Galaxy Chirality Classifier (bamfai/galaxy-chirality-v2)

Companion model for the current Paper IV manuscript — An Observed-Label Chirality-Dipole Null in 949,584 High-Confidence DESI Spirals and an 8.5-Million-Galaxy Catalog.

Current manuscript source: Hubify-Projects/bigbounce. No immutable repository tag or DOI for the current Paper IV manuscript is claimed.

Model overview

• Architecture: ViT-Small/16 with mirror-equivariant test-time augmentation. Z₂ 2-fold flip is the production TTA mode (D₄ 8-fold tested on hold-outs only; see Paper IV §sec:tta).
• Output classes: CW, CCW, NS (not-spiral).
• Production Catalog C uses raw softmax ranking scores with flip-swap averaging; these scores are explicitly uncalibrated. Per-class Platt scaling and Catalog B are retained only as a historical diagnostic and are not part of the released production Catalog C transform.
• Equivariance suppression factor 2.98× (raw asymmetry +1.576% → equivariant asymmetry −0.529%).

Catalog scale

Applied to the DESI Legacy DR8 8.47M-galaxy footprint via the canonical Paper IV pipeline:

• 8,474,531 galaxies in the science-facing catalog
• 1,592,107 CW + 1,609,053 CCW = 3,201,160 chirality-relevant spirals
• Catalog-wide CW fraction (post-TTA equivariant) 0.4974 ± 0.000279, a −9.47σ classifier-residual monopole relative to 0.5 in raw-count binomial units, not a cosmological dipole. The distinct primary fixed-occupancy dipole statistic is consistent with its null (z=+0.7053169638; one-sided empirical-rank p=0.2246775322).

Current Paper IV result

The primary high-confidence observed-label statistic is consistent with zero under fixed-occupancy label randomization (z=0.7053169638, one-sided empirical-rank p=0.2246775322). It is not a calibrated true-spin, physical-amplitude, or primordial-parity bound. WLS and harmonic results use different supports or nulls and are retained as systematics diagnostics rather than headline detection tests.

Historical v1.0.123–v1.0.125 changelog

The following entries document an older manuscript state. They are preserved for provenance and do not describe the current manuscript or establish a current clean-review verdict.

• v1.0.123 closed the Houston-shared external review on v1.0.122 (ChatGPT MAJOR REVISIONS + Grok MINOR REVISIONS + Gemini MAJOR REVISIONS; 9 BLOCKER + 10 MAJOR bundled hard-fix; 2 BLOCKERs audit-falsified). Stripped all internal-review scaffolding (Perplexity R22, P4-EXT, P4-INT, R20-Grok-B2 closure, etc.); fixed 3 undefined section references; tightened Shamir 2022 framing; softened "three-interpretation closure" → "diagnostic"; added explicit "headline uses p_CW-weighted maps not argmax labels" defense against the 21% D4-TTA flip-rate concern. HF dataset README also pushed to v1.0.123 in the same fire.
• v1.0.124 closed ChatGPT MAJOR M3: new mask pixel-count-threshold robustness sweep (5 cells × MASTER coupling matrix × N=200 monopole-only null) shows the canonical-mask ℓ=1 σ is robust at +6.31 to +8.26 across n_total > {1,5,10,20,50}. Signal does not attenuate at higher thresholds — rules out the low-count-edge artifact interpretation. New §IX.B "Mask robustness" subsection
  • Table V + pre-specified estimator hierarchy paragraph.
• v1.0.125 closed Gemini MAJOR M-2: new DECaLS [0.5,0.6) stratum-specific cross-spectrum C^{An}_ell on the 938,563-spiral stratum (f_sky=0.279). Result: r_ℓ=1 = −0.70 (σ = −1.68), same negative-correlation sign as full canonical −0.49 with larger magnitude. The DECaLS-stratum excess is anti-correlated with its own pixel-density at the same scale as the canonical depth-systematic signature — directly ties the stratum's family-corrected ~2.4σ excess to the depth-correlated systematic family (interpretation ii), not a separate DECaLS-specific physical signal.

Historical R-round record

At the historical Paper IV v1.0.125 state, the project recorded: 3 consecutive 5/5 clean R-rounds (R23 + R24 + R25) across DeepSeek-V4-Pro

• Gemini-3.1-Pro + GPT-5 + Grok-4.3 + Perplexity-Sonar-Pro (45 of 45 reviewers returned 0 BLOCKER / 0 MAJOR). This historical record is not a review result for the current manuscript.

Usage

from huggingface_hub import hf_hub_download
import torch
import torch.nn as nn
import timm

ckpt_path = hf_hub_download(
    repo_id="bamfai/galaxy-chirality-v2",
    filename="chirality_model_v2_best.pt",
    revision="237d021c451d75cf86a875e86d4de498b74e2f12",
)

class Head(nn.Module):
    def __init__(self):
        super().__init__()
        self.h = nn.Sequential(
            nn.LayerNorm(384), nn.Linear(384, 512), nn.GELU(), nn.Dropout(0.3),
            nn.Linear(512, 256), nn.GELU(), nn.Dropout(0.2), nn.Linear(256, 3),
        )
    def forward(self, x):
        return self.h(x)

state = torch.load(ckpt_path, map_location="cpu", weights_only=True)
encoder = timm.create_model("vit_small_patch16_224", pretrained=False,
                            num_classes=0)
head = Head()
encoder.load_state_dict(state["enc"])
head.load_state_dict(state["head"])
encoder.eval()
head.eval()

The checkpoint is a PyTorch dictionary with enc, head, val_acc, epoch, and n_classes entries, not a standalone safetensors model. Exact production preprocessing and equivariant inference are documented in run_eq_dataloader.py. Because the historical environment and training realization are not fully pinned, this loading example does not claim standalone exact reproducibility.

Companion catalog dataset

Per-galaxy CW/CCW/NS labels + per-class equivariant ranking scores: bamfai/galaxy-chirality-catalog (see the dataset card for its pinned data/provider commits; no current manuscript tag is claimed here).

Citation

@misc{golden_chirality_2026,
  author = {Houston Golden},
  title  = {An Observed-Label Chirality-Dipole Null in 949,584 High-Confidence
            DESI Spirals and an 8.5-Million-Galaxy Catalog},
  year   = {2026},
  url    = {https://github.com/Hubify-Projects/bigbounce/blob/main/pipelines/p2_chirality/chirality_catalog_paper.tex},
  note   = {Paper IV, current manuscript; no DOI assigned}
}

Bias and limitations

• Applied to DESI Legacy DR8 imaging. Generalization to other surveys (HSC, KiDS, future LSST) requires re-validation.
• The historical training realization is not exactly reproducible: the committed README reports 92.10% three-class validation accuracy, while an immutable-repository audit reconstructs 93.6878% with 21,293 training and 5,323 validation rows. Conflicting total/non-spiral counts, absent retained object/split membership, and missing random-state/run receipts prevent either metric from being treated as a resolved, exactly replayable validation.
• A residual catalog-wide CW-fraction offset of −0.002647 (a −9.47σ deviation from 0.5 at N = 3.2M chirality-relevant spirals) is attributable to classifier-residual bias and is documented in Paper IV §VI.E.
• For environment-dependent chirality, see the companion Paper V (pipelines/p5_desi_chirality/paper/p5_desi_chirality.tex).

Provenance

• Current manuscript: Paper IV source linked above.
• Audited model revision: 237d021c451d75cf86a875e86d4de498b74e2f12.
• Historical card state: v1.0.125 framing is retained above only as a labeled changelog; it is superseded by the current manuscript framing.
• Publication identifiers: no immutable tag for the current manuscript, arXiv identifier, or DOI is claimed.