comicbox.formats.base.online.signals¶

• normalize_series — Lowercase, strip volume suffixes, collapse non-alphanumeric runs.

• s_series — Series-name similarity via rapidfuzz WRatio, normalized.

• s_issue — Issue number match. Integer-equal first; fall back to string-equal.

• s_year — Year match with smooth decay beyond the original 1.0/0.7/0.4 anchors.

• s_publisher — Score publisher equality after normalization.

• s_pages — Page-count match: 1 if equal, 0.7 within 10%, 0.3 within 25%, else 0.

• both missing: 0.5 (weak agnostic prior — we have no signal)
• asymmetric (one side has year, other doesn’t): 0.3 (the previous 0.6 over-credited unknown candidates and let wrong-volume picks coast through the auto-write band when CV’s BasicIssue lacked a cover_date)

Phase F (2026-05-14) replaced the original cliff (0.0 for any diff ≥ 3) with linear decay. The cliff treated modern reissues of older series as having no year signal at all — fine for the common case where year discriminates between volumes, but bad when the right candidate IS the same series at a different year (e.g. The Boys 2009 fixture’s expected vol was at year=2006, gap of 3 years). Smooth decay gives the older candidate ~75% of the one-year-off signal, enough to compete with cover-hash signal when present. Beyond 7 years, the score still hits zero — the long tail of 20+ year reissues (Conan 1973 vs 2025) shouldn’t benefit from year signal.

See tasks/online-tagging/calibration-notes/2026-05-14-bigmedia-247.md for the failure-mode analysis that motivated this change.