Papers→ Paper 4
Paper 4
An Observed-Label Chirality-Dipole Null in 890,069 Quality-Controlled High-Confidence DESI Spirals and an 8.5-Million-Galaxy Catalog
Do spiral galaxies' apparent handedness directions cluster? An 8.5M-galaxy test (result: no dipole)
Classifies 8.47M galaxies. The strict safe-sample observed-label statistic remains null-consistent (z=0.63465, p=0.23768), and the v1.0.268 CE-composition audit retained the GZ1-core realization while documenting the honest-negative CE-inclusive retrain. v1.0.274 expands Expected Calibration Error (ECE) at first use without changing the science; exact confirmation and Houston review remain separate gates.
0B/0M/0m/0C open · August 3, 2026
The galaxy chirality catalog: 8.47M galaxies classified CW/CCW by a ViT-Small ensemble with flip-equivariant TTA. Two real-compute pod-campaign closures land in v1.0.268. (1) G1 manifest-retained ViT retrain COMPLETE on RunPod A4000 (<$1 total): trained on 8,637 objects (6,637 GZ1-core + 2,000 synthetic; ce_resnet_present=false — the Jia CE-ResNet catalog still needs external re-provisioning, so the 826-vs-846 sub-conflict stays open and released Catalog C labels are UNCHANGED), every object ID/split/seed retained in the committed manifest, best_val_acc=0.9931 at epoch 47, checkpoint backed up to 3 verified locations. (2) G2 training-disjoint validation: accuracy=0.9867 / Cohen's kappa=0.9733 on 3,000 GZ1 confident spirals disjoint on both the object-ID and label-source axes (overlap counts 0), presented with an explicit like-for-like distinction vs the historical kappa=0.40 human-vote figure — a genuinely different measure, not a replacement for it. Caveat sites narrowed honestly at abstract/intro/discussion/conclusions/data-availability. No science number changed; the primary remains null-consistent and harmonic results remain systematics diagnostics only. CE-ResNet re-provisioning, the MASTER-decoupled/full-likelihood covariance legs, complete metadata, a DOI-backed archive, exact v1.0.268 confirmation, and human review remain open.
- Draft completedone
- Historical automated reviewdone
Receipt-backed evidence retained; labels are version-specific
- Latest-version review coveragein progress
Version-specific evidence retained; final-hash bounded confirmation may still be pending
- Independent human reviewin progress
Required; automated-model labels are not journal decisions
- Archive and venue checksqueued
Immutable archive/DOI plus journal-specific scope and format checks
- Author submission decisionqueued
arXiv endorsement and journal submission remain author-controlled
- Null-consistent strict safe-sample observed-label dipole
- 8.47M-galaxy released classification catalog
- Manifest-retained training and disjoint validation evidence
- Bounded final-hash confirmation, systematics metadata, and Houston ApJS sign-off
- Largest spiral chirality catalog: 3.2M DESI Legacy spirals (471,049 high-confidence at p_CW^eq > 0.9)
- Headline ℓ=1 chirality-dipole observable: real-space dipole at +0.41σ (empirical-rank p=0.31; A_dip < 6.8\times10^-3 at 95% UL) — the earlier -0.12σ subsample-mask pseudo-C_1 null was withdrawn in v1.0.166 after a provenance audit
- Four-null systematic battery + direct cross-spectrum disfavors clean-dipole interpretation: ℓ=2 broadband, no quartile scaling, σ=-2.89 depth cross-spectrum at ℓ=2
- Falsification criterion: >5σ detection at ≳ 0.75\% amplitude in future ≥ 10^7-galaxy systematics-modeled survey
External peer review kit
Houston-driven manual round · paste prompt into any frontier LLM web UI with the PDF attachedOne click to copy a referee prompt scoped to this paper. One click to download the latest PDF. Paste both into Claude / GPT-5 / Gemini / Grok / Perplexity — return findings here and the autonomous cron will close them in the next bundled hard-fix wave.
preview prompt
You are an external referee for MNRAS / Physical Review D / JCAP (target journal depends on paper). Attached: Paper 4 v1.0.274 — "An Observed-Label Chirality-Dipole Null in 890,069 Quality-Controlled High-Confidence DESI Spirals and an 8.5-Million-Galaxy Catalog" Source: pipelines/p2_chirality/chirality_catalog_paper.tex PDF: PDF · 32 pp · v1.0.274 · updated Aug 3, 2026 · md5 6c7de2b81dfa3d7af2a7414214d57cfc · sha256 2641a228af1e3decf17d18341570c4e779483a823267421fe041aade1375e0d7 — Expected Calibration Error (ECE) expanded at first use; no scientific claim, number, or caveat changed. Read the FULL PDF end-to-end. Produce a referee report in MNRAS format with: 1. Recommendation: ACCEPT / MINOR REVISIONS / MAJOR REVISIONS / REJECT 2. BLOCKERS (must fix before publication) — list each with section/line + proposed fix 3. MAJORS (should fix) — same format 4. MINORS (polish) — same format 5. Strengths (>= 3 bullet points) 6. Specific scrutiny on: - Null-consistent strict safe-sample observed-label dipole - 8.47M-galaxy released classification catalog - Manifest-retained training and disjoint validation evidence - Bounded final-hash confirmation, systematics metadata, and Houston ApJS sign-off CALIBRATION (do not burn findings on these known classes): - The current date is June 2026. arXiv identifiers of the form 25xx.xxxxx and 26xx.xxxxx are VALID, already-published preprints — do not flag them as "future-dated" or "nonexistent". Verify a citation against arXiv/ADS before claiming it does not exist. - Correction notes, retraction notices, and "an earlier version stated X" disclosures in the text are DELIBERATE transparency policy. Flag them only if their content is wrong, never for existing. - Companion-paper citations marked "posted concurrently on arXiv" are deliberate placeholders; real arXiv IDs are inserted during the coordinated submission sequence. - Explicitly labeled conservatism allowances, scaling estimates, ansatz/heuristic status labels, and disclosed queued follow-up computations are deliberate scoping, not oversights — flag only if the label itself is inaccurate. - PDF text extraction can mangle math (square roots, fractions, superscripts). Before flagging "garbled" or "wrong" math, consider extraction artifacts; flag only what is visibly wrong in the rendered PDF. VERDICT STANDARD (apply the SAME high bar a first-pass Physical Review D / MNRAS referee would — this is one of the most rigorous journals in the world): - Assign each finding's severity (BLOCKER / MAJOR / MINOR) by your own independent referee judgment. Do NOT default to any tier, and do NOT soften a finding because the rest of the paper is strong. Do not echo this prompt's context. - A reporting choice that headlines the more favorable of two numbers, an unstated assumption, an uncontrolled systematic, or an internal inconsistency IS a real finding — classify it honestly (MINOR at minimum), not as mere "style" or "opinion". - Truth-audit any claim that seems off by checking it against the published .tex / on-disk artifacts before flagging (this only filters out genuine extraction artifacts — it does not lower the bar on real defects).
- 018.47M galaxies classified (1,592,107 CW / 1,609,053 CCW / 5,273,371 NOT_SPIRAL)
- 02Strict release-safe primary: N_selected=890,069, N_support=887,472, z_mom=+0.63465, one-sided empirical-rank p=0.23768
- 03G1 manifest-retained ViT retrain: 8,637 objects (6,637 GZ1-core + 2,000 synthetic; CE-ResNet component absent pending external re-provisioning), every object ID/split/seed retained, best_val_acc=0.9931 @ epoch 47, checkpoint backed up to 3 verified locations — proves the exact training realization is regenerable
- 04G2 training-disjoint validation: accuracy=0.9867 / Cohen's kappa=0.9733 on 3,000 GZ1 confident spirals disjoint from G1 training on both the object-ID and label-source axes (overlap counts 0); presented like-for-like against, and explicitly NOT as a replacement for, the historical kappa=0.40 GZ1 human-vote inter-rater figure — a different measure (model-vs-independent-labels agreement, not human-vote agreement)
- 05Coverage-calibrated OBSERVED-LABEL 95% sensitivity upper limit on the primary dipole: A_95^obs~=0.98% (linear-interp; logistic cross-check 0.955%), from 2,000 random-axis injections/amplitude through the exact primary estimator + exact fixed-occupancy null; NOT a physical parity-amplitude bound (that remains gated on the morphology transfer function)
- 06Exact 24,087-pixel FSC support: fixed-occupancy harmonic z=6.923, p=0.001996; apodized z=7.033; 10,000-draw binomial z=7.207, p=0.00059994; systematics diagnostics only
- 07Corrected public ApJS bundle: immutable HF data commit db110233 contains the safe catalog, 249,066-row quarantine, fixed-occupancy primary array f6360f4b, distinct pixel-permutation diagnostic 62bb1c01, manifest, schema, checksums, and reproducer; receipt e535b262 verifies remote paths and byte sizes
- 08Release-integrity closure: HF dataset revision 43fc8a5b publishes the complete clean-bootstrap semantic contract; HF model revision 6f113097 publishes the production-accurate model card
- 09Catalog C semantic validation streams all 8,474,531 primary and 249,066 quarantine rows and proves exact object-ID/HC-flag equivalence with zero failed gates; Catalog B remains historical and unreleased
- 10Galaxy Zoo 1 human-vote cross-check: 46,017 matched inputs, exactly 4,963 on the 394-pixel N_pixel>10 support and 41,054 excluded; null-consistent at z=−0.539, p=0.666 under its legacy pixel-permutation convention
- 11ℓ=1 MASTER decoupling numerically stable: full coupling-matrix condition number 3.17 (3.11/3.25 at 1°/3° apodization)
- 12Bin-by-bin CW flatness audited across 4 morphology axes; residuals are classification correlations, orthogonal to dipole tests
- 13Finite-grid injection pilots (20 axes, 100 injections/amplitude) report score-pass fractions only and do not themselves establish calibrated recovery thresholds; the separately-computed 2,000-axis A_95^obs upper limit above supersedes them for the observed-label sensitivity floor, but establishes no physical bound
- 14Edge-on TTA equivariance check: CW fraction 0.4975 ± 0.0006 on 785,859 edge-on galaxies, indistinguishable from catalog-wide
- 15Platt calibration: raw +0.79%/28.8σ → calibrated +0.4%/14.6σ → equivariant -0.26%/9.5σ
- 16Equivariance suppression factor 2.98× (raw +1.58% → equivariant −0.53% in asymmetry-A units; +0.79% → −0.26% in f_CW units)
- 17Harmonic injection scores are pipeline-response diagnostics on a different field, support, and null; they are not calibrated completeness or physical exclusions
- 18Hemisphere look-elsewhere null: p_LEE < 10⁻⁴ (0/10,000 MC nulls reach data)
- 19100,000-bootstrap CW/CCW asymmetry: A_obs=1.5757%, 95%CI=[1.471%, 1.685%], σ_stat = 28.80σ
- 20Two support-bound FS-C diagnostics show apodization robustness and broadband low-ℓ structure; six different-support or support-unproven calculations are excluded from the strict synthesis
- 21No physical or primordial amplitude bound is claimed; matched-footprint independent-estimator analysis and a spatial transfer model remain open
- 22R25conf round 2-of-2 CLEAN (93 findings audited); one substantive catch — App A field-convention description corrected from artifacts, no number changed (v1.0.170)
- 23Pod recompute wave: unthresholded-sample injection floors A50=0.36% / A95=0.63% (the 0.57% excess sits between — honest disclosure); area-uniform axis curve P3σ=0.59 @ 0.75%; threshold sweep reproduces canonical exactly; T7 quantified by confidence (0.267 vs 0.383); training-acc semantics corrected — 93.7% = 3-class val, 94.9% = CW per-class (v1.0.171)