The license-family table aggregates the source table into five families and
claims the aggregation is complete. Nothing checked it: the prose sweep skips
partition subtotals on purpose, so a family could drift or go missing without
any check noticing.
Each released source is now assigned a family from the registry and the
subtotals are compared against the printed billions, along with the total and
the set of families itself — so a family present in the corpus but missing a
row fails, not just a wrong number.
Families need a finer split than license comparison does: norm_license folds
CC0 and public domain together because the source table prints either for the
same source, while this table reports them separately. license_family keeps
them apart and decides from the source's own license wording.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UXd7RSpYjF6SX2fFMygxic
Extends the paper verifier past names and token counts: every source row in
the composition table is now checked against src/sources.py and against the
source's own datasheet, and the table's footnote markers are checked in both
directions (used but unexplained, explained but on no row).
License strings are compared by (family, version), so the paper may abbreviate
"CC-BY-4.0" to "CC-BY" and print public domain as either PD or CC0 without
tripping the check, while a genuine family or version mismatch fails.
The audit's one real hit: parliamentary is public-domain content (PL art. 4
pr. aut.) redistributed by IPI PAN under CC-BY-4.0, so the registry's `license`
and `license_spdx` disagree by design. A one-line cell that prints only one of
the two layers now has to carry a footnote.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SLqJTEbW4oU78vmBMPVQs2
fix: make release totals registry-driven, not "whatever is on main"
make_docs.py counted every source with a stats.json, so a source merged to
main silently joined the "current release" and the README total row drifted
one release behind (4,246,429 docs / 9,597.9M tokens vs the actual v0.2.5
4,319,200 / 9,639.1M).
- src/sources.py: each entry records `release` (the version it was admitted
to, or None) and `added` (real date from its datasheet, replacing the single
global ADDED stamp). Reproduces 11/12/13/16/17/18 for v0.2.0..v0.2.5.
- src/make_docs.py: release totals count only sources whose `release` is set
and not newer than VERSION; the rest render as "on main, not yet released".
Also preserves existing CRLF line endings on regeneration, orders versions
numerically, and marks hand-authored datasheets `custom_datasheet: True` so
they are not clobbered by the template.
- README.md / CHANGELOG.md: regenerated totals, plus an Unreleased section for
wiktionary_examples, sejm_api and parlamint_pl.
- src/verify_paper_claims.py (+ test): chec…
- Add in-tree PII scrub and wire it into HPLT and Sejm fetchers (f02380ada54f72c34dcf195464c6639e840eb731)
Co-authored-by: Paweł Puzio <ppuzio@users.noreply.huggingface.co>
