Independent fact check and research extension
Audit date: 6 September 2026. Original repository baseline: e23a3d2b84abf1aa4b2e1f09d8404ebfc03e6d7a. This is an additive review by Codex, with parallel research checks and an independently reviewed numerical implementation. All original research under 01-independent/, 02-drive/, 03-critic-optimist/, and 04-summary/ remains unchanged. Model names are the repository's own attribution labels; this audit does not authenticate which service executed the original runs.
Main finding
The research has scientific value as a synthesis and source of testable hypotheses, but it does not demonstrate a new astrophysical discovery or a validated improvement in cosmological inference. The common physical picture is largely sound: optically thin polarized emission adds as Stokes vectors; spectral differences between emitting structures can change their relative contributions; geometry and grain physics therefore matter. The promising contribution would be a controlled test of whether uncertain, correctly associated 3D information improves predictions beyond an equally flexible empirical foreground model. See the novelty assessment and research protocol.
The strongest recurring problems are treating restricted results as universal theorems, confusing a sightline's polarization residual with a sky's BB residual, and converting toy residual fractions into experimental r forecasts without the likelihood and instrument. The primary references support narrower claims than some essays and summaries make. Agreement between the LLMs does not independently validate those claims.
Coverage and method
The review read all 27 substantive research outputs: three independent essays, six driven essays, twelve critic/optimist reports and six syntheses. It also examined all six driver briefs and three summary documents. Prompts and manifests establish provenance; raw execution logs and failed-run artifacts were not treated as additional scientific essays. All three participating models and all role assignments receive individual appraisals in the two detailed audits.
This is a substantive-claim audit, not a guarantee that every sentence, number or citation has been verified. Sources were checked using primary journal papers, arXiv records/full text and official project material where relevant. Recent 2026 references were checked instead of rejected on date alone. The literature search is bounded and cannot prove priority or absence of prior art. The two JSON ledgers retain claim-level verdicts, model/mode attribution, original paths and supporting sources; repeated claims in different stages can appear separately. Counts are not accuracy scores.
| Model, using repository attribution | Scientific strengths in this corpus | Main limits |
|---|---|---|
| GPT-5.6 Sol | Careful extinction/emission distinction, grain-emissivity integrals, hierarchical inference and conditional claims | A real author-attribution error, some insufficiently qualified factorization/bias claims, and acceptance of selected overgeneralizations |
| Grok 4.6 | Broad systematic inventory and layered model structure; useful hypotheses | Ordinary ILC incorrectly described as dust nulling, overgeneralized mixing/rotation statements, unit and quantitative assumptions needing correction |
| Fable 5.1 | Quantitative examples, tomography methods and explicit inference ideas | Reversed temperature derivative in one driven essay, reversed grain-growth interpretation in another, fragile optical-to-microwave transfer, overreach from local residuals to BB/r |
These describe the particular outputs, not reliable rankings of model capability. A single prompt, shared sources, unblinded assessment and interdependent later stages do not support a controlled claim that one collaboration mode is generally better.
Corrections that matter scientifically
- Two-cloud rotation: distinct SEDs and distinct physical angles are not sufficient for smooth rotation. The exact numerator contains
sin(2Δψ); at 90° the Stokes vectors are antiparallel. Multilayer terms can also cancel. The derivations state the exact conditions. - Known-angle inversion: two dust-only amplitudes can be recovered at one frequency only under specified assumptions. The matrix is singular at both 0° and 90°, and its exact condition number is not
1/|sin(2Δψ)|. - Temperature derivative: the MBB slope is
β + 3 − x/(1−exp(−x)). A sign error reverses the predicted temperature-driven rotation. The shared simulation kernel is checked against a finite difference. - Spatial spectra: unchanged pixel angles do not imply unchanged BB morphology. The synthetic map experiment supplies a concrete counterexample using spatially varying real frequency scalings.
- ILC: ordinary ILC preserves a CMB response and minimizes variance; explicit dust nulling requires additional constraints. See the primary-method sources in the detailed audits.
- Criticism can create errors: Pelgrims et al.'s full paper does specify a ≥60° selection threshold. Its absence from the abstract was not grounds for discarding it. Some criticism also promoted a temperature linearization into an unsupported negative physical result.
- Recent references: Guillet, Mehandiratta, McCallum and GRADE-POL papers cited with 2026 dates are real. Their existence validates bibliographic identity, not every inference the essays make from them. Details and corrected bibliographies are in the literature audit.
Added artifacts
| File | Purpose |
|---|---|
| Independent and driven audit | Individual appraisal of all nine essays and six briefs; factuality and scientific value |
| Critic–optimist audit | All 18 dialogue-stage outputs; successful corrections and regressions |
| Literature audit | Primary sources, source identity corrections and closest prior art |
| Novelty and research program | Distinguish synthesis from discovery; falsifiable validation protocol |
| Exact derivations | Rotation, coherence bound, inversion conditioning and spectral nullspace test |
| Simulation methods | Reproducibility, units, synthetic sky definition, tests and limitations |
The website adds Fact check, Simulation lab and Novelty routes. Original prose is preserved and marked as original model work. The simulations generate educational synthetic outputs; no observational maps, experiment-specific forecasts or new cosmological measurements were produced.