**1. Precise restatement** Model: reading Y = θ + E, with E a random variable whose distribution does not depend on θ (state-independence), E[|E|] < ∞. A "reporting scale" is a strictly increasing f: ℝ→ℝ. "Mean calibration on scale f" means E[f(θ+E)] = f(θ) for every θ ∈ ℝ. "Calibration within tolerance ε" means sup_θ |E[f(θ+E)] − f(θ)| ≤ ε. "Arbitrarily nearly linear" means sup_x |f(x) − x| and sup_x |f'(x) − 1| can be made as small as desired. Claim: (i) there exist two such scales f₁, f₂ (as close to the identity as desired) such that exact calibration on both forces E = 0 a.s.; (ii) for those same f₁, f₂ and any ε > 0, there are error distributions E calibrated within ε on both with Var(E) as large as desired. Hidden assumptions: additivity; independence of E from θ; finite first moment (so E[Y] exists); calibration required for all θ ∈ ℝ (an unbounded range — essential, since the Fourier argument needs linear independence of 1, sin θ, cos θ over the whole line); the specific incommensurate choice of frequencies 1 and √2. **2. Step check and independent derivation** Monotonicity: f₁' = 1 + a cos x ≥ 1 − |a| > 0 for |a| < 1; f₂' = 1 + a√2 cos(√2x) > 0 for |a| < 1/√2. |f_j(x) − x| ≤ |a|, |f_j'(x) − 1| ≤ √2|a|. So "arbitrarily close in value and derivative" is correct. Exact calibration on f₁: E[θ + E + a sin(θ+E)] = θ + a sin θ. Using sin(θ+E) = sin θ cos E + cos θ sin E (bounded, so expectations exist): E[E] + a(E[cos E] − 1) sin θ + a E[sin E] cos θ = 0 for all θ. Since 1, sin θ, cos θ are linearly independent functions on ℝ, E[E] = 0, E[cos E] = 1, E[sin E] = 0. Because cos E ≤ 1 pointwise, E[cos E] = 1 ⇒ cos E = 1 a.s. ⇒ E ∈ 2πℤ a.s. Same for f₂: E ∈ (2π/√2)ℤ = √2πℤ a.s. If 2πm = √2πk with integers, then k = √2 m, impossible unless m = k = 0. So E = 0 a.s. The derivation is correct; the text's statement that convexity plays no role is accurate (neither scale is convex). Tolerance part: E = ±2πn w.p. ½. E[E] = 0, E[sin E] = E[sin √2E] = 0 by symmetry, E[cos E] = cos 2πn = 1, so f₁-calibration is exact. On f₂ the discrepancy is a(cos(2π√2 n) − 1) sin(√2θ), with sup over θ equal to |a||cos(2π√2 n) − 1|. By Dirichlet's approximation theorem there are infinitely many n with |√2 n − m| < 1/n for some integer m, so |cos(2π√2n) − 1| ≤ 2π²/n² → 0 along such n. Var(E) = 4π²n² → ∞. Correct. **3. Attempts to break it** - Could the Fourier matching fail because θ ranges only over a bounded interval? Yes — if calibration is only required on a finite interval, the claim's exact-calibration argument still works (1, sin, cos are linearly independent on any interval of positive length), so no problem. - Does one need a ≠ 0 only? Yes; any nonzero a works, including arbitrarily small ones. No lower bound on a needed. - Commensurate frequencies (e.g., 1 and 2) would *not* give the result (E ∈ 2πℤ survives), so the incommensurability is essential — a construction detail, not a flaw. - Numerical sanity check: n = 5, √2·5 = 7.0711, cos(2π·0.0711) ≈ 0.9028, error ≈ 0.097a. n = 12, √2·12 = 16.971, cos(2π·(−0.0294)) ≈ 0.983, error ≈ 0.017a. n = 29, √2·29 = 41.012, error ≈ 0.003a; Var = 4π²·841 ≈ 33,200. Trend confirmed. - Is "noiseless" overclaimed? E = 0 a.s. is exactly noiseless. Fine. I find no gap; the argument is complete. **4. Novelty** The mechanism is standard: E[f(θ+E)] = f(θ) for all θ with f containing a Fourier mode e^{itx} forces the characteristic function φ_E(t) = 1, which (classical lattice-distribution fact) means E is supported on (2π/t)ℤ; incommensurate t's intersect only at 0. The tolerance instability is Dirichlet/Kronecker approximation. The qualitative moral — that exact mean-unbiasedness across nonlinear re-expressions forces degeneracy — is familiar from the Jensen/convexity argument (unbiased on x and x² ⇒ zero variance). This is a neat non-convex variant, but a special case of known characteristic-function facts rather than a new result. **5. Fairness** The sentence is a fair summary. Small caveats: the result depends on a deliberately engineered pair of scales with incommensurate oscillation frequencies (not generic "nearly linear" scales), on calibration being demanded for all θ, and on the additive independent-error model with finite mean. The text makes all of these explicit, and "can force" appropriately signals existence rather than generality. {"verdict": "correct", "novelty": "special case or variant of known results", "fair_summary": "yes", "confidence": 5, "one_line": "The argument is valid: exact calibration on x + a sin x and x + a sin(√2x) forces φ_E(1)=φ_E(√2)=1 hence E=0 a.s., while ±2πn errors with √2n nearly integral give arbitrarily good tolerance calibration with divergent variance; it is a clean application of standard characteristic-function and Diophantine facts rather than a new result."}