kami
c19a726455
Answering whether the caption merge breaks the page-space key in eval/chapter-truth.json. It does not. _merge_plan only groups adjacent fragments and slice_webtoon cuts at gutter midpoints, so a vstacked panel's [x, y0, w, sum(heights)] is the exact page region. The defect is one line up. slice_webtoon skipped any segment shorter than min_seg instead of absorbing it, so this chapter's 116 panels have a 62px hole between panel_order 70 and 71 and lose 18px at the strip's end. A merge spanning that hole reports a bbox 62px short of the page span it covers, and every consumer mapping a detection back to the page is then wrong by that much, silently. The 62px held the credits line, found at p_050.png row 338. A short band is now held and absorbed by the next segment, or by the previous one when it is last. min_seg still suppresses a tiny panel and no strip pixel is discarded. The self-check asserts the bboxes tile the strip end to end. Checks: worker_crop.py self-check passes and fails on the old code, where the first band starts at 49 instead of 0. ruff check . exits 0. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
4.5 KiB
The measurement spine
eval/chapter-truth.json plus the purity and fragmentation section of audit_registry.py. This is
ARCHITECTURE.md step 0, and nothing below it in that build order is measurable without it.
The truth file keys on page-space geometry
Closed, 2026-08-13. Reproduces the baseline on the 19:44 run of 2026-08-12.
A truth row cannot name a panel_id or a character_id. A re-crop remints panel ids, and
/characters/reset remints character ids. The fifth cycle does both, so a truth file keyed on either
would die the moment it was first used.
Each row carries [x1, y1, x2, y2] in page space, which is the panel's bbox offset plus the
detection box. The page strip is the fetched artifact and does not change when panels are re-sliced, so
the key survives. audit_registry.py:_score matches a row to an assignment on the same page at
IOU_MIN = 0.5, greedily, best overlap first, each row and each assignment used once.
The 19:44 panel_order, local_id and character_id are kept per row under from_19_44, for
debugging only. Nothing matches on them.
A slice under min_seg is absorbed, not dropped
Closed, 2026-08-13. worker_crop.slice_webtoon, checked in its __main__.
The page-space key above needs the panel bboxes to tile the strip. They did not. slice_webtoon skipped
any segment shorter than min_seg, so its pixels reached nothing and the y chain gained a hole. On the
2026-08-12 chapter it happened twice. 62px went missing between panel_order 70 and 71, holding the
credits line 글 | H2J · 그림 | 이루리다, and 18px at the bottom of the strip.
62px is small and that content was worthless, which is why it went unnoticed for the whole project. The
rule is not harmless. The same threshold drops a 62px caption. merge_faceless_captions vstacks
adjacent fragments, so a merge spanning a hole reports a bbox 62px shorter than the page span it covers.
Every consumer mapping a detection back to the page is then wrong by that much, silently.
A short band is now held and absorbed by the next segment. The last band is absorbed by the previous one.
min_seg still does its job, which is to suppress a tiny panel, and no strip pixel is discarded.
The self-check asserts the bboxes tile the strip end to end and that no panel is under min_seg.
Purity is a share, so it never names an id
Closed, 2026-08-13.
Purity of a cluster is the largest share of one true person inside it. The lead's 16 assignments hold 14 of him, so 0.88, and the truth file never has to say which id he was. Fragmentation of a person is the number of clusters holding their occurrences. Woman A has 15 occurrences over 2 ids, so 2.
So both numbers survive a reset. The definitions in ARCHITECTURE.md were not explicit about that.
A cluster is reported under its dominant person. Every other person in it prints as a wrong: line,
which is where the photograph and the chibi show up on the lead.
The baseline block in the truth file holds the 19:44 numbers, and each printed line says
= baseline or the delta. That makes step 0's DoD a check rather than an eyeball.
The truth is 38 assignments, and holds no dialogue yet
Open, 2026-08-13. The character half is done, the dialogue half is not written.
The eyeball pass covered three characters, 38 of the run's 60 assignments. The other 10 characters were never checked and carry no truth row. The audit prints them as out of scope rather than as correct. Anything claiming chapter-wide accuracy off this file is overreaching.
The prose in NEXT.md left two rows unnamed. They were identified on 2026-08-13 by re-cropping p030
and p032: panel_order 31 and 33 on woman B's id are both woman A. That closes the recorded
arithmetic of 7 correct and 2 wrong, and it is why woman A's fragmentation is 2.
One judgement is recorded in the file rather than hidden. The chibi of woman B at panel_order 28
counts as her, because it is a stylised rendering of the subject. The chibi of the lead at
panel_order 20 does not, because it is a drawing inside the scene. not_people names the second kind.
ARCHITECTURE.md also asks for 30 dialogue lines with their true speaker, typed
visible | offscreen | narrator | unknown. dialogue is an empty array. No such pass exists, because
NEXT.md records no per-line speaker truth. Writing it means watching the chapter panel by panel. The
audit prints nothing about speakers until it is filled.