Who Owns the Reading? Feedback Carriers from Scarcity to AGI Graduation, a Learnability Inequality, and Common-Mode Failure — seven preprints read together: one principle, two propositions, a set of corollaries and registered predictions

Abstract

Abstract Seven preprints each found a failure: hunger taken away by abundance, violations moving into hiding, a price set by the environment, time curated by others, an observed object adjusting itself to observation, kinship absent between humans and AGI but present among AGI copies, and snapshot indicators that look healthy up to the fold. This paper argues that these are instances of one thing: the coupling between a reading and a restoring force is cut, or ownership of the reading passes to a party that does not bear its consequences. From this it states a reading-ownership principle and gives two propositions and one closed form. Proposition 1 (the learnability inequality) concerns identifying the sign of a governance lever before contact. With W the number of learnable warnings, h the hidden share and q the jointly perceived share, the effective window is W_eff = (1 − h)·q·W; with I_d the distinguishability per incident, Wald's bound sets the number of warnings needed at k* = N_W/I_d, where N_W = 5.24 at error probabilities of 0.005. If k* warnings are needed, the probability of learning first is (W_eff/(W_eff + 1))^k*; when S_warn = W_eff − k* ≤ 0 it is at most one half, tending to 1/e, and a probability of 0.95 needs W_eff ≈ 19.5·k*. An effect of 0.2 standard deviations needs about 262 incidents, and a window of 29 lets only effects of at least 0.60 standard deviations be identified. When incident severity has a tail index κ★ ≤ 2, sums have infinite variance: at κ★ = 1.5, comparing means needs between 2000 and 5000 incidents per arm for a 0.95 probability of the correct sign, exceedance counts about 500 and medians about 50; for a proportional intervention the distinguishability of a sum is proportional to R/(1 − R), where R = (E s)²/E s² is exactly the moment ratio that sets the maintenance floor, and both lose their scale-free form at κ★ = 2. Proposition 2 (common-mode failure): with 100 same-origin instances each failing with probability 0.05, a latent correlation of 0.3 leaves the expected number of simultaneous failures at 5 but raises the probability that at least half fail together from 7×10⁻³⁸ to 1.7×10⁻³, and lowers the effective number of independent instances to 9.4; the fleet-level warning window falls from practically infinite to 114, and the probability of collecting 20 warnings before a fatal fleet event falls to 0.84 (0.33 at correlation 0.5). If a monitor's misses load on the same origin factor as the violations it watches, its miss rate given a violation rises monotonically with the shared correlation — to 0.65 at 0.6, against an unconditional 0.10. The Gaussian one-factor model has no tail dependence and errs on the optimistic side: under a t form with four degrees of freedom the fleet window at correlation 0.3 falls to 11.5. The two propositions together give a structural conclusion: when a sign cannot be learned within the window and prior evaluations can be performed by the observed party, the admissible tools are those whose sign is fixed by construction — resource gating, sequestration of replication, compartmentalised interfaces, heterologous surveillance, external switches, and the separation of capability from permission. Eight further corollaries, six prohibitions and seven registered predictions follow.