An index of files from our own published DOI archives that have since been replaced. We keep documenting them, clearly marked as superseded, rather than deleting them from view — so that if you ever come across one of these files on its own (say, from an older DOI's Zenodo archive), you'll know it's outdated and where to find the current version, instead of unknowingly reproducing a result that's since been corrected.
create_alignment_db_single.py| Old file(s) | create_alignment_db_single.py, and an earlier multi-process variant, create_alignment_db.py |
|---|---|
| Published in | The DOI 1.10 archive (Merutan_DOI1.10.zip) |
| Superseded by | build_notta_tta_comp_db.py |
Computed SSIM for every block and wrote it to a database. The
comp_* features and
delta_ssim were pulled in afterward from a
separate, external source database
(meltan_integrated_viewer.duckdb) by matching
image/row/column coordinates — the join step described on
the Technology page. That
source database's schema has been public from the very start:
DOI 1.0's
own appendix documents its blocks,
meta, and axis5_history tables in
full, column by column (Appendix A.2, "Integrated Viewer DB").
The raw database file itself is large enough that it wasn't
distributed directly — DOI 1.0 says so explicitly
(Appendix A.3.3) — but that's the same practical,
file-size reason later verification databases in this series
aren't distributed either, not a matter of anything being kept
quiet. What isn't separately published is the multi-step Python
pipeline that builds that database (the block-level system
referenced in the origin story on About);
two of its scripts —
sorted_phase_v13_full.py
and
generate_axis5_tier3_full_db.py
— were released alongside DOI 1.7 and are documented under
Pipeline & Code. This patchwork of separate scripts and
databases is a mark of the project's early, exploratory period,
where tools accumulated one need at a time rather than being
designed as a single system from the start. The single-pass
rewrite below is what that experience led to.
That join step was a repeated source of silent data-integrity
bugs at the scale of hundreds of millions of rows — rows that
stayed at zero when they shouldn't have, or received another
block's values. Two concrete, published corrections trace back
to it: Upscayl Standard's
sigma_stat
(published as 0.0800, corrected to 0.101 in
DOI 1.12),
and Upscayl High-Fidelity's
mu_stat/sigma_stat
(corrected in DOI 1.14).
build_notta_tta_comp_db.py — computes SSIM and all comp_* features together, in a single pass per image, with no cross-table join anywhere in the process.
This index will grow as more superseded files are identified and documented, each getting its own section like the one above. If a specific correction interests you, the technical report tied to it (linked above) always has the full mechanism, affected values, and evidence.