Publications

Publications

Everything below is published on Zenodo, a free, permanent research repository operated by CERN. Each entry has a DOI (Digital Object Identifier) — a permanent, citable reference number that ensures the record can always be found, even years from now. See the FAQ for more.

The series is numbered sequentially (DOI 1.0, 1.1, 1.2, ...) in the order each report was completed, and listed newest first below. When a report corrects or supplements an earlier one, both entries say so explicitly — the full history is summarized in the correction trail at the bottom of this page.

Every entry includes the exact SQL queries used, cryptographic hashes of the underlying data, and — for more recent reports — unedited screen recordings of those queries actually running, so every published number can be independently checked.

Preprint

• SSRN Closed-Form Critical Points of SSIM: Deriving and Validating the Boundaries of Metric Reliability SSRN, abstract=7291266

Technical report series (Zenodo, DOI 1.0–1.21)

• DOI 1.212026-10-10A higher-precision pipeline for measuring sigma_gt/mu_gt's distinct-value countDetails ▾Hide ▴

  sigma_gt/mu_gt take only finitely many distinct values because of the GT image's 8-bit quantization -- that property itself is correct. However, the specific figures reported in Technical Reports v1.16/v1.18/v1.19 (approximately 2.3 million distinct values for sigma_gt, 68,649 for mu_gt) turn out to have been inflated by about 1.95x and 4.2x respectively, due to a separate issue unrelated to quantization: floating-point addition's non-associativity producing "ghost duplication" (the same true value recorded as slightly different bit patterns depending on the exact computation path). This report demonstrates the mechanism with concrete examples and presents a higher-precision pipeline -- tolerance-based adjacent-gap analysis (the "LAG method") -- for computing the true, quantization-derived distinct-value count (sigma_gt≈1.18 million, mu_gt≈16,320), confirmed byte-for-byte identical across all four models.

  Includes: PDF report (EN/JA), 10 screen-recorded videos, 10 verification SQL scripts, 2 Python DB-construction scripts, file hashes.

  View full report on Zenodo: https://doi.org/10.5281/zenodo.23271571

• DOI 1.20.12026-09-29A ground-truth-free binary classifier for Structural Masking riskDetails ▾Hide ▴

  comp_contrast, one of this series' eight independently-measured quality features (Technical Report v1.8), turns out to be RMS contrast computed on the raw 8-bit pixel scale rather than the [0,1]-normalized scale used for sigma_gt/sigma_notta throughout this series -- dividing by 255 recovers sigma_gt from comp_contrast_gt exactly, with zero error across all four models' full sigma_gt≤0.16 populations (81,564,041 blocks each). By the same construction, comp_contrast_notta/255 is an exact, algebra-free measurement of sigma_notta -- not an approximation, unlike an earlier candidate approach (a quadratic-equation recovery with an unresolvable root-selection ambiguity). A single untrained threshold, sigma_notta≤sigma_stat, classifies Structural Masking risk at 98.83-99.53% accuracy across all four models, with masking-region recall at 99.09-99.95%. Recall on the minority non-masking class is lower and model-dependent (42.79-78.76%): because sigma_notta is measured exactly here, this reflects real reconstruction behavior, not measurement error.

  Plain-language explainer: Catching SSIM's blind spots without ground truth

  Includes: PDF report (EN/JA), 1 screen-recorded video, verification SQL (consolidated script), documentation markdown (EN/JA), file hashes.

  View full report on Zenodo: https://doi.org/10.5281/zenodo.23023733

• DOI 1.202026-09-26A ground-truth-free binary classifier for Amplification Phase riskDetails ▾Hide ▴

  This series has established μ_stat as a model-specific, empirically-calibrated transition point marking SSIM's Amplification Phase, the region where the luminance term becomes hypersensitive to small absolute deviations between ground-truth mean brightness μ_gt and a reconstruction's own mean brightness μ_notta. Diagnosing this has so far required μ_gt itself, i.e. ground truth. This report tests a ground-truth-free alternative: reframing the question as binary classification (is this block at risk, yes or no) rather than exact value estimation sidesteps an identifiability problem tied to μ_gt's 8-bit quantization. A single untrained threshold, μ_notta ≤ μ_stat, with μ_notta recovered algebraically from a stored SSIM component, classifies Amplification Phase risk at 96–100% accuracy across all four models, with recall at or above 99.5% throughout. Offered as a validated diagnostic method for a narrow use case, not as a claim that ground-truth-free SSIM diagnosis is possible in general.

  Plain-language explainer: Catching SSIM's blind spots without ground truth

  Includes: PDF report (EN/JA), 1 screen-recorded video, verification SQL (consolidated script + 4 original exploratory scripts), file hashes.

  View full report on Zenodo: https://doi.org/10.5281/zenodo.22968572

• DOI 1.19.12026-09-19Release of the DB-building pipeline, for reuse on any datasetDetails ▾Hide ▴

  Releases, unedited and in its 2026-09-01 state, the exact code pair (build_notta_tta_comp_db.py and heatmap_tool.py) that generated every number published in Technical Reports v1.12 through v1.16. Anyone who reads the formulas and methods already published in this series could in principle reimplement equivalent code themselves; this release is for those who don't want to spend the effort building it from scratch. It accepts any GT/notta/tta image folder, not just this project's own dataset, so anyone can build the same kind of database and run the same kind of SQL investigation used throughout this series on their own data and models. Hardcoded values (σ_stat) are included and meant to be rewritten for one's own dataset and model.

  Includes: PDF report (EN/JA), 2 Python scripts, 1 screen-recorded video, requirements.txt, LICENSE.

  Documentation on this site: docs/build_notta_tta_comp_db.html (see also the beginner-friendly walkthrough of the database-building code)

  View full report on Zenodo: https://doi.org/10.5281/zenodo.22838600

• DOI 1.192026-09-02Four-model video evidence supplement for two discreteness claimsDetails ▾Hide ▴

  This is a supplement, not a correction — no value in DOI 1.16 or 1.18 changes. Both reports state that a ground-truth-only quantity (σ_gt for DOI 1.16, μ_gt for DOI 1.18) is quantized identically across all four models, since it never depends on any model's reconstruction — but neither report backed that specific claim with this series' usual screen-recorded, hash-verified evidence. DOI 1.16's text stated it as an assumption with no cross-check at all; DOI 1.18's text claimed an improvement (an actual Standard/SwinIR comparison), but that check covered only two of the four models. This supplement closes both gaps: the identical query for each quantity, run against all four models in one continuous screen recording per quantity, confirms byte-for-byte identical results across Standard, SwinIR, Upscayl High-Fidelity, and Digital-Art in both cases.

  Includes: PDF report (EN/JA), 2 screen-recorded videos, 2 SQL files, file hashes.

  SQL walkthrough on this site: docs/sql/doi-1-19.html

  View full report on Zenodo: https://doi.org/10.5281/zenodo.22252098

• DOI 1.182026-09-01Completing the SSIM sub-term picture: the luminance termDetails ▾Hide ▴

  SSIM's structure term S and contrast term C have each been shown to lock to exactly 1 under an identifiable condition regardless of image content (DOI 1.11, 1.15, 1.16). This report completes the picture with SSIM's third and final multiplicative term, luminance L: a previously-unwritten proof shows L=1 if and only if μ_gt=μ_notta, a mathematical equivalence mirroring the contrast-term theorem exactly. Searching all four models finds real μ_gt=μ_notta ties at rates from 0.218% to 6.512% of blocks, the large majority genuinely different content that merely shares the same mean brightness; μ_gt is shown to be even more sharply quantized than σ_gt. For two of the four models, concrete blocks exist where L=1 exactly yet the aggregate SSIM score is negative — a coincidental tie in one sub-term provides no protection when the remaining terms are simultaneously degraded. Finally, the report closes a further question: since L, C, and S are each bounded above by exactly 1 and strictly positive, their product can equal exactly 1 only if all three do — so a coincidental joint alignment of all three terms for genuinely different content is not merely unobserved, it is algebraically impossible. Confirmed directly: of 4,734,946 blocks with SSIM=1.0 exactly across all four models, all but 1,442 (a SwinIR-only floating-point rounding artifact, diagnosed and confirmed harmless) also had all three terms bit-exact at 1.0.

  Includes: PDF report (EN/JA), 3 screen-recorded videos, 6 SQL files, file hashes.

  SQL walkthrough on this site: docs/sql/doi-1-18.html

  View full report on Zenodo: https://doi.org/10.5281/zenodo.22230202

• DOI 1.162026-08-20Generalizing the masking identity beyond σ_gt=0Details ▾Hide ▴

  Technical Report v1.11 proved that SSIM's structure term S is mathematically pinned to exactly 1 whenever σ_gt=0. Technical Report v1.15 extended this to SSIM's contrast term C, proving C=1 exactly whenever σ_gt=σ_notta, for any shared value, not only zero. This report completes that result in two ways: a previously-unwritten converse proof that C=1 is not merely correlated with σ_gt=σ_notta but mathematically equivalent to it, and a direct database search across all four models finding that nonzero σ_gt=σ_notta ties occur at a real, quantified rate (0.019%–0.342% of blocks), the large majority genuinely different content that merely shares the same local variance by coincidence. The root cause is traced empirically to 8-bit pixel quantization: only 2,308,415 distinct σ_gt values exist across 79,321,389 blocks per model, and 76.1% of blocks share their exact value with 100 or more others.

  Includes: PDF report (EN/JA), 1 screen-recorded video, 3 SQL files, file hashes.

  SQL walkthrough on this site: docs/sql/doi-1-16.html

  View full report on Zenodo: https://doi.org/10.5281/zenodo.22029346

• DOI 1.152026-08-19Testing the converse: does a feature match guarantee SSIM=1.0?Details ▾Hide ▴

  Technical Report v1.13 asked whether an SSIM score match guarantees content agreement, and found it does not, always. This report tests the converse: if a reconstruction's eight comp_* features match the ground truth's exactly, does SSIM still equal 1.0? The answer is again conditionally yes: SwinIR sits at 99.97%, but Upscayl High-Fidelity has 63.05% of its feature-matched blocks below SSIM=0.5. Tracing the mechanism: the SSIM contrast term C is proven — algebraically — to equal exactly 1.0 whenever σ_gt=σ_notta (a new general theorem), and the structure term S stays near 1.0 throughout; only the luminance term L remains free to vary, and its spread is the entire content of the observed divergence. Root cause: none of the eight comp_* features measures luminance at all. Together with v1.13, this establishes the argument in both directions — neither an SSIM match nor a comp_* match alone guarantees the other.

  Includes: PDF report (EN/JA), 1 screen-recorded video, 6 SQL files, file hashes.

  SQL walkthrough on this site: docs/sql/doi-1-15.html

  View full report on Zenodo: https://doi.org/10.5281/zenodo.22013446

• DOI 1.142026-08-19Four-model verification of the Failure Fingerprint statisticsDetails ▾Hide ▴

  Errata and supplement to Technical Reports v1.10 and v1.12, completing the Failure Fingerprint correction for all four models. Standard's σ_stat is corrected from 0.0800 to 0.101 (confirming v1.12); SwinIR and Digital-Art are confirmed fully unaffected; Upscayl High-Fidelity is severely affected on both axes — μ_stat corrected from 0.1400 to 0.0100, σ_stat from 0.0800 to 0.0860. No two models share the identical impact pattern, confirming the bug's consequences are model-specific rather than a single fixable constant. A numerical coincidence between High-Fidelity's and SwinIR's corrected μ_stat values is investigated and resolved as a boundary -condition artifact, not a genuine connection. Together with v1.12, this demonstrates the join-based pipeline's risk was real and model-specific, and the new join-free pipeline is a validated correction.

  Includes: PDF report (EN/JA), 2 screen-recorded videos, verification SQL, file hashes.

  SQL walkthrough on this site: docs/sql/doi-1-14.html

  View full report on Zenodo: https://doi.org/10.5281/zenodo.22012409

• DOI 1.132026-08-16Two empirical tests of score-vs-content agreementDetails ▾Hide ▴

  SSIM aggregates three multiplicative components (L·C·S) into a single number, a lossy compression by construction: two blocks with different underlying content can share the same score. This report tests, at full dataset scale, whether that theoretical possibility actually matters in practice, via two independent comparisons at increasing perturbation size. Finding 1 (same model, TTA on/off): feature agreement holds almost always for three of four models (98.75%–99.28%), but only 94.21% for Upscayl High-Fidelity. Finding 2 (different models, same coordinate): restricting to exact score ties, feature agreement is 99.85%–99.98% for pairs not involving High-Fidelity, but only 67.14%–81.93% for pairs that do. Both tests converge: SSIM score agreement does not, by itself, guarantee content agreement — and Upscayl High-Fidelity is consistently the least predictable of the four models.

  Includes: PDF report (EN/JA), 3 screen-recorded videos, 17 SQL files, file hashes.

  SQL walkthrough on this site: docs/sql/doi-1-13.html

  View full report on Zenodo: https://doi.org/10.5281/zenodo.21960019

• DOI 1.122026-08-11sigma_stat correctionDetails ▾Hide ▴

  Errata and supplement to Technical Report v1.10. This document corrects the empirically-detected texture-axis transition point σ_stat reported for Upscayl Standard, computed against a database later found to contain a data-integrity bug: a comp_* feature -extraction pass had attached a different, non-zero block's values to the wrong row — producing plausible-looking but incorrect numbers the existing is_comp_zero flag could not detect. An independent, from-scratch verification database (no join step at all) was built for both Upscayl Standard and SwinIR to re-derive σ_stat under the exact original methodology. Result: Standard's σ_stat is corrected from 0.0800 to 0.101 (+0.021); SwinIR's is confirmed unaffected at 0.0800. Both models were verified via live SQL execution, screen-recorded and hash-verified. Upscayl High-Fidelity and Digital-Art are outside this report's scope.

  Includes: PDF report (EN/JA), 2 screen-recorded videos, 4 SQL files, file hashes.

  SQL walkthrough on this site: docs/sql/doi-1-12.html

  View full report on Zenodo: https://doi.org/10.5281/zenodo.21894585

• DOI 1.12.12026-08-12Verification SQL supplement for the SSIM=1.0 transitivity corollaryDetails ▾Hide ▴

  A supplement, not a correction — no value in DOI 1.12 changes. Section 4.4 of that report used a corollary of the Pixel Identity Theorem as a cross-check: blocks where SSIM(GT,NOTTA)=1.0 and SSIM(GT,TTA)=1.0 hold simultaneously are forced to be pixel-identical to GT on both sides, and therefore to each other. DOI 1.12's own text reported this holding exactly for Upscayl Standard (702,108 blocks, zero exceptions) but the SQL behind it was not included in that report's bundle, and no SwinIR figure was stated. This supplement adds the SQL, re-runs it, and extends the check to SwinIR — matching DOI 1.12's own two-model scope — finding the identical pattern across 2,657,100 blocks, again with zero exceptions.

  Includes: PDF report (EN/JA), 2 screen-recorded videos, 4 SQL files, file hashes.

  View full report on Zenodo: https://doi.org/10.5281/zenodo.21903024

• DOI 1.112026-08-07The masking blind spot theoremDetails ▾Hide ▴

  This report proves a closed-form identity in SSIM: when a ground-truth image block is perfectly flat (σ_gt=0), the SSIM structure term S is mathematically forced to exactly 1, regardless of what the reconstructed block contains — not an approximation, an exact algebraic consequence of the covariance formula. Validated against every σ_gt=0 block in the project's full evaluation database (655,884,288 blocks across all four models under both noTTA and TTA combined): 19,688,824 such blocks, zero exceptions across all eight model×mode combinations. It then checks whether reconstructions actually place non-trivial, hallucinated content into blocks where this blind spot applies — they do, to substantially different degrees across models: Upscayl High-Fidelity's flat-ground-truth population shows hallucinated detail with a maximum nearly 23× SwinIR's maximum, a difference S is mathematically incapable of registering.

  Includes: PDF report (EN/JA), 1 screen-recorded video, verification SQL, file hashes.

  SQL walkthrough on this site: docs/sql/doi-1-11.html

  View full report on Zenodo: https://doi.org/10.5281/zenodo.21837260

• DOI 1.102026-07-29Correction of the C1/C2/C3 constant bug Consolidated correction covering DOI 1.0/1.4/1.9/1.9.1Details ▾Hide ▴

  Errata and supplement to Technical Reports v1.0, v1.4, v1.9, and v1.9.1. This document corrects an implementation error in the SSIM stabilization constants C1, C2, C3 used throughout the database generation pipeline since the project's inception, and recomputes every value in the four reports above that depends on the theoretical critical points μc and σc. The constants had omitted the squaring required by the standard Wang et al. (2004) convention Ci=(KiL)², giving values roughly 100×/33× too large; the correction gives μc≈0.0071, σc≈0.0212 (previously reported as μc≈0.0707, σc≈0.1225). All four affected models were regenerated in full (5,004 images and 163,971,072 blocks each) and re-verified via live SQL execution, screen-recorded and hash-verified. A second, independent finding made during this full-dataset regeneration — a content-population discontinuity in the texture axis beyond σ_gt≈0.16 — is also documented and corrected for.

  Includes: PDF report (EN/JA), 4 screen-recorded videos, verification SQL, 2 Python scripts, file hashes.

  SQL walkthrough on this site: docs/sql/doi-1-10.html

  View full report on Zenodo: https://doi.org/10.5281/zenodo.21658621

• DOI 1.9.12026-07-20Visual Failure Fingerprint plot for High-Fidelity and Digital-Art Supplement to DOI 1.9; later consolidated into DOI 1.10Details ▾Hide ▴

  A supplement to Technical Report v1.9 that visualizes its already-verified Δμ/Δσ values relative to the theoretical critical points μ_c, σ_c established in DOI 1.0, adding two simple trigonometric derivations (axis angle and inter-vector angle). No new data is introduced — only a diagrammatic view of existing verified numbers.

  Includes: PDF report (EN/JA), file hashes.

  View full report on Zenodo: https://doi.org/10.5281/zenodo.21445635

• DOI 1.92026-07-19Live SQL verification of Failure Fingerprint values (High-Fidelity / Digital-Art) Later consolidated into DOI 1.10Details ▾Hide ▴

  Documents dataset statistics and Failure Fingerprint values for two additional models — Upscayl High-Fidelity (142.8 million blocks) and Upscayl Digital-Art (164 million blocks) — computed via live SQL execution against the source data, with unedited screen-capture video as direct evidence the values weren't fabricated or adjusted after the fact.

  Includes: PDF report (EN/JA), 2 screen-recorded videos, verification SQL, file hashes.

  SQL walkthrough on this site: docs/sql/doi-1-9.html

  View full report on Zenodo: https://doi.org/10.5281/zenodo.21433900

• DOI 1.82026-07-19The comp_* feature framework: definitions and implementationDetails ▾Hide ▴

  Formalizes the methodology behind eight block-level quality-component features (plus two supplementary frequency-domain descriptors) extracted from every 8×8 image block — features that had previously existed only in reference code, not as a documented specification. This is the definitional source for the comp_* feature framework referenced throughout the rest of this series and described on the Technology page.

  Includes: PDF report (EN/JA), file hashes.

  View full report on Zenodo: https://doi.org/10.5281/zenodo.21433447

• DOI 1.72026-07-18Axis5 collapse intensity and Tier3 state distribution by Axis1 polarityDetails ▾Hide ▴

  Examines Axis5 (a Mahalanobis-distance-based collapse -intensity measure) feature weights for four architecturally distinct AI super-resolution models, and how Tier3 state distribution varies by Axis1 polarity classification. Approximately 82 million blocks from 5,004 chest X-ray images, across four upscaling architectures including SwinIR and Real-ESRGAN, with all data re-derived to ensure accuracy.

  Includes: PDF report (EN/JA), 4 screen-recorded videos, 2 Python scripts, file hashes.

  SQL walkthrough on this site: docs/sql/doi-1-7.html

  View full report on Zenodo: https://doi.org/10.5281/zenodo.21425511

• DOI 1.62026-03-17Universal Tier3 classification coverage, with a DOI 1.0–1.6 research overviewDetails ▾Hide ▴

  Validates complete (100%) coverage of a four-state block classification (Stable, Persistent, Escape, BlindSpot) across the full 81.9-million-block database, all four models under both no-augmentation and test-time-augmentation conditions, via three independent SQL diagnostic checks confirming no contradictory assignments. Also includes a Research Architecture Overview mapping the entire DOI 1.0–1.6 series across four research phases and explicitly documenting the correction relationships between versions — the direct ancestor of the "Correction trail" summary below.

  Includes: PDF report (EN/JA), 2 screen-recorded videos, file hashes.

  SQL walkthrough on this site: docs/sql/doi-1-6.html

  View full report on Zenodo: https://doi.org/10.5281/zenodo.19058174

• DOI 1.52026-03-16Viewer systems and a corrected collapse-type classificationDetails ▾Hide ▴

  Documents two visualization tools built for this project: a local 3D multi-axis viewer for browsing collapsed blocks across SSIM's component axes (with heatmap and animation views for all four models), and a local web application for interactively browsing the DuckDB databases via a REST API and per-image heatmaps. Alongside these, presents a corrected block-classification analysis across all four models (81,975,528 blocks each): certain collapse types ("Silent Dual" and "Silent Multiple") turn out to exist universally across models, and Upscayl High-Fidelity shows an anomalous 86.53% prevalence of one classification ("Silent None") that the other models do not share.

  Includes: 5 PDF documents, 16 screen-recorded videos, file hashes.

  SQL walkthrough on this site: docs/sql/doi-1-5.html

  View full report on Zenodo: https://doi.org/10.5281/zenodo.19040901

• DOI 1.42026-03-13SQL corrections and two sigma_stat approaches Errata and supplement to DOI 1.0; later consolidated into DOI 1.10Details ▾Hide ▴

  Errata and supplement to Technical Report v1.0, covering four areas: corrects a quotation-mark rendering issue in one SQL query from PDF conversion; clarifies that computing μ_stat requires two sequential stages (an SQL step, then a Python two-segment optimal -split procedure); archives an unedited screen recording of eight DuckDB CLI queries as video evidence; and documents two distinct σ_stat approaches (average SSIM per σ_gt bin vs. standard deviation of SSIM differences per bin), with failure-fingerprint maps for both Upscayl Standard and SwinIR under each approach.

  Includes: PDF report, 2 screen-recorded videos, the corrected SQL, file hashes.

  SQL walkthrough on this site: docs/sql/doi-1-4.html

  View full report on Zenodo: https://doi.org/10.5281/zenodo.18994113

• DOI 1.32026-03-11Axis1 bias direction is architecture-dependent, not universalDetails ▾Hide ▴

  Empirically validates an internal bias direction ("Axis1") across four AI upscaling architectures, 328 million blocks from 5,004 chest X-ray images. Finds two architecture-dependent groups with opposing bias polarity: Upscayl Standard and SwinIR show positive Axis1 driven by artifacts (ringing, blocking), while Upscayl High-Fidelity and Digital-Art show reversed polarity, where positive Axis1 instead reflects sharpness enhancement. The conclusion: this bias direction cannot be interpreted uniformly across different architectures, correcting and extending earlier single-model findings.

  Includes: 2 PDF documents, 4 screen-recorded videos, file hashes.

  View full report on Zenodo: https://doi.org/10.5281/zenodo.18949243

• DOI 1.22026-03-09Structural failure theory: visible and hidden collapse across four architecturesDetails ▾Hide ▴

  Examines fundamental limitations in SSIM using 328 million blocks across four image restoration architectures (three Upscayl variants and SwinIR). Identifies two systematic biases — luminance instability and structural masking — that create detection blind spots, and a "two-layer collapse structure" distinguishing observable degradation (visible collapse) from severe internal degradation masked by a high SSIM score (hidden collapse). 95.30% of SwinIR's hidden persistent failures occur specifically in regions where SSIM becomes mathematically unresponsive.

  Includes: PDF report, 6 screen-recorded videos, file hashes.

  View full report on Zenodo: https://doi.org/10.5281/zenodo.18917078

• DOI 1.12026-03-06Statistical infrastructure for the 82-million-block SSIM blind-spot databaseDetails ▾Hide ▴

  Establishes the statistical infrastructure behind the project's block-level analysis, moving from a single scalar score to multidimensional structural diagnostics. Covers four AI upscaling models, each run through two processing paths (no-augmentation and test-time augmentation), for roughly 656 million data points in total, analyzed via a DuckDB backend built to make internal failure modes invisible to conventional aggregate metrics tractable at this scale. SHA-256 hashes of all raw output files are included to establish priority and data integrity.

  Includes: PDF report, 5 screen-recorded videos, file hashes.

  View full report on Zenodo: https://doi.org/10.5281/zenodo.18883267

• DOI 1.02026-03-02First report in the seriesDetails ▾Hide ▴

  This study introduces the SSIM Limit Theory, establishing a mathematical framework for analyzing the degradation of structural similarity. To capture the underlying distributional shifts that drive SSIM collapse, we propose Statistical Breakdown Fingerprints (Δμ, Δσ), a concise representation of these changes. The dataset accompanying this work includes the full theoretical manuscript, database schema, reproducible SQL resources, and integrity-verified data blocks that support transparent scientific validation. These results are built upon an exhaustive analysis of over 82 million records and are presented alongside a demonstration video verifying the system's operation.

  This series is registered in Zenodo's curated Artificial Intelligence and Machine Learning community.

  Includes: PDF report, 2 screen-recorded videos, 2 verification databases, file hashes.

  View full report on Zenodo: https://doi.org/10.5281/zenodo.18830286