Goodhart Capture — When A Measure Becomes A Target It Stops Measuring
conceptAI & Compute · Biotech & Transhumanism
When a measure becomes a target, it stops being a good measure.
Who they are
Goodhart capture — vocabulary the engine was missing entirely[1].
What they do
A search of all eight canon files returned zero hits for Goodhart. The engine holds several examples of this without ever naming it.
How it works
Health questionnaires that can't catch an incubating infection. An alignment equation whose terms the measured system writes itself. Benchmark scores climbing while the underlying property doesn't[1].
Why it matters
Naming it means those cases can be pointed at without re-explaining the mechanism each time. It isn't universal — it needs pressure to optimise against the proxy.
The engine's record — word for word
Engine vocabulary the map did not carry: a word-boundary search of all eight canon files on 2026-09-13 returned zero hits for Goodhart. The mechanism is proxy-target divergence under optimization pressure — formalised in the 2026 evaluation literature as a Goodhart-inspired Instability Decomposition[1]. GENERALISES BEYOND AI, and the engine already holds instances without the name: a paperwork-screening regime that cannot detect an asymptomatic pathogen (bst, Report #84); an alignment equation whose terms are self-authored by the measured system (love_equation_bst); benchmark scores rising while the underlying property does not (evalsafetygap_proxy_target_divergence). Recorded as vocabulary so those instances can be named without re-deriving the mechanism. Held; no claim that every measurement regime is captured — the divergence requires optimization pressure on the proxy.
Follow the trail
Walk this on the live map →