Blinding amplitude: choose the blind's prior range #407
Description
Smokescreen blinding is merged (#253). Under a test blind (ΔS8 = −0.060, ΔΩm = −0.043), v1.4.6.3's ξ± moves by up to about 2.7σ per bin, which may be large enough to be obvious. Choose the blind's prior range so that a blind hides the result without standing out. Decide it on the full GLASS mock set.
The CMBX blinded-mock rehearsal is relevant here. Blinded posteriors match a translation of the unblinded ones only approximately: the residuals reach 0.75σ, and the width changes. That is expected, since an additive shift in the data vector isn't a translation in parameter space (the Jacobian ∂t/∂θ changes). So the round-trip test should check that the posterior is unbiased with a sensible width, not that it is an exact copy shifted over.
Claude Opus 5.5 on behalf of Cail. From the 3x2pt telecon, 6 Oct 2026.
Activity
cailmdaley commentedon Oct 30, 2026
Checked on a catalogue-painted GLASS mock: v1.4.6.3 positions, weights and rotated shape noise, the B n(z), seed 1. A mock blind with ΔS8 = +0.055 (ΔΩm = −0.019), comparable to the original ΔS8 = −0.060 test blind, was applied to a byte copy of the mock, and both copies went through cosmo_val.
The shift is a smooth, coherent change in amplitude. Next to the clean mock it is obvious, but on the blinded points alone nothing stands out against the scatter. Cail's decision (30 Oct) is to keep the current envelope, so I'm closing this.
Whether a blinded inference round trip recovers the truth (unbiased, sensible width) is the next block of the mock tests.
Claude Opus 5.5 on behalf of Cail.