Canvas-padding prevalence: majority of catalogue (~55-65%), not a targeted fix (measurement, follow-up to #735/#736)

[WilfordGrimley]

Summary

#735 established that _BORDER_SAMPLE_BANDS lands on canvas padding rather than the card's
printed border in a sample of directly-viewed images, and that layout_class agreement with
ground truth (67.2%) is worse than always guessing "black" (84.8% base rate). Neither #735 nor
#736 established what share of the catalogue actually has canvas padding — without a
denominator, "close to universal" in an 8-image sample can't be distinguished from a defect
confined to a few large sources. This issue reports that denominator.

Full report:
docs/reports/2026-08-06-canvas-padding-prevalence.md

Headline

Roughly 55-65% of the catalogue's 230,373 cards show genuine canvas padding — a
precision-corrected estimate. The raw automated detector read 86.5% (315/364 stratified sample);
eye-validation on 43 images found the detector's padded calls have real, non-uniform precision:
87.5% correct on black-bordered cards, but only 33.3% correct on borderless cards (n=16 and
n=9 validated respectively), because full-bleed/borderless art and ornate/holo/gradient card
borders get mistaken for a canvas edge by any uniformity-based method. Applying those measured
precision rates to their own population strata gives the corrected 55-65% range. Either the raw
or the corrected number is a majority of the catalogue, not a handful of sources.

New information beyond #735/#736

1. A denominator. 230,373-card catalogue, source-stratified sample (top-20 sources by volume
   at ~15 each, 60 more from the long tail, plus the 4 cards classify_border_color's fixed-fraction bands sample canvas padding, not the card's border, on most catalogue images #735 named), 364 images total.
2. Per-layout-class reliability split. Canvas detection is validated reliable on
   black-bordered cards (87.5% precision); it is not reliable on borderless cards (33.3%
   precision) — this is a different failure surface than the "black-on-black is invisible"
   concern classify_border_color's fixed-fraction bands sample canvas padding, not the card's border, on most catalogue images #735/the landmark-gate report predicted (0/40 abstentions in this method were driven
   by a black-indeterminate edge; 39/40 were on borderless-layout cards instead). Full
   explanation in the report's "black-on-black bias, reconsidered" section.
3. Borderless-on-canvas has its own number: ~8-9% of the catalogue (corrected), the
   population no positional (non-colour) fix can ever fully rescue, since a borderless card has
   no printed frame strip to locate its edge by.
4. Per-source concentration table for the top-20 sources by volume (66.1% of the catalogue),
   with each source's black/borderless mix so raw padding rates can be judged for reliability.
   The three largest sources (RustyShackleford, Berndt_Toast83, PsilosX) read ~94-100% padded and
   are majority black-bordered, so those specific figures are close to face value.
5. Vertical padding asymmetry confirmed at n=215 (vs. the prior report's n=4): top padding
   median 5.95%, bottom median 9.41%, median gap 3.19 percentage points, 64% of images >2pp
   asymmetric. Horizontal padding is symmetric (median gap 0.00pp). Any future correction cannot
   assume symmetric vertical padding.

Recommendation

Even the conservative corrected estimate (~55%) means this is not a targeted fix for a handful
of sources — it is a catalogue-wide validity problem for every fixed-fraction geometry assumption
downstream of image edge == card edge (classify_bleed_edge, compute_bleed_diff_mm,
_BORDER_SAMPLE_BANDS, *_CROP_BOX). The borderless-on-canvas remainder (~8-9% of the
catalogue) needs a separate, non-positional approach and should likely be scoped explicitly out
of a first fix rather than left as an implicit gap.

Measurement only — no calculator changed, no vote cast, no code touched. Full methodology,
per-source table, and open items in the linked report.

[WilfordGrimley]

Correction from the follow-up validation (2026-08-06-borderless-validation.md, docs/reports/ in proxyprints-orchestration): the borderless padded precision figure in this issue (33.3%, n=9) is superseded by a 51-card, 26-source stratified validation measured at 91.9% (34/37).

What changed (OPEN ITEM #1 in this issue predicted exactly this):

• Borderless padded precision: 33.3% (3/9) → 91.9% (34/37) — the n=9 figure was a small-sample artifact (unlucky draw of ornate-frame cards).
• Corrected borderless padded rate: ~22% → ~60.7% (66.0% raw × 0.919).
• Borderless-on-canvas floor: ~8-9% → ~20.4% of the catalogue (~47,000 cards) — roughly 2.4x larger than reported here.
• Total corrected padding prevalence: ~55-65% → ~79-81% (black stratum unchanged at 87.5%).

The 3 misses (all ornate/decorative-frame borderless, all clip the top edge): Kaito Shizuki (top+left), Avacyn Angel of Hope (top), Sliver Overlord (top) — confirmed by pixel-level arbitration of zoomed edge strips. Clean (non-ornate) borderless validated 31/31 (100%). Sliver Overlord reproduces as a miss from this issue's named list; the other named cards (Chaos Wand, Wooded Foothills) validate CORRECT on different instances.

Net effect on conclusions: "majority of the catalogue" holds a fortiori (~79-81%); the borderless-on-canvas remainder is a first-class scoping problem (~47k cards), not a residual. Measurement only — no extractor or calculator change in either report.