PUBLICATION-STAGE LITERATURE, ATTRIBUTION AND ORIGINALITY REVIEW
18 October 2026 (America/Chicago). Independent AI review. WORKING RECORD.

Source family and scope
The retained analytic baseline was read and its SHA256 verified:
D:/LUMINARY/research/self_transport_nonuniform_revision2_20261018/revised_stopping_theorem.txt
7d2ed0072718a9fbb84581ed4c67fe75d4781468592bafa3198e163bbcd12edd.
The stopped calibration branch was read; its findings source hash was verified:
D:/LUMINARY/research/self_transport_finite_calibration_20261018/FINITE_CALIBRATION_FINDINGS.txt
db4525f5344d5add32130cbd253bb3106a8fe90ed86187bd0a83c474c7171de8.
Prior FINAL_ORIGINALITY_AND_VALUE.txt, the literature/application ledger,
project-context family/overlap/publication records and public predecessor page
were inspected. This reviewer owns only manuscript/references.tex and this
review. It does not independently replace the analytic proof reviewers or
perform external human peer review. Final manuscript source binding is pending.

Fresh discovery and access
Actual Chrome Google discovery was performed, and visible results read, for:
  "self-transport" "stopping" "score"
  Sinkhorn score estimation statistical stopping asymptotic normality
  "Sinkhorn" "residual" "central limit" stopping
The first query was weak; broadening led to the Maeda bridge-iteration work,
Belomestny empirical-potential work, Goldfeld et al. limit theory and a new
partial-transport optimization paper. Snippets and AI overviews were leads,
not theorem evidence. Supplemental web search (engine unspecified) checked
bibliographic metadata; it is not described as Google. Primary versions and
sections were read as identified below. The earlier broad research audit was
reused explicitly, rather than misrepresented as a fresh exhaustive reread.

Primary comparisons and attribution
1. Alex Kokot, Variational Problems in Feature Selection and Data Compression,
University of Washington PhD dissertation (2026), publicly deposited 16 Sept
2026 according to the preserved primary repository metadata.
https://hdl.handle.net/1773/57893
The archived primary PDF/text and metadata remain in the first nonuniform
folder. Chapter 3 Theorem 3.6, Corollary 3.3 and Lemma 3.29, including their
proofs, were reread: scaling differentiation, the self-diagonal plus-resolvent
and fixed-epsilon Hilbert CLT are direct antecedents. Chapter 4 Section 4.4.1,
Lemma 4.4 and Corollary 4.3 were reread: two population symmetric updates
already recover the converged first-order expansion. These mechanisms require
explicit manuscript attribution. Fresh UW web fetches failed; public deposit
metadata is authenticated reuse from the preserved primary record, not a newly
successful live download. The fixed-epsilon derivative/population expansion
does not itself give the present joint iid trajectory first-hit theorem.
The Chapter 4 conclusion was also checked: it explicitly leaves estimator
variance and joint sample--bandwidth asymptotics as further research. Its
population two-update result therefore is not a hidden empirical stopping
theorem, although it remains a direct antecedent of the initialization work.

2. Gilles Mordant, arXiv:2412.12007v2 (29 April 2026).
https://arxiv.org/pdf/2412.12007v2
https://arxiv.org/html/2412.12007v2
Theorem 12 and Algorithm 1 were refreshed, with their assumptions and model
distinctions retained from the preceding audit. Decreasing-regularization
score CLTs and half-self computation are established. The Euclidean compact
convex model is distinct from the exact periodic Gaussian here. The candidate
does not improve the exact-estimator bandwidth window in dimension two.
Bibliographic detail: the abstract-page metadata title is "The entropic optimal
(self-)transport problem: Limit distributions for decreasing regularization
with application to score function estimation"; PDF page 1 and HTML print
"The statistics of entropic optimal transport with decreasing regularisation".
The bibliography uses the title printed on the exact cited PDF, not a made-up
hybrid. The version ID and link remove ambiguity.

3. Jean Feydy et al., PMLR 89 (2019), 2681--2690.
https://proceedings.mlr.press/v89/feydy19a.html
https://proceedings.mlr.press/v89/feydy19a/feydy19a.pdf
Section 3.1, equation (25), was read directly: the symmetric averaged
log-potential update is the algorithm used here after exponentiation. It is
not introduced by this paper. Its brief stopping discussion is numerical
background, not a joint statistical first-hit CLT.

4. Wormell--Reich, SIAM J Numer Anal 59(3),1687--1734 (2021),
doi:10.1137/20M1344093; arXiv:2006.02037v3.
https://arxiv.org/html/2006.02037v3
Section 4 was refreshed: positivity/PSD, local contraction, stopping and the
ASSA filter -P(I-P)/2 are established. Its local theorem is numbered Theorem 9
in this HTML rendering; a previous author-PDF rendering used Theorem 4.1.
Use the exact source or section reference. The present half-self filter is
(I-P)/2. Related normalized-kernel and torus operator estimates are prior work.

5. Cheng--Landa, Information and Inference 13(4),iaae026 (2024),
doi:10.1093/imaiai/iaae026; arXiv:2206.11386v3.
https://arxiv.org/html/2206.11386v3
Theorem 4.2 and the constrained approximate normalization definitions were
refreshed. Approximate normalization can retain nonuniform graph-Laplacian
statistical consistency; that broad idea cannot be claimed new. Its graph
operator statistic, norm and scaling constraint differ from a fixed-query
score Gaussian law at the actual unrestricted first hit.

6. Celisse--Wahl, JMLR 22(76),1--59 (2021).
https://jmlr.org/papers/v22/20-358.html
https://www.jmlr.org/papers/volume22/20-358/20-358.pdf
The primary metadata and Proposition 13/Theorem 15 were refreshed. Effective
dimension, discrepancy stopping, smoothed residuals and random-design norm
comparison are established. Those regression/inverse-problem results do not
directly prove the present nonlinear self-scaling trajectory.

7. Maeda--Yao--Nitanda, arXiv:2510.22560v1 (26 October 2025).
https://arxiv.org/html/2510.22560v1
Theorems 1 and 2 were refreshed: empirical intermediate bridge iterations
already have statistical plus algorithmic error bounds, for path total
variation. This prevents claims to the first statistical analysis of actual
Sinkhorn iterates. It is not this half-self residual first-hit score result.

8. Fresh recent checks beyond the core bibliography:
Belomestny, arXiv:2608.29152v1, explicitly excludes optimization error from
terminating finite Sinkhorn iterations; the exclusion was read in the primary
text. https://arxiv.org/html/2608.29152v1
Truong et al., Accelerated Sinkhorn Algorithms for Partial Optimal Transport,
arXiv:2601.17196v3, Theorems 1--3 and its problem definition were inspected.
https://arxiv.org/html/2601.17196v3
It addresses deterministic partial-transport feasibility/optimization
complexity, not density-score inference or the current joint iid first hit.
Other previously checked close sources (Greco et al. gradient convergence,
Goldfeld et al. fixed-regularization functional inference, fixed-support
low-temperature paths and spectral count arithmetic) remain recorded in the
preserved baseline review; their full proofs were not newly reread here.

9. Prior author publication, actually read live in Chrome:
Scott Schubr, Residual Certificates and Empirical Sharpness for Entropic
Self-Transport Scores, Luminary Research Archive, edition 1,18 October 2026.
https://luminaryoddities.com/researcharchive/residual-certificates-empirical-sharpness-entropic-self-transport-scores/v/1
The live page identifies the author/affiliation/date, edition and permanent
URL. That work handles arbitrary positive candidate residual-to-score
certificates and sharp witnesses; its scope explicitly excludes a solver
trajectory lower bound or new CLT. The new manuscript should identify itself
as a follow-up addressing the actual half-self trajectory, and preserve the
same result-family relationship rather than inflate editions into discoveries.
This was an observed public-page read, not a fresh complete public ZIP audit.

10. Finite-coverage limitation, if retained in the manuscript:
Dufour, Econometrica 65(6),1365--1388 (1997),
doi:10.2307/2171740, author-hosted full primary paper was refreshed:
https://www2.cirano.qc.ca/~dufourj/Web_Site/Dufour_1994_ImpossibilityTheorems_W.pdf
Locally almost unidentified/discontinuous targets and bounded honest confidence
set obstructions are established mechanisms. The paper's elementary sine
example may clarify the distinction between fixed-density asymptotics and
unrestricted-class finite-sample honesty; it is not claimed as a new main
contribution. No conservative finite-class salvage is marketed as new.

Bounded originality and value assessment
No checked primary theorem subsumes the combined actual first-hit conclusion:
fixed positive C6 nonuniform density on T^d,d=1,2; shrinking exact periodic
Gaussian; same iid data; ordinary b0=1 half-self iterates; every fixed strict
0<gamma<d/4; first residual hit with attained lower bound for any c<1/2,
leading iteration count, negligible point-score numerical error and retained
true-score Gaussian law; uniform-density endpoint counterexample for the
specified coefficient-one threshold. This is a meaningful focused specialist
contribution if the independent analytic review verifies the exact final proof.

The identifiable increment is the joint nonlinear empirical trajectory and
residual crossing control. The iteration count is logarithmic inversion once
two-sided residual control is proved. Slutsky studentization/confidence regions
are standard useful consequences. Neither the algorithm, RKHS resolvent,
population two-step mechanism, spectral discrepancy principle, score inference
in general nor geometric count arithmetic is new. No worldwide priority
guarantee, new minimax result, adaptively selected query guarantee, optimal
bandwidth theorem, finite-sample calibration, global diffusion/path guarantee
or practical runtime speedup is supported.

Fresh final-manuscript attribution check and source hashes: PENDING.
