END-TO-END POSTER AUDIT — 11 OCTOBER 2026 Research companion for Duho Kim's KAS 2026 Fall poster, v68. Scope This audit follows the poster from theory and retained data through methods, inference, maps, wording and delivery. It reuses saved classifications and fitted vectors. It does not repeat image classification, velocity fits or random streams. Model reviews are checks of code and interpretation, not conventional peer review or independent observational confirmation. What changed on the poster The two displayed galaxy maps use their own full resolved/assigned samples: 30,910 for SpArcFiRe and 49,007 for Fourier, from 49,211 targets. A separate comparison uses the same 30,910 galaxies; 18,791 labels agree (60.8%). The revised footnote and first takeaway distinguish that matched comparison from the unequal-sample panels. No map values, signs, images or methods changed. The historical SpArcFiRe calibration combines Galaxy Zoo with a development-exposed CzSL expert-labelled comparison. Its estimate -0.06% [-4.44%, +4.32%] assumes transferable shared class-conditional errors and equal CW/ACW class resolution (d=0). The wider [-19.07%, +11.10%] range permits cohort-specific errors under a different model and joint confidence construction. The wider interval is not solely a measure of sample heterogeneity, and neither interval calibrates the current Fourier classifications. New diagnostic 1: motion-method agreement The regression and half-plane methods use the same saved systems. We measured the directed angle between their projected maximum-velocity vectors, without folding opposite directions together. Frozen categories are less than 45 degrees, 45 through 135 degrees, and greater than 135 degrees. All counted vectors were defined; none lay within the predeclared boundary tolerance. For the 1,722 displayed systems, 1,705 (99.0%) differ by less than 45 degrees, 17 lie in the middle category, and none differ by more than 135 degrees. The median angle is 10.43 degrees. Among the 1,713 retained groups the counts are 1,696 / 17 / 0; all nine retained clusters fall below 45 degrees, with a median of 3.42 degrees. Among all 1,929 fitted groups, 1,909 fall below 45 degrees, 20 in the middle, and none above 135 degrees; median 10.46 degrees. These are descriptive agreements of two correlated projected-motion estimators. They are not accuracy against human labels, independently detected rotations, or full three-dimensional spins. New diagnostic 2: screening and cluster addition H is the dimensionless pair-handedness statistic, not a CW/ACW fraction. The following are pooled H values multiplied by 10,000; they are not the hemisphere-map DeltaH values. Set N Regression Half-plane A: all fitted groups 1929 -2.741947 -3.205278 B: groups retained by joint screen 1713 +0.273061 -0.322342 C: all fitted groups plus all clusters 1939 -2.066542 -2.439433 D: retained groups plus clusters 1722 +0.673972 +0.210983 The screen is the original joint group-plus-cluster proximity screen. B is not the result of applying a new group-only screen. It excludes 216 groups and one cluster; D retains 1,713 groups and nine clusters. Along A -> B -> D, screening is sufficient to change the regression sign, while half-plane H remains negative until the retained clusters are included. Along A -> C -> D, adding all ten clusters to the unscreened groups leaves both H values negative; screening then changes both signs. Each set uses its own N(N-1)/2 denominator, so the increments depend on the path. This is a reconstruction of the arithmetic effect of the fixed sample choices, not causal attribution or evidence that one selection is physically correct. For D, the contributions using the same full D pair denominator are: Pair class Regression Half-plane Pair count Group-group +0.270213 -0.318980 1,466,328 Group-cluster +0.445736 +0.569609 15,417 Cluster-cluster -0.041977 -0.039645 36 Total +0.673972 +0.210983 1,481,781 The positive combined result is not evidence that the cluster-only statistic is positive. The cross-pair term matters, and the three pair classes are dependent. The producer and independent verifier each ran once. All 13 verification checks passed, including direct pair-kernel recomputation, exact pair counts, frozen coverage, saved-value anchors, directed-angle summaries and ownership checks. No new significance test, uncertainty interval, refit or randomization was performed. Statistical interpretation Neither galaxy method establishes a statistically significant preferred sky direction under the retained conditional tests. Bonferroni-adjusted p=1 is a capped multiplicity correction; it is not a 100% probability of randomness, exact symmetry or correct labels. The two methods can differ descriptively without a significant directional detection. The source of their disagreement remains unresolved. Galaxy colour is DeltaA in percentage points: the winding imbalance in the toward hemisphere minus that in the opposite hemisphere. It differs from a fitted dipole amplitude and from a calibration-adjusted pooled estimate. Group/cluster colour remains 10^4 times DeltaH; it is not the percentage of rotating systems. The displayed 1,722-system sample has no evaluated global significance or measurement confidence interval. Existing leave-one-system ranges are sensitivity only. Axis reversal swaps the halves, so contrast changes sign by construction. Antipodal pairs and neighbouring map cells are correlated, not independent observations. Theory, sources and illustrations Parity-symmetric LambdaCDM permits gravitationally generated spin correlations without requiring a population handedness preference. Primordial-chirality simulations and rotating black-hole-universe proposals motivate questions; this study neither fits nor confirms those cosmological models. Image winding is a proxy and does not determine full three-dimensional angular momentum. For the H geometry used here, simultaneous spatial inversion of data directions (n,g) -> (-n,-g) reverses H. At a FIXED candidate dividing axis, it also swaps the hemisphere memberships, so DeltaH is unchanged. If the candidate axis is inverted too, that last conclusion changes. The research-page statement now specifies the fixed-axis convention. The four galaxy examples are retained development illustrations with stored arm fits; they are not independent validation. Their recorded r magnitudes are 15.051883, 15.96085, 17.63897 and 17.689562, in CW, ACW, CW, ACW order. No new private image inspection was performed. Both cluster motion estimators use the original selected cluster membership. The inspected source did not support a reviewer's suspected membership mismatch. Historical estimator-specific direction adapters were traced; no sign was changed. Independent review and adjudication Claude Code supplied the initial source/math review, diagnostic and editorial source drafts in separate authoring calls, and independent source/control reviews. Fresh review contexts give procedural independence, not independence from the model family or from shared data. Root checked the original source and retained results before accepting or rejecting findings. Gemini supplied a limited text and primary-reference review; it did not inspect the full packet or rendered poster. A conflation of historical SpArcFiRe/CzSL calibration with current Fourier references was rejected from the original evidence. Kimi's one existing-credit request ended at its completion cap without a usable answer. It supplied no new opinion or approval. The conservative cost bound was $0.2499948, below the $0.25 cap; no retry was made. Earlier Kimi feedback remains historical context only. A reviewer's claimed pypdf keyword incompatibility was disproved by the actual pinned pypdf 6.13.2 source. The accepted report preflight nevertheless checks the entire PDF edit, serialization and readback in memory before consuming the report stage. Other genuine binding and ownership checks were strengthened. Remaining limits and print conditions The historical compiled classifier/configuration and absolute-orientation chain is not fully established by the retained synthetic and relative-image checks. Calibration transport and class-dependent resolution across the full population remain assumptions. Selection, footprint, measurement error and dependent estimators prevent a physical spin-parity detection claim. An audit finding no demonstrated sign bug is not proof that all classifications are physically correct. The poster remains A0 portrait. Its Standard 14 Helvetica/Helvetica-Bold fonts are not embedded; it is not certified PDF/X. A printing service should accept or preflight those fonts. Release requires exact new/retained text checks, identical images and links, zero pixel changes outside the three approved text regions, actual visual inspection, and working current and archived downloads. No phone viewer or mean map is generated. Contributions Duho Kim provided the original idea and supervised the study. The Codex-led AI workflow performed implementation, analysis, validation checks, figures and writing, with the other models' actual contributions recorded above. Observational data, human reference labels and existing software remain credited to their original sources. Primary sources Lee & Pen (2001): https://arxiv.org/abs/astro-ph/0008135 Shim et al. (2025 revision): https://arxiv.org/abs/2406.06080v2 Poplawski (2025 revision): https://arxiv.org/abs/1910.10819 Research and original references: https://nebulamind.net/studies/spin-parity-2026.html