Replication package for: Response Conductance and Decision Distance — A Cross-Domain Theory of Endpoint Loss

Replication package for the SSRN working paper … (Tanaka 2026, SSRN abstract_id=6741738). Endpoint loss arises when a system's actions track signals other than the ones that capture what its principals or beneficiaries value. We introduce Response Conductance Theory (RCT) as a response-conversion framework: information becomes system-relevant only when it changes action, and harm depends on the joint presence of optimization pressure M, proxy-endpoint divergence DPE, proxy response dominance $K = [\log(G_P/G_E)]+$, weak guardrails $H_d$, and exposure $V_d$, combined through a domain-specific aggregation function $L{E,d} = \phi_d(M, D_{PE}, K, H_d, V_d)$. Decision Distance Theory (DDT) is the institutional special case, where GE depends on price, exit, and voice, GP on intermediary objectives, and the removal test identifies whether an entity raises GE/GP. We test the framework at three layers. Institutional scoreboard: across 22 domain-outcome pairs spanning six independent families, removal-test-satisfying proxies predicted the correct direction in 13 of 13 preregistered tests; within-country panel evidence in tax administration (β=&minus;0.138), healthcare (β=+0.157), and judicial enforcement (β=+0.086, rising to +0.190 at lag 5) confirms direction; orthogonal-shock tests (24/24) show that GDP-controlled estimates over-attenuate. Catalog tests: blinded rater replications on 27 institutional and 15 AI alignment cases reject any universal multiplicative score R=M⋅DPE⋅K, but reject it in different ways---institutional severity is dominated by DPE (pooled LOO R2=0.97); AI severity is dominated by M (pooled LOO R2=0.90). Boundary tests: the universal-curve form is rejected; cross-country GE/GP averages do not predict β; expert protocols for conductance ratios fail (three attempts, all ρ<0.2); the discriminator cell required to distinguish plain Goodhart from RCT is empty in both catalogs. The catalog evidence supports a universal component grammar---each component is jointly relevant across domains---but rejects a fixed multiplicative aggregation; identification of ϕd is reserved for decision-level data where each component can be measured independently of realized loss. Systems optimize what they can hear.