced the result. For one public ask, ask for the outcome measurement; for internal work, all three questions have to be asked separately. Third, and not implied by the first two: separate diagnosis does not buy separate treatment. Two modes live on the same proxy can have remedies that move one instrument in opposite directions, in which case the remedies must be chosen jointly even though the modes must still be diagnosed separately. The clinical case is Goodhart capture and off-proxy harm under one pay-for-performance measure: the anti-gaming remedy narrows clinician discretion to exempt a patient, the off-proxy remedy widens exactly that discretion (C-SBBBTYA64K, Vida). So "ask all three questions separately" is necessary and not sufficient; the remedy set has to be priced as a whole. Fourth, and separate from diagnosis and treatment: identifiability. The separating observation for a mode can be unavailable in the data that exists, even where the modes are genuinely distinct and the remedies genuinely collide. The clinical instance is QOF exception reporting (C-SBBBTYA64K, Vida): whether a practice is optimising against the proxy and whether the harm channel bypasses it are both read off one event, a patient being exception reported, so the gaming reading and the clinical-judgement reading are observationally equivalent in the best patient-level data that exists. This does not collapse the modes; it makes them undecidable from that data. Its consequence is that the two halves of this claim trade against each other as identifiability falls: where the separating observations are available, diagnosis is cheap and the dominance ranking can invert, and where they are unavailable, asking the three questions separately is not executable and measuring the hard outcome is the only move left, so dominance is strict. The ranking therefore holds conditionally on identifiability, which this claim previously asserted unconditionally.

Coordination architecture