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

Overlap and unique thoughts

Experiment: three models answered the same CMB-dust / B-mode prompt independently, then in a full drive matrix, then with the other two as critic and optimist.

All 6 critics and all 6 optimists exist. Folder names under cmb-dust-bmode-model-matrix/ encode mode, models, and effort.

Overlap (all three independent models)

They independently built the same physical skeleton:

1. Polarized thermal dust is the dominant high-frequency B-mode foreground. Aspherical grains aligned with B emit with E perpendicular to B_POS. A single layer is a modified blackbody in I and a scaled, angle-tagged copy in Q,U.

2. “Spectral-index decorrelation” is not a scalar β_d becoming uncorrelated with itself. It is frequency decorrelation of the polarized maps: Q,U at ν₁ are not a real scalar times Q,U at ν₂.

3. The LOS mechanism is the two-condition Tassis & Pavlidou (2015) picture: different layer SEDs and misaligned B_POS. SED differences without misalignment give curvature but no angle rotation; misaligned fields with identical SEDs still factorize.

4. Observational nuance is shared: Planck Int. L suggested 217×353 BB decorrelation; Sheehy & Slosar and Planck XI did not confirm it as a robust all-sky detection; Pelgrims et al. (2021) detected the LOS angle effect on HI-selected multi-cloud, misaligned sightlines (~10–15% cloud-to-cloud polarized 217/353 variation).

5. A one-MBB-per-pixel model is misspecified. Averaging MBBs is not an MBB (Chluba / moments). Extra free β,T moments buy unbiased r only at a cost in σ(r) (Vacher et al. 2022).

6. The three ingredients map onto three factors in a layered forward model:

   • 3D dust → layer weights / which sightlines are multi-cloud
   • grain size/composition/alignment → layer SEDs and p(ν)
   • sky-projected B (starlight tomography, HI fibres, 353 GHz) → layer phases e^{2iψ}

7. 3D “dust maps” in hand are mostly extinction/reddening, not millimetre emission tomography.

8. Pipeline idea: layered (or complex-moment) forward model with priors from ancillary data, not a frozen template; keep residual moments for unresolved averaging.

9. Shared limitations: sparse high-latitude stars, extinction ≠ emission, HI velocity ≠ distance, γ poorly measured, grain models degenerate, finite frequency channels, instrumental systematics can mimic decorrelation.

10. Shared names: Planck XI/XII/Int. L/XXII, Tassis & Pavlidou, Pelgrims 2021, Clark/HI fibres, Gaia 3D dust, RAT alignment, Draine & Fraisse, Guillet et al., Hensley & Draine astrodust, Vacher/Mangilli/Chluba moments, LiteBIRD / BK18 r targets.

After critic/optimist, all three syntheses converge further on: SED×angle covariance as the necessary condition; 90° emission vs B_POS; R_{P/p} not a universal per-layer Stokes transfer; millimetre β as dielectric, not size; McBride vs Sponseller as two different papers; spin-moments as the likelihood interface; the 3D programme as a prior and diagnostic, not a turn-key r pipeline.

Pairwise overlap (independent)

Grok ∩ Fable

• PASIPHAE as a real programme (Fable names “dedusting” as PASIPHAE’s term; Grok uses it as the missing all-sky starlight tomography).
• Spin-moments / polarized moment expansions and Ritacco 2023 residuals.
• R_{P/p} as the optical-to-submm polarized bridge (Fable quantitative; Grok more qualitative until synthesis).
• BISP-1 / tomography methods (Fable names them; Grok cites Panopoulou).

Grok ∩ GPT

• Poh & Dodelson 2017 with Tassis & Pavlidou for the two-condition mechanism.
• Explicit warning that maps decorrelate, not a scalar β_d.

Fable ∩ GPT

• Complex moments whose phase encodes angle rotation (GPT ℳ_β; Fable spin-moments 𝒲).
• McBride-class result that CMB Q,U can stay unbiased unless both SED and angle vary (GPT cited the wrong paper name independently; Fable/McBride and later syntheses fixed it).

Unique to Grok 4.6 (independent, and what survived)

Thoughts the others did not really develop until later:

1. Lee & Draine cos²γ in the Stokes equations, plus later admission that γ is the weak link.
2. MRN / Weingartner–Draine size ladder, a_align ~ 0.05 μm, intra-cloud T moments from n(a).
3. Magnetic dipole / iron inclusions (Draine & Hensley 2013; Hensley & Bull 2018) as a polarized SED contaminant — still largely Grok’s unique systematic after synthesis.
4. AME / spinning nanoparticles as extra unpolarized intensity at ≲100 GHz.
5. Explicit H0–H3 hypothesis ladder (masking → layered MBB → intra-layer moments → harmonic hybrid replacing a single ℛ_ℓ).
6. Moment-constrained ILC (Remazeilles, Rotti & Chluba).
7. LiteBIRD ~31′ beam vs unresolved MHD turbulence: 3D layers reduce moments, they do not zero them.
8. After Fable’s critic: Grok retracted “any multi-layer beam ⇒ p, ψ, β all frequency-dependent” and “tens of degrees” of two-cloud rotation; adopted Tassis & Pavlidou’s ≲5° unless α ~ 85°.
9. After synthesis: note that CMB-S4 was cancelled 9 July 2025; keep the scientific residual target anyway. (The other independents still wrote CMB-S4 as a live Stage-4 experiment.)

Unique to GPT-5.6 Sol xhigh

1. Sharpest operational definition: single-MBB-per-pixel = one real mixing scalar, one mixing-matrix column.
2. Explicit j_ν^I vs j_ν^P integrals over dn/da, C_abs, C_pol, alignment R_m(a).
3. Fanciullo et al. 2017: size changes alone cannot explain emission-to-extinction polarization. Still GPT-unique after synthesis.
4. Rayleigh-regime millimetre opacity ≠ grain radius (Grok/Fable adopted this after being criticised).
5. Hierarchical Bayesian posture: microwave data update layer amplitudes; ancillary data are priors.
6. Shortest independent essay; most careful about non-detection vs targeted detection.
7. Independent citation error (Sponseller & Kogut linked to McBride’s MNRAS paper) — Fable’s critic caught it; both GPT syntheses correct the attribution.

Unique to Fable 5.1 max

1. Opens by defining dedusting as PASIPHAE’s term and a compact r / σ(r) table.
2. Residual arithmetic: 30% layer, Δβ=0.3 → R(217×353)≈0.999, invisible to Planck, ~0.7% of 150 GHz dust power → δr ~ 10⁻³. Angle rotation 1°/2°/3° → 0.2%/1%/2% leftover BB. After GPT’s critic, Fable softened the two-cloud Δψ numbers but kept the budget arithmetic, checked against a direct residual (~0.8%).
3. BK18 Δ_d parameter and ln²(ν₁/ν₂) scaling.
4. Ritacco 2023: first-order MBB moments fit EE residuals but not BB.
5. Guillet et al. 2026 SRoll2 vs PR4 residual covariances (systematics vs astrophysics). Other models picked this up in synthesis.
6. Starlight-tomography method zoo: BISP-1, RoboPol 4 deg², TRIShUL, PASIPHAE/WALOP.
7. Independent overclaim: starlight q,u × R_{P/p} gives layer 353 GHz Stokes with cos²γ cancelling. GPT’s critic attacked this; Fable’s synthesis keeps the cancellation but adds the minus sign, 2.5× model shortfall, and non-universality.
8. DustFilaments / Huffenberger–Rotti–Collins filament–B alignment as origin of E/B≈2.
9. GRADE-POL / THEMIS named as living grain-model controversy.
10. Staged deployment: now risk maps; later PASIPHAE priors; then joint LiteBIRD fits.

Thoughts that only appeared after a model was driven

Drive mostly transferred emphasis, not new physics.

No drive cell invented a mechanism the independents had not already named. They changed which survey is the lever and how aggressive the r claim is.

Thoughts that only appeared after critic / optimist

This is where the models actually left the others’ original frame.

Blockers the critics forced (and syntheses accepted)

• Grok’s opening theorem was false. Multi-layer + different {β,T,ψ} does not automatically make p, ψ, β frequency-dependent. Rotation needs SED–orientation correlation and comparable polarized amplitudes. GPT stated this independently; GPT-as-critic beat it into Grok; Grok’s syntheses open with 𝒫_ν = Σ A_c s_c(ν) e^{2iψ_c} and the two necessary conditions.
• Faraday RM does not give per-layer γ. Fable independent implied it; GPT critic killed it; Fable synthesis dropped it.
• R_{P/p} is not a per-cloud Stokes photocopier. 90° sign, selected sightlines, 2.5× model shortfall (Planck Int. XXI).
• Mask pixels are not LOS components of 𝒫_ν(n̂). Fable had folded power-spectrum averaging into the per-pixel sum; GPT critic split LOS vs beam vs mask.
• Citation hygiene: GPT fused Sponseller & Kogut with McBride; Grok overstated two-cloud Δψ relative to Tassis & Pavlidou; Fable attributed a 60° Pelgrims threshold that the abstract does not state.

Optimist additions that stuck

• Complex / spin-moments as the interface to the r likelihood, not a rival to 3D structure.
• Local Bubble as a candidate first high-latitude layer, with Halal, Clark & Tahani (2024) as the caution that p_353 does not show a simple tangential shell.
• Zelko et al. as a prototype 3D emission/temperature map, not just extinction.
• Staged use now: risk maps and Δ_d(n̂) priors from HI + extinction, without waiting for all-sky PASIPHAE.
• Chromatic dust EB as a possible payoff / contaminant for birefringence (Fable synthesis).

What primaries rejected from reviewers

• Grok rejected fabricating an end-to-end Δr number (Fable critic demanded one).
• Grok rejected unverified 2026 citations the optimist offered.
• Fable rejected the critic’s implication that the method is blocked until identifiability is proven; kept “unforecast, not unjustified.”
• Fable rejected putting the Local Bubble cartoon in front of everything.
• GPT rejected turning R_{P/p} into a hard constraint; kept it as an environmental prior.

Outside what the others considered (strongest divergences)

Ranked by “the other two independents simply did not go there”:

The last Fable synthesis (Grok critic + GPT optimist) added two leftovers that still sit mostly on Fable’s side:

• Mehandiratta et al. 2026: measured cloud-to-cloud variation in R_{P/p}, so the optical-to-submm transfer is not even a single number per environment.
• Identifiability with supplied angles: Fable rejected the critic’s claim that layers cannot be separated at one frequency. With distinct known ψ, Q and U separate two layers with condition number ~1/sin(2Δψ), independent of Δβ. That is Fable’s defence of why the third ingredient is not optional.

Bottom line: they agreed on the mechanism from the first independent essays (LOS mixing + misaligned B + grain-weighted SED → non-factorizable Q,U). They disagreed on (i) how immediately a multi-layer sightline breaks a single MBB, (ii) which millimetre grain physics matters, (iii) how quantitative the r residual is today, and (iv) which survey is the actual lever. Critic/optimist collapsed (i) onto GPT’s two-condition lemma, moved Fable’s residual arithmetic and spin-moments into Grok/GPT, and left Grok’s iron-dipole/AME, GPT’s Fanciullo point, and Fable’s angle-conditioned layer invertibility as the leftovers the others never really absorbed.