The Shadow Theorem

We introduce the Shadow Theorem, a family of spectral bounds on the ability of a lower-capacity agent to estimate the capabilities of a higher-capacity agent. The core result is an application of the Eckart–Young theorem to intelligence estimation: if agent \(A\) can represent \(R_A\) independent capability dimensions and agent \(B\) can represent \(R_B > R_A\), then \(A\)'s optimal model of \(B\) has irreducible error at least \(\sum_{k=R_A+1}^{R_B} \sigma_k^2(B)\), where \(\sigma_k\) are the singular values of \(B\)'s capability operator. We prove four structural results: (1) the Shadow Bound on estimation error, (2) the Domain Blindness Theorem showing \(R_B - R_A\) capability distinctions invisible to \(A\), (3) the Observation Saturation Theorem showing the error floor is independent of observation time, and (4) the Calibration Impossibility Theorem showing \(A\) cannot accurately assess its own estimation error about \(B\). We validate all four predictions empirically using the Latent of Latents framework (Nagy, 2026) applied to four models: GPT-2 Small (\(R = 10\), \(d = 768\)), GPT-2 Medium (\(R = 6\), \(d = 1024\)), GPT-2 Large (\(R = 31\), \(d = 1280\)), and TinyLlama 1.1B (\(R = 31\), \(d = 2048\)). Rank-10 reconstruction of TinyLlama's 41-domain knowledge structure loses 33.7% of variance; domain pairs undergo complete sign reversals (psychology–law: \(+0.14 \to -0.96\)) at low rank; the cross-architecture estimation asymmetry is 83-fold; and at every rank below full, 100% of the estimation error is invisible to the estimating agent. The GPT-2 family traces the complete Unification Pulse: \(R = 10 \to 6 \to 31\) as parameters increase from 124M to 774M, with the depth signature \(\delta\) confirming this trajectory is genuine (unification at Medium: \(\delta = 0.038\); re-discrimination at Large: \(\delta = 0.247\)). This provides direct empirical evidence for the Unification Paradox: a phase-2 (discriminating) observer cannot distinguish a phase-1 (naive) system from a phase-3 (unified) system, because simplicity and profundity project to the same shadow. We discuss implications for AI safety, where the simpler agent is human and the more complex agent is a potential superhuman AI.