Documentation
Every verification SQL file we've published, grouped by the
technical report it belongs to. A handful of the earliest reports
published their queries only inside the PDF and the
screen-recorded video, with no separate SQL file to link to — that's
noted where it applies, rather than left silent. For the recurring
patterns these queries use, see
How to read SQL; for what
symbols like sigma_gt actually mean, see the
Glossary. Where the original video
exists, each expanded entry also embeds the screen-recorded run
of that exact query.
View DOI 1.0 on Zenodo: https://doi.org/10.5281/zenodo.18830286
No standalone SQL file was published with this report — the queries behind the SSIM Limit Theory and Statistical Breakdown Fingerprints are shown only in the PDF and the two screen-recorded videos.
View DOI 1.1 on Zenodo: https://doi.org/10.5281/zenodo.18883267
No standalone SQL file was published with this report — the queries are shown only in the PDF and the four screen-recorded per-model videos.
View DOI 1.2 on Zenodo: https://doi.org/10.5281/zenodo.18917078
No standalone SQL file was published with this report — the queries are shown only in the PDF.
View DOI 1.3 on Zenodo: https://doi.org/10.5281/zenodo.18949243
No standalone SQL file was published with this report — the queries are shown only in the PDF and the four screen-recorded per-model videos.
View DOI 1.4 on Zenodo: https://doi.org/10.5281/zenodo.18994113
DOI1_4_sql.txt — The corrected queries themselves: dataset/model counts, Δssim spread by μgt bin, improvement/degradation rate by μgt bin, the noTTA→TTA L/C/S component shift for σgt=0 blocks, and the S-term spread by σgt bin — the raw material behind this report's two-stage μstat computation. Walkthrough →View DOI 1.5 on Zenodo: https://doi.org/10.5281/zenodo.19040901
DOI1_5_Silent_sql.txt — Classifies near-zero-Δssim blocks into Silent/Near-Silent (by how close Δssim is to exactly zero) crossed with None/Single/Multiple_or_Dual (by how many comp_* features dominate the change). Walkthrough →View DOI 1.6 on Zenodo: https://doi.org/10.5281/zenodo.19058174
DOI_1_6sql.txt — Three independent completeness checks: the overall Tier3 state distribution, a noTTA×TTA cross-tabulation, and a diagnostic query isolating the small number of NULL-state rows to show what data they're actually missing. Walkthrough →View DOI 1.7 on Zenodo: https://doi.org/10.5281/zenodo.21425511
axis5_weights_query.md — Confirms the published Axis5 feature-weight table matches what's actually persisted in each model's database, by reading back the stored config_json record (run once per model). Walkthrough →tier3_axis1_crosstab_query.md — Tier3 state distribution restricted to the positive-Axis1 region (Axis1 > +0.1), run once per model. Walkthrough →View DOI 1.8 on Zenodo: https://doi.org/10.5281/zenodo.21433447
This report documents the comp_* feature framework itself, implemented in Python rather than verified via SQL query. See import_ground_truth_data.py under Pipeline & Code.
View DOI 1.9 on Zenodo: https://doi.org/10.5281/zenodo.21433900
verification_queries.md — The dataset-scale and "Merutan condition" block-count queries run live on camera against the High-Fidelity and Digital-Art databases, annotated with explanatory notes. Walkthrough →source_script_en.md — The same script with no annotation added — the literal text pasted into the terminal in both videos, visible in each recording's paste-confirmation dialog. Walkthrough →View DOI 1.9.1 on Zenodo: https://doi.org/10.5281/zenodo.21445635
A visualization supplement to DOI 1.9 — no new queries. It plots DOI 1.9's already-verified values against the theoretical critical points established in DOI 1.0.
View DOI 1.10 on Zenodo: https://doi.org/10.5281/zenodo.21658621
verification_sql_final_en.md — The consolidated final verification script, already incorporating the corrected C1/C2/C3 constants and the σgt≤0.16 restriction from the start. Run once per model, four screen recordings total. Walkthrough →sigma016_verification_sql.md — The query used to export the four models' σgt≤0.16 evidentiary CSVs (Section 6.4 of the report). Walkthrough →create_alignment_db.py / create_alignment_db_single.py — The database-builder scripts used to regenerate this report's four model databases with the corrected C1/C2/C3 constants (the latter a single-process variant for WSL memory limits). Later superseded by the JOIN-free pipeline — see Superseded Files.sigma016_check_std.csv / sigma016_check_swinir.csv / sigma016_check_high.csv / sigma016_check_digital.csv — The four models' σgt≤0.16 evidentiary CSVs, actually exported by the query above.View DOI 1.11 on Zenodo: https://doi.org/10.5281/zenodo.21837260
query_doi111_verification_windows.sql — Verifies the σgt=0 ⇒ S=1 identity against all four models in a single session (all four databases ATTACHed together, results combined with UNION ALL), confirming zero exceptions. Walkthrough →View DOI 1.12 on Zenodo: https://doi.org/10.5281/zenodo.21894585
standard_mu_stat_exact_match_recheck.sql — Re-derives Standard's μstat from scratch against the from-scratch, no-join database, confirming it independently of the retired is_comp_zero flag. Walkthrough →standard_sigma_stat_exact_match_recheck.sql — Re-derives Standard's σstat the same way — the query that found 0.101 instead of the originally published 0.0800. Walkthrough →swinir_mu_stat_exact_match_recheck.sql — The same μstat re-derivation for SwinIR, confirming its value unaffected. Walkthrough →swinir_sigma_stat_exact_match_recheck.sql — The same σstat re-derivation for SwinIR, confirming its value unaffected. Walkthrough →View DOI 1.12.1 on Zenodo: https://doi.org/10.5281/zenodo.21903024
A supplement to DOI 1.12: adds the SQL behind Section 4.4's Pixel Identity Theorem cross-check (previously omitted from that report's bundle) and extends it to SwinIR, matching DOI 1.12's own two-model scope. Standard: 702,108 blocks, zero exceptions. SwinIR: 2,657,100 blocks, zero exceptions.
View DOI 1.13 on Zenodo: https://doi.org/10.5281/zenodo.21960019
cross_model_notta_exact_match_recheck.sql (+ 6 single-pair variants) — Re-checks whether two different models' exactly-matching SSIM scores, at the same coordinate, also imply matching comp_* features — across all 6 model pairs. The single-pair files split this into one file per pair so each can be run and screen-recorded independently, without ATTACHing all four databases at once. Walkthrough →cross_model_notta_exact_match_sigma_breakdown.sql (+ 6 single-pair variants, + a _sequential all-pairs-in-one-file version) — Mechanism follow-up: breaks the exact-match population down by σgt to test whether the masking blind-spot theorem (DOI 1.11) explains the jump in match rate once the old filter is removed. Walkthrough →notta_tta_same_score_same_content_check.sql (+ a _4model combined version) — For the same model at the same coordinate, checks whether an identical noTTA/TTA SSIM score also means identical comp_* content. Walkthrough →View DOI 1.14 on Zenodo: https://doi.org/10.5281/zenodo.22012409
verification_sql_final_en.md — A consolidated verification script extending DOI 1.12's to all four models, computing both μstat and σstat directly in SQL (no external Python post-processing step), and adding a Part 5 for the SSIM=1.0 pixel-identity transitivity check. Walkthrough →View DOI 1.15 on Zenodo: https://doi.org/10.5281/zenodo.22013446
standard_gt_feature_match_ssim_correlation_check.sql — Asks the converse of the Pixel Identity Theorem: if a reconstruction's 8 comp_* features match ground truth's exactly, does that force SSIM=1.0, or can it still fall far short?standard_gt_all8_match_ssim_distribution.sql — Follow-up: for the group whose 8 features all match, how are the remaining (non-1.0) SSIM values actually distributed?gt_feature_match_sigma_gt_mechanism_check.sql — Follow-up: tests whether the masking blind spot (σgt=0, DOI 1.11) explains the dense SSIM plateau found in that group.gt_feature_match_lc_breakdown_check.sql — Breaks the SSIM product apart into its L and C terms individually, within the σgt=0 subgroup, to see which term actually drives the plateau's spread.gt_feature_match_lc_breakdown_both_sigma_check.sql — Generalizes the C-term check beyond σgt=0: tests whether C=1 whenever σgt=σnotta, for any shared value, not just zero — the theorem this report is named for.verification_sql_final_4model.sql — The consolidated script running the full chain above against all four models in one file. Walkthrough →View DOI 1.16 on Zenodo: https://doi.org/10.5281/zenodo.22029346
nonzero_sigma_gt_eq_notta_discovery.sql — Searches the full database (not restricted to the comp_*-match subgroup) for concrete, nonzero cases where σgt=σnotta, using the SSIM contrast term's own signature (the schema has no direct σnotta column). Walkthrough →sigma_gt_discreteness_check.sql — Tests whether σgt is drawn from a small, heavily-repeated set of discrete values (consistent with 8-bit pixel quantization) — the mechanism behind the previous file's matches. Ground-truth-side only, so the result is shared across all four models. Walkthrough →verification_sql_final_4model.sql — The consolidated script running both checks above against all four models in one file. Walkthrough →View DOI 1.18 on Zenodo: https://doi.org/10.5281/zenodo.22230202
nonzero_mu_gt_eq_notta_discovery.sql — The luminance-term analogue of DOI 1.16's discovery script: searches the full database for concrete, nonzero cases where μgt=μnotta, using the SSIM luminance term's own signature (the schema has no direct μnotta column). Walkthrough →mu_gt_discreteness_check.sql — Tests whether μgt is drawn from a small, heavily-repeated set of discrete values, the same 8-bit quantization mechanism as DOI 1.16, checked here for μgt instead of σgt. Walkthrough →verification_sql_final_4model.sql — The consolidated script running both checks above against all four models in one file. Walkthrough →joint_term_identity_check_4model.sql — Tests whether SSIM's three sub-terms (L, C, S) ever align to exactly 1 simultaneously for a block that isn't pixel-identical, across all four models. Walkthrough →joint_term_identity_swinir_exception_diagnostic.sql — Inspects the only cases found where the product equaled 1.0 without all three terms individually bit-exact, confirming they were floating-point rounding, not a real exception. Walkthrough →joint_term_identity_check.sql — The original single-database version of the joint-term check, included for completeness. Walkthrough →View DOI 1.19 on Zenodo: https://doi.org/10.5281/zenodo.22252098
sigma_gt_discreteness_check_4model.sql — DOI 1.16's σgt discreteness query, extended to run against all four models in one file with screen-recorded video evidence (DOI 1.16's own text only stated the GT-only-dependency argument, with no cross-model check). See DOI 1.16 →mu_gt_discreteness_check_4model.sql — The same treatment for DOI 1.18's μgt discreteness query (previously checked against only two of the four models). See DOI 1.18 →View DOI 1.20 on Zenodo: https://doi.org/10.5281/zenodo.22968572
verification_sql_final_4model_comp.sql — The consolidated, copy-paste-ready binary classifier query (mu_notta ≤ mu_stat, mu_notta recovered algebraically), covering all four models sequentially. Walkthrough →curiosity_binary_classification_validation_comp_standard.sql, _swinir.sql, _high.sql, _digital.sql — The original, per-model exploratory scripts the consolidated file was assembled from, included for completeness. Walkthrough →View DOI 1.20.1 on Zenodo: https://doi.org/10.5281/zenodo.23023733
verification_sql_final_4model_comp.sql — The consolidated, copy-paste-ready sanity check, correlation, and binary classifier queries (sigma_notta ≤ sigma_stat, sigma_notta measured exactly via comp_contrast_notta/255), covering all four models sequentially. Walkthrough →View DOI 1.21 on Zenodo: https://doi.org/10.5281/zenodo.23271571
ghost_duplication_concrete_example_4model.sql — Shows one concrete pair of blocks sharing the exact same true sigma_gt, whose stored bit patterns nonetheless differ by exactly 2^-64. Walkthrough →DOI1.21_verify_against_published_comp_db_4model.sql — Recomputes sigma_gt/mu_gt's true distinct-value count against the exact published database and population, alongside the originally published (inflated) figures. Walkthrough →DOI1.21_section8_noise_signal_gap_4model.sql — Validates the 1e-9 tolerance threshold by measuring the empty band between floating-point noise and genuinely distinct values. Walkthrough →proxy_c_eq_1_robustness_check_4model.sql, comp_vs_sigma_db_naive_count_diff_4model.sql, sigma_mu_notta_real_correlation_check.sql, sigma_mu_notta_real_correlation_check_v2_tolerance.sql, DOI1.21_correlation_ratio_4model.sql, DOI1.21_section4_validity_checks_standard.sql, sigma_db_content_hash_4model.sql — Supporting checks (robustness, population reconciliation, correlation-ratio recalculation, threshold validity, database content hashing) cited in the full report but not walked through individually here. See the report's own Section 9 for the complete list.