Independent audit of the independent and driven research

Audit date: 2026-09-06. This is a separate contribution; none of the original essays or briefs was edited. It assesses scientific content, not the general capabilities of a model. Model names are the repository's labels.

Finding

The corpus is a useful, unusually connected literature synthesis and a starting point for a research pipeline. It does not demonstrate a new astrophysical effect, a validated foreground-removal method, or an achieved reduction in tensor-to-scalar-ratio bias. Its central mechanism and the proposal to use distance-resolved polarimetry were already described in the cited literature. The most promising extension is a quantitatively tested way of selecting and regularizing dust spectral degrees of freedom using uncertain external information.

Several equations and bibliographic statements need an accompanying errata layer before implementation. The most consequential are a reversed temperature derivative, an incorrect description of ordinary ILC, mismatched amplitude units, and repeated transfers from a local polarization-vector example to a harmonic B-mode or experimental r forecast without the intervening calculation.

Coverage and method

All nine complete answer files and all six complete driver briefs were read. Source checks used primary paper records, abstracts and, where needed, full papers. Algebraic identities and several numerical examples were independently recomputed from the Planck function. This is a substantive claim audit, not a certification of every sentence or every bibliography entry. Forecasts and instrument specifications are treated as stated design assumptions, not verified current schedules. A statement marked unsupported lacks sufficient support for the asserted inference; it is not necessarily false.

The following six complete briefs were included in the provenance audit:

The six driver-brief.md files adjacent to the driven answers were also checked. Their status is provenance, not evidence. The two Fable briefs explicitly supply quantitative scenarios that reappear downstream; matching results are not independent discoveries. Grok’s briefs contain problematic grain-growth and dust-power assertions. GPT driven by Grok explicitly repairs the grain-growth reasoning, whereas Fable driven by Grok repeats a lower-β grain-growth interpretation. GPT’s briefs encourage appropriate observational caution, but that does not prevent the downstream temperature-sign error. Drivers are therefore distinguished from answering models in the companion JSON.

What survives checking

The complex line-of-sight sum, Pν = Σ Aᵢfᵢ(ν)exp(2iψᵢ), is the right organizing model in the optically thin approximation, with a consistent polarization convention. Unequal cloud SEDs plus non-collinear Stokes vectors can rotate the resultant with frequency. Identical SEDs factor out even for a tangled geometry. Per-pixel changes in SED can also lower map correlations without rotating any pixel. These are established results, not new mechanisms. Tassis & Pavlidou: two-cloud foreground decorrelation (2015), Vacher et al.: spin-moment expansion (2023)

The essays’ apparently conflicting observational claims mostly concern different selections and statistics. Planck’s broad-area null result is compatible with Pelgrims’ H I-selected angle-difference excess and with SPIDER’s regional SED differences. None establishes a universal sky-wide decorrelation amplitude. Planck 2018 XI: polarized dust foregrounds, Pelgrims et al.: selected-sightline evidence (2021), SPIDER: large-scale polarized dust analysis (2025)

Extinction tomography, H I morphology, and starlight polarization provide complementary information, but none directly supplies the complete polarized microwave emissivity. H I velocity is not a unique local distance coordinate. The published 3D temperature reconstruction is model-conditioned and its two-cloud Cepheus temperature separation is only about 1σ, so descriptions of those temperatures as decisively resolved overstate the evidence. Zelko, Finkbeiner, Lee & Green: 3D dust temperatures (2022), Clark & Hensley: H I-based Stokes templates (2019)

The 2026 references are real, not inherently suspicious because of their dates. Guillet’s SRoll2 residual covariances are compatible with temperature-only variation; the PR4 conclusion differs. MilkyWayAtlas infers a local density/velocity reconstruction by combining H I and extinction information. GRADE-POL II models optical polarization with grain alignment physics. These are useful inputs with distinct validation domains, not demonstrations of a primordial-B cleaning pipeline. Guillet et al.: complex polarized dust residuals (2026), McCallum et al.: MilkyWayAtlas (2026), Tram et al.: GRADE-POL II (2026)

Individual findings

I-GPT: independent GPT-5.6-sol

I-GPT infers frequency-independent polarization fraction from a common Q/U scaling. A positive scalar scaling fixes ψ; it fixes p = |P|/I only when intensity shares that scaling. This also affects similar wording in Grok→Fable.

The temperature-moment expression in I-GPT uses ∂T ln Bν after defining reference-normalized SEDs. For amplitudes defined at ν₀, the temperature basis is ∂T ln[Bν/Bν₀]. An unnormalized convention can absorb the missing constant into the zeroth moment, so this is a convention ambiguity rather than evidence that all moment reasoning is wrong. Alignment also appears in both the polarized grain integral and a subsequent amplitude factor; an implementation must specify which carries it to avoid double counting.

The final reference in I-GPT labels MNRAS 519, 4370 as Sponseller & Kogut. It is McBride, Bull & Hensley. The linked work’s broad message is relevant; the authorship is wrong. McBride, Bull & Hensley: dust and CMB recovery (2023)

Scientific value: the hierarchical forward model, auxiliary-data priors and explicit identifiability caveats are productive research design. No data fit or validated gain is supplied.

I-Grok: independent Grok-4.6

I-Grok says ordinary ILC nulls one dust SED. Ordinary ILC minimizes variance subject to unit CMB response; a dust-null constraint is an extra condition in constrained ILC. Non-MBB dust does not automatically bias every ILC implementation as asserted in the introduction. Remazeilles, Delabrouille & Cardoso: constrained ILC (2011), Remazeilles, Rotti & Chluba: constrained moment ILC (2021)

I-Grok defines Aᵢᴾ = p₀ᵢ cos²γᵢ τᵢ, which is dimensionless, then multiplies it by a dimensionless normalized SED to obtain Q/U. Either define Aᵢᴾ as the reference-frequency polarized intensity, including Bν₀(Tᵢ), or use an unnormalized Planck function. As written the equations cannot produce intensity units.

I-Grok makes changing angle/fraction/index inevitable whenever multiple physical properties differ. That is false: common SEDs factorize. The statement in I-Grok that a phenomenological MBB is not falsifiable is also false; it makes spectral predictions testable with additional bands. The paper’s own proposed single-cloud control is only a statistical control: one catalogued cloud may contain unresolved gradients or missing gas phases.

I-Grok overstates the consequences of a single grain composition. Size-selective alignment and differing temperatures can weight I and P differently even in a composite-grain model. A population model may make a nearly flat FIR polarization spectrum plausible without making intensity-derived T and β exact polarization parameters. Hensley & Draine: astrodust+PAH model (2023)

Scientific value: the map/tracer inventory and failure-mode discussion are useful. The promise of uniquely predicting the sky from a few layers is too strong without a posterior and held-out tests.

I-Fable: independent Fable-5.1

The local two-cloud examples and the uncorrelated-layer covariance calculation in I-Fable reproduce numerically under their assumptions; see the numerical checks below. But a Gaussian distribution of a spatial index does not by itself imply the displayed correlation law for every harmonic Rℓ: the angular covariance and correlations with amplitude must also be specified.

I-Fable says BK18 marginalizes its decorrelation parameter. The baseline fixes no decorrelation; varying the parameter is an extension in Appendix E.2. The BK XVI report of Δd > 0.98 at 68% is a real published bound and should not be dismissed. BICEP/Keck XIII: BK18 analysis and Appendix E.2, BICEP/Keck XVI: H I filament correlation

The emission/extinction conversion near I-Fable needs a 90° angle rotation (a minus sign in complex polarization when conventions otherwise match), and the empirical Rᴾ/ₚ transfer has environmental and column-matching assumptions. It does not remove all alignment and inclination uncertainty exactly per cloud. Planck Int. XXI: emission/extinction polarization ratios

The identifiability argument in I-Fable mixes two parameterizations. With independent Q and U amplitudes per layer, the design matrix is I₂ ⊗ F: actual cloud angles do not improve its singular values. If external angles are treated as known, each layer can instead have one real amplitude and a column fᵢ ⊗ (cos2ψᵢ, sin2ψᵢ); angles may then improve conditioning. Counting 30 data values is not a rank or noise analysis. The angle bound in I-Fable is a useful toy diagnostic, but its claimed residual below 10⁻⁴ in r is not established by that bound alone.

Scientific value: strong literature integration and reusable worked examples; the inference design should be upgraded with explicit parameterization, prior uncertainty and map-level validation.

Fable→GPT

The Fable→GPT essay handles the two-cloud vector geometry and limiting cases carefully. Its β₁ = 1.45, β₂ = 1.65, T = 20 K, amplitude ratio 0.5 and 45° separation reproduce a 353→150 GHz rotation near 1.86°. The numerical agreement is a successful calculation, not independent evidence for its occurrence in a real target field: the brief supplied the example.

The Gaussian decorrelation calculation is valid for the explicitly idealized local ensemble with amplitude-independent Gaussian index fluctuations. Use its result as that statistic unless a spatial covariance model has also been supplied. External optical templates can remove direct shared microwave instrumental-noise contamination, but do not automatically guarantee an unbiased microwave estimator: tracer mismatch, selection and chance sky correlations still matter.

Scientific value: one of the most reusable tutorials and proposed workflows in the corpus; its value is explanatory and methodological, not a measured foreground-removal achievement.

Fable→Grok

The Fable→Grok precision statement asks for a percent-of-a-percent level in dust power while its own numbers imply a tensor-to-dust ratio around 7×10⁻⁵ / 0.009 ≈ 0.78%. A desired fractional bias budget must be stated before a dust-residual requirement can be derived; 0.01% is not the consequence of those numbers.

The local 1 − cos(2δψ) loss and the later power-spectrum correlation are different observables. The formula 1 − R² applies to an optimally fitted scalar template and does not say that knowing R removes that residual. A 15% 353/217 ratio change at fixed temperature corresponds to Δβ = ln(1.15)/ln(353/217) ≈ 0.287, not automatically a spatial σβ of 0.1–0.2.

The bibliography assigns Ritacco et al. to A&A 672, A22. The primary record gives A&A 670, A163. Ritacco et al.: Planck residual dust SED (2023)

Scientific value: useful scenario explorer and research checklist. Its precision target and local-to-harmonic translation need explicit assumptions before numerical results are used in a forecast.

GPT→Fable

The temperature term in GPT→Fable has the wrong sign. Since d ln Bν(T)/d lnν = 3 − g(x), where g(x) = x/(1 − e⁻ˣ), the correct ratio slope is

d ln ρ / d ln ν = (β₂ − β₁) + g(x₁) − g(x₂),    ρ = P₂ f₂ / (P₁ f₁).

Its displayed expression reverses temperature-driven rotation. Equal-temperature examples cannot detect this bug. The preceding angle derivative is otherwise correct. The statement in GPT→Fable that a 90° separation gives no rotation needs an exception: when ρ crosses one, polarization passes through zero and its angle jumps by 90°. The given 0.8→0.67 example does not cross and remains valid.

The claim that distance-resolved models are the only kind able to predict decorrelation is false; statistical and moment models also predict frequency covariance, although they do not uniquely assign it to individual clouds. Six or more bands and two layers are not a guarantee of identifiability. Vacher 2022’s cited numerical bias/uncertainty results are simulation-specific fits of BB cross-angular spectra, not a general Q/U pixel reconstruction result. Vacher et al.: dust moments and LiteBIRD simulations (2022)

Scientific value: useful project architecture, but the derivative must be independently implemented from the corrected identity and unequal-temperature limits tested.

GPT→Grok

The GPT→Grok essay makes a valuable distinction between Planck’s broad-area null and selected-sightline evidence. It also handles the near-cancellation 90° flip correctly.

GPT→Grok says misalignment makes the total polarization fraction smaller than each individual cloud’s fraction. This is not generally true if the intrinsic fractions differ. The bound is p_total ≤ Σ Iᵢpᵢ/Σ Iᵢ ≤ max pᵢ; cancellation reduces the first bound but need not put the result below the least-polarized layer.

GPT→Grok calls an incorrect scalar astrophysical SED a calibration error and infers scalar-multiple maps from fixed angles. A per-pixel scalar law is not a single global proportionality if cloud mixtures vary across the sky. The stated 15–25 K temperature range gives an 11.38%, not 15%, change in the 353/217 Planck-function ratio at fixed β. That is a modest numerical overstatement, not a refutation of temperature-driven decorrelation.

Scientific value: a substantially more disciplined synthesis than independent Grok, with useful likelihood and auxiliary-data uncertainty discussion. Requiring every analysis to show a visible increase in σ(r) is too strong: sufficiently informative or irrelevant nuisance parameters can leave it practically unchanged.

Grok→Fable

Grok→Fable cites grain accretion/coagulation as lowering β by several tenths. In the cited Köhler et al. model, accretion raises β from 1.5 to 1.8 and coagulation raises it to about 2. This reverses the cited physical trend. Real grains need not follow one universal β trend, but that paper does not establish the one asserted here. Köhler, Ysard & Jones: grain evolution (2015)

The thermal/index sensitivity derivatives are approximately correct, but derivatives alone do not prove which process dominates: variances, covariances, amplitudes, angular selection and systematics matter. Guillet’s temperature-only compatibility for SRoll2 is a reason to test both, not impose index dominance as an observational conclusion. Guillet et al.: complex polarized dust residuals (2026)

The end-stage claim at Grok→Fable that high-latitude sightlines have at most three layers and all dust within about 500 pc is not a valid universal cutoff. Published tomography detects multiple structures extending to about 2 kpc, and the extended dust distribution matters in complexity studies. Pelgrims et al.: degree-scale magnetic tomography (2024), Halal, Clark & Tahani: dust complexity and polarization (2024)

The limitations section’s weighting τᵢTᵢ^(4+β) is a bolometric approximation, not monochromatic microwave emission. At a specified frequency the weight is τᵢ(ν/ν₀)^β Bν(Tᵢ). Finally, Grok→Fable explicitly “corrects” Zelko et al. to a two-author paper; the record has four authors: Zelko, Finkbeiner, Lee and Green. Zelko, Finkbeiner, Lee & Green: 3D dust temperatures (2022)

Scientific value: a detailed synthesis with a correct two-cloud ratio derivative. The grain-evolution prior, assumed map completeness and confidence in predicted r reductions require revision in any new implementation.

Grok→GPT

The Grok→GPT essay repairs the brief’s main grain-physics misconception: ordinary submicron interstellar grains remain in the Rayleigh limit at CMB wavelengths, and grain growth does not universally flatten β. The complex-derivative explanation is consistent with the correct two-cloud result. The proposed constrained-ILC conditions explicitly distinguish unit CMB response from dust-subspace deprojection.

Scientific value: careful treatment of probabilistic layers, model selection, nuisance parameters, bandpasses and held-out validation. The forward model remains a proposal. Assigning plausible grain and field priors does not yet demonstrate that a particular telescope can recover r with a smaller error or bias.

Independently recomputed numerical checks

These are exact toy-model calculations using h = 6.62607015×10⁻³⁴ J s, k = 1.380649×10⁻²³ J/K, c = 299792458 m/s and T_CMB = 2.7255 K; they are not measurements. Frequencies are monochromatic. A real instrument requires bandpass integration.

For a residual in a specified band or harmonic bin, with target x and template y,

C_res(a) = C_xx - 2a C_xy + a² C_yy
         = C_xx(1 - R²) + C_yy(a - C_xy/C_yy)².

Thus 1 − R² is the minimum variance fraction under scalar regression, not a residual guarantee for a fixed or uncertain calibration. A pointwise vector projection uses a different statistic from a harmonic BB cross-spectrum.

For a spatially uniform small angle error α and negligible EB, the leading residual BB power is 4α² CEE. If CEE/CBB = 2, this is 8α² of the original BB, or 0.244%, 0.975% and 2.19% at 1°, 2° and 3°. For varying α(n), E/B mixing is nonlocal and requires a map transform. An r bias then additionally needs the tensor template, multipole weights, mask, instrument response, foreground fit and likelihood.

Scientific value and novelty boundary

The original corpus has no demonstrated new observation or novel physical law. Combining grain physics, 3D dust maps and field geometry is a useful synthesis with substantial prior art. Analytical examples illustrating cancellation and extrapolation risk are deductions from the cited model. Independently reproducing them raises confidence in the calculation, not in novelty.

A credible next contribution is a test of whether uncertain geometric priors improve identifiability and out-of-sample CMB recovery. Compare free Q/U layer amplitudes with uncertain angle-constrained amplitudes, using the same noise, frequencies and sky realizations. Report noise-weighted singular values, posterior calibration, held-out frequency prediction, and bias/coverage for r after full E/B analysis. Include missing layers, wrong extinction-to-emission conversion and correlations between grain properties and geometry. A complexity mask must also be tested for changed sky selection and E/B leakage.

The result could be scientifically publishable if it establishes a reproducible advantage or failure boundary beyond existing moment and layered baselines. The present essays provide the motivation and candidate ingredients, not that result. The companion claim ledger separates confirmed statements, conditional deductions, errors and proposals without assigning arbitrary model scores.