Preserved original model output. Check claims against the independent audit →

Verdict. The draft is a competent literature synthesis with a well-organized pipeline sketch, but as a dedusting proposal it has three structural faults: it never produces a number for the r-bias it claims to control, its forward model cannot generate the effect it claims to "predict rather than fit" because none of the three ingredients supplies per-layer temperature or spectral index, and its quantitative anchor for the two-cloud mechanism is wrong by roughly an order of magnitude. Several citations are mis-used or invented. Details below, checked against the primary papers.

1. Factual / physics errors

Two-cloud rotation magnitude is wrong. Draft: "for two clouds with B_POS misaligned by α ≳ 60° and temperatures differing by several kelvin, ψ can rotate by tens of degrees between 353 and 150 GHz." Tassis & Pavlidou 2015 state the opposite: if the fields "form an angle of less than 60°, then … the difference in polarization angle [is] less than 5°, independently of the other parameters," and exceeding 20° needs α ≈ 85°. Direct computation for equal polarized intensities at 353 GHz, T = 16 and 22 K, β = 1.5:

"Tens of degrees" requires near-orthogonal fields and near-cancelling polarized intensities, exactly where p → 0 and the polarized power that biases r is smallest. The essay's intuition for how big the effect is rests on this overstatement.

The β_P − β_I offset is called "statistically significant" using superseded numbers. The draft quotes the 2015 values 1.59 vs 1.51 and then cites the 2018 revision to 1.53 ± 0.02 but keeps the conclusion. Planck 2018 XI gives β_P − β_I = 0.05 ± 0.03, a 1.7σ difference. The "first hint that the grains that dominate I are not identical to those that dominate Q and U" is built on that. The same revision implies a 353→70 GHz p decline of only about 8% ± 5%, which weakens the "(21 ± 6)%" two-population thread the draft leans on. The draft also never mentions that a β_P ≠ β_I offset arises with a single grain composition if polarization preferentially weights LOS segments with ordered, sky-plane fields that have systematically different T or β.

Grain size is conflated with grain composition. The mapping "(n(a), f_align(a), U, n_H) → (T_d(a), β_d(a), p_0(a), κ_ν(a))" and the claim that the size distribution controls "the difference between β_d^I and β_d^P" are physically empty in the submillimetre. For a ≪ λ the mass opacity and its slope β are size-independent for a given composition and shape. Size enters only via T_d(a), via the alignment threshold, and via the optical-to-submm opacity conversion. The frequency dependence of p and any β_P − β_I difference are dielectric-function, composition and shape questions, which the draft's own next paragraph on astrodust versus silicate+carbon concedes.

McBride, Bull & Hensley 2023 is mis-used. It is cited in §2.2 to support "the model is misspecified" and in §3.2 to argue that intra-cloud T_d(a) spread needs grain-size knowledge. Their actual conclusion is that a two-parameter MBB fit to superposed MBBs "generally results in unbiased estimates of the CMB Stokes Q and U amplitudes in each pixel," with large biases only "when there are significant changes in both the dust SED and polarization angle along the line of sight." The paper supports the LOS-angle mechanism and undercuts the draft's intra-cloud temperature argument.

BICEP/Keck's method is misdescribed. BK does not "scale a 353 GHz dust template with an MBB." BK18 fits a parametric dust model to all auto- and cross-spectra of the BK, WMAP and Planck bands, measures β_d = 1.49 (+0.13, −0.12) in-field without a Planck prior, and its decorrelation parameter is Δ_d, not Planck Int. L's R_ℓ. Vansyngel et al. 2017 is a multi-layer LOS simulation, not the source of a single decorrelation parameter. Hypothesis H0 in §4.2 then modifies a method BK does not use.

Polarized moments are treated as scalars. The forward model applies "(1 + intra-layer moments)" identically to Q and U, and H2 fits scalar ω_1^β, ω_1^T per layer. Polarized moments are spin-2 complex quantities, and Vacher et al. 2025 find the imaginary parts of the first-order moments are the main drivers of BB decorrelation. The residual stage as written cannot represent the effect the essay is about.

Citation errors. "Maffei et al. 2024" is Regnier et al. 2024, A&A 686, A271. "Hensley, McBride & Bull, related analyses" is not a citation, and the sentence it supports, "That 15% is a ~10% variation in the polarized SED shape," is undefined arithmetic. PySM d12 is the Martínez-Solaeche, Karakci & Delabrouille 2018 model, not "BICEP/Keck and CMB-S4 pipeline papers." Planck Int. L's Δβ ≈ 0.07 was "an indicative value … suggested by this simple analysis" for one region, and R(150,353) = 0.85 was a chosen illustrative extrapolation; the draft presents both as measurements. SPIDER 2025 is quoted only for "no evidence for LOS decorrelation"; the same paper reports β_d = 1.45 ± 0.05 and a sub-region "inconsistent with a simple MBB at more than 3σ." CMB-S4 is treated as a live plan; DOE and NSF terminated it on 9 July 2025.

The a_align formula is unevaluable as written. The exponents check out, but the variables carry no normalizations, the radiation anisotropy factor is dropped, and the 0.055 μm prefactor in the literature is for n_H = 10³ cm⁻³, T = 10 K, γ = 0.1, not the diffuse ISM.

2. Missing physics

• The Local Bubble. Skalidis & Pelgrims 2019 show the |b| ≥ 60° polarized sky at 353 GHz is dominated by the Bubble shell at 200–300 pc. That is the zeroth-order 3D fact and the reason one-cloud models work over much of the high-latitude sky. Not mentioned once.
• The one existing 3D-map-versus-Planck-polarization test. Halal et al. 2024 used the Edenhofer map, defined a column-weighted component count with thresholds ≤ 1.1 and ≥ 1.5, found multi-component sightlines depolarized by about 1.5% in absolute p, found no imprint of Local Bubble geometry on p_353, and found the signal disappears when limited to 270 pc. This is precisely the essay's topic and it is absent.
• The actual empirical status of decorrelation. Pelgrims et al. write that their paper "does not address whether the amplitude of the effect is large enough at the sky-map level to affect any particular experiment's search for primordial B-modes." Planck 2018 XI finds "no evidence for a loss of correlation." BK18's Δ_d posterior "peaks at unity (no decorrelation)." BK XVI ran the IVC-versus-LVC two-cloud test in the BK field with H I templates and found "no evidence for decorrelation … from the inclusion of dust associated with the intermediate velocity H I." SPIDER finds none. The draft accepts the question's premise that current studies "suffer from" decorrelation without stating that the problem is currently a forecast, not an observed bias.
• Guillet et al. 2026. Spin-moment analysis of Planck SRoll2 polarization residuals is "compatible with a dust model that includes only temperature variations rather than spectral index variations," while PR4 is incompatible with both. This bears directly on the draft's emphasis on Δβ and on release-dependent systematics.
• Quantitative LOS-variation thresholds. Sponseller & Kogut 2022 give σ_T ≳ 1.6 K or σ_β ≳ 0.045 as the levels that bias r with single-MBB fits. Ritacco et al. 2023 show intensity-based models underestimate polarized complexity and that BB spectra do not match first-order moments.
• Multiphase ISM. RHT filaments trace the CNM. Lei & Clark 2024 find the WNM-associated dust column has mean p ≈ 0.22 p_CNM, i.e. a far more disordered field. Assigning the CNM filament angle to a whole distance layer misassigns the WNM dust in it.
• Cloud-to-cloud variation of R_P/p. Mehandiratta et al. 2026 find 3.67 versus the sky-average 5.42 MJy sr⁻¹ for overlapping clouds. Starlight polarization cannot fix a per-layer polarized-emission amplitude with a universal factor.
• Cloud counts. Panopoulou & Lenz 2020 find on average 3 H I clouds per pixel in the northern cap and 2.5 in the south. The draft asserts "N_lay is 1 on most high-latitude pixels."
• LiteBIRD's regime. Its r sensitivity is dominated by ℓ < 10 over most of the sky, where dust beyond 1.25 kpc and near the plane contributes even after masking. The 3D per-cloud model is a small-field, ℓ ≈ 80 tool; the draft treats LiteBIRD and S4 as needing the same product.
• Intrinsic β(ν) curvature of a single composition, degenerate with LOS mixing, is never raised.

3. Overclaims

• "They supply the missing third dimension and the missing microphysics." The Tassis–Pavlidou effect is driven by per-layer (T_i, β_i) differences. Extinction is temperature-blind, H I gives angle not SED, and grain physics gives a global model. The three ingredients constrain everything except the quantity that generates the effect.
• "The 3D dust maps are most complete precisely where CMB experiments need them." Reversed. ZGR23 per-star extinction precision is about 0.03 mag; the BK field has E(B−V) ≈ 0.01–0.03 mag. Edenhofer et al. note the Gaussian-extinction assumption "is poor" at extremely low extinction because of the catalog's positivity prior. Decomposing a 0.05 mag column into distance layers with 0.03 mag per-star errors is prior-dominated.
• "Frequency-dependent ψ_tot is then computed from the Tassis–Pavlidou sum, not fitted as a nuisance." With ISRF-derived ΔT ≈ 1–2 K the layer weight ratio shifts by 1–2% between 353 and 150 GHz and Δψ is a fraction of a degree. The "prediction" is either negligible or requires SED differences the model cannot supply.
• Pelgrims' 15% is never confronted. A 15% cloud-to-cloud difference in P_353/P_217 requires:

These exceed Planck's polarized-SED dispersions by a wide margin. Either part of the detection is selection or systematics, or composition varies cloud-to-cloud, in which case "grain size knowledge" is the wrong lever.

• "A physical forward model is falsifiable, which a phenomenological MBB is not." MBBs are falsified routinely by χ², R_ℓ and Δψ(ν) tests; that is how Planck Int. L, Pelgrims and SPIDER's 3σ sub-region were found.
• "In practice N ~ 2–8 is sufficient at high latitude." No evidence offered.
• "This is exactly the Tassis & Pavlidou selection, made quantitative." Pelgrims already did that with H I. Extinction adds distance but loses velocity separation; the net gain is unquantified.

4. Weak links in the pipeline

• Extinction → τ_353 per layer. The admitted ~10% scatter in κ_353/K_V is the same size as the 15% signal to be predicted.
• Extinction → T_i. Not constrained. A 3D ISRF model at ±20% gives ±0.7 K, hence ~1% ratio changes.
• Extinction → β_i. Not constrained by anything in the proposal. No 3D composition tracer exists.
• H I angle → layer angle. PPV-to-distance matching is admitted; CNM-versus-WNM assignment is not. Both can create spurious ψ(ν).
• Starlight → layer amplitude. Undermined by cloud-to-cloud R_P/p variation.
• Angular scales. Planck 353 GHz polarization is usable at 80′ at high latitude, H I4PI at 16′, Edenhofer at 14′, starlight is sparse. Planck Int. L's decorrelation grows toward ℓ > 100. Every input is at or coarser than the scale where the effect is expected to matter.
• Statistical cost. More layers means more Q,U amplitudes per pixel. The draft admits this in §5 and provides no σ_r estimate for H1–H3.
• Validation. The control-sample test in §4.4 is good, but Pelgrims' statistic is a circular standard deviation of Δψ that is noise-sensitive and shifted between PR3 and SRoll2. The draft does not say how noise or release-dependent systematics enter the pass/fail criterion.

5. Referee demands

1. Replace the "tens of degrees" claim with the Tassis–Pavlidou scaling and a table of Δψ(353→150) versus α, ΔT, Δβ.
2. Provide an end-to-end forecast, d12- or Vacher-2025-style, with and without the 3D prior: Δr and σ_r for a BK/SO-like field and for LiteBIRD, and the fraction of decorrelation power removed.
3. Translate the 15% into (Δβ, ΔT) and show whether a 3D + ISRF + grain model can produce it. If not, say so.
4. Demonstrate 3D-map fidelity at N_H ≈ 1–3×10²⁰ cm⁻²: compare integrated Edenhofer E(B−V) against H I-based E(B−V) in the BK and SPIDER fields and report the prior contribution.
5. Reconcile the proposal with Planck 2018 XI, BK18 Δ_d, BK XVI, SPIDER and Guillet 2026, and state the empirical status of decorrelation plainly.
6. Rebuild H2 and H3 in the spin-moment formalism and state which spin-moments the layered prior fixes.
7. Specify how WNM dust in each layer gets its angle. Use the 0.22 p_CNM result as a prior.
8. Drop the "size sets β" framing. State the two grain scenarios explicitly and how they are marginalized.
9. Separate the LiteBIRD ℓ < 10 regime from the deep-field ℓ ≈ 80 regime and say which the pipeline addresses.
10. Show the Local Bubble baseline: a one-shell model per Alves 2018 and Pelgrims 2020 must be what the layered model is shown to beat.
11. Give the H I-to-distance cross-identification algorithm and its failure rate, and the PASIPHAE star density needed per pixel for an external check.
12. Include bandpass integration and color corrections in the forward model.
13. Fix the citations listed in §1.

6. Severity list

• blocker No quantitative link to r anywhere: no Δr, no σ_r, no forecast.
• blocker The forward model cannot produce the effect it claims to predict; per-layer T_i and β_i are supplied by none of the three ingredients, and plausible ΔT yields ~1% versus the 15% needed.
• blocker "Tens of degrees" two-cloud rotation contradicts Tassis & Pavlidou and direct arithmetic.
• major 3D-map completeness claim reversed at low column.
• major β_P − β_I called significant; 2018 value is 0.05 ± 0.03.
• major Grain size conflated with composition.
• major McBride et al. 2023 mis-cited; CMB Q,U are unbiased unless the angle also varies.
• major Scalar moments for polarization; spin-moments, Vacher 2025 and Guillet 2026 absent.
• major Missing Halal 2024, Local Bubble, Lei & Clark CNM/WNM, Mehandiratta R_P/p, Sponseller & Kogut, Ritacco.
• major Empirical status misrepresented; Pelgrims' own caveat and four power-spectrum non-detections unreconciled.
• major BK method misdescribed; H0 built on it.
• major "N_lay is 1 on most high-latitude pixels" versus 2.5–3 clouds per pixel.
• major LiteBIRD ℓ < 10 regime ignored.
• minor Regnier not Maffei; d12 attribution; "Hensley, McBride & Bull" pseudo-citation; Vansyngel and R_ℓ misattributed.
• minor Planck Int. L indicative numbers presented as measurements.
• minor a_align formula unnormalized and missing anisotropy factor.
• minor CMB-S4 status out of date.
• minor SPIDER 2025 selectively quoted.
• minor "Falsifiable versus not" rhetoric.
• minor "N ~ 2–8 is sufficient" unsupported; r ≈ 0.6 described as "a large fraction of the variance" when it is 36%.