Documentation
Plain-language explanations of the code and SQL behind our published results — what each file does, without requiring you to read the code itself. Every file we've ever published is meant to end up here eventually, including superseded versions, clearly marked, so an old file is never mistaken for the current one.
A plain-language walkthrough of two ground-truth-free binary classifiers for catching SSIM's blind spots — brightness hypersensitivity and texture loss — documented in DOI 1.20 and 1.20.1.
A plain-language explanation of 8-bit quantization and a floating-point trap called ghost duplication (DOI 1.16 and 1.21) — a structural property of any model trained on 8-bit images, not specific to this project.
The script that builds the no-join verification databases used from DOI 1.12 onward.
A beginner-friendly, section-by-section walkthrough of the same script: what each library does and why the code is written the way it is.
The reference implementation behind the eight comp_* diagnostic features documented in DOI 1.8.
The full experimental pipeline behind DOI 1.7's Axis5 collapse-intensity model and the Tier3 v13 definition.
The inference-only pipeline that produces axis5_full, the table behind DOI 1.7's Axis5/Tier3 results.
A walkthrough of the SQL patterns used repeatedly across the published verification queries.
Plain-language definitions of the symbols and variable names used throughout the technical reports and SQL.
An index of every verification SQL file published alongside the technical reports (DOI 1.0–1.21).
An index of old, replaced pipeline files, each pointing at its current replacement.
A higher-precision pipeline for measuring sigma_gt/mu_gt's distinct-value count, correcting for floating-point ghost duplication, validated across all four models.
A ground-truth-free binary classifier for SSIM Structural Masking risk (sigma_notta ≤ sigma_stat, sigma_notta measured exactly via comp_contrast_notta/255), validated at 98.8–99.5% accuracy across all four models.
A ground-truth-free binary classifier for SSIM Amplification Phase risk (mu_notta ≤ mu_stat), validated at 96–100% accuracy across all four models.
Closing an evidentiary gap in DOI 1.16 and DOI 1.18: proper video-recorded, four-model confirmation of two GT-only discreteness claims.
Completing the SSIM sub-term picture: the luminance-term identity, real ties across four models, and why all three terms can never coincidentally align at once.
Real nonzero sigma_gt=sigma_notta ties across four models, traced to 8-bit pixel quantization.
Testing the converse of the Pixel Identity Theorem, and proving why the contrast term C pins to 1.0.
Four-model re-derivation of mu_stat/sigma_stat plus the SSIM=1.0 pixel-identity corollary, all in SQL.
Two independent tests of whether an SSIM score tie also means the comp_* features agree.
sigma_stat correction: the from-scratch re-derivation queries for Standard and SwinIR.
The masking blind spot theorem, verified across all four models.
Correction of the C1/C2/C3 constant bug.
Live SQL verification of Failure Fingerprint values (High-Fidelity / Digital-Art).
Axis5 collapse intensity and Tier3 state distribution by Axis1 polarity.
Universal Tier3 classification coverage, with a DOI 1.0–1.6 research overview.
Viewer systems and a corrected collapse-type classification.
SQL corrections and two sigma_stat approaches.
