Measured with the image check: full-body turnarounds 0/5 -> 10/10 (Karyla),
12/12 on Karyla, Darrag and Eolan after the face fixes. The 4-step model
loses the tail of a ~3000-char prompt. Scars are left out of image prompts:
any scar wording, even "fully healed", rendered as fresh gashes. v1.40.0.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Bulk delete/type for a click- or rectangle-selected set; every destructive
region change returns an undo token (one level, refused once something newer
happened); text inside a region that becomes a picture is folded into it.
TODO: model-based layout detection (PP-DocLayout, LayoutReader) with licence
notes. v1.39.0.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Split at the cursor and merge-with-next keep article membership right; typed
letters in the editor never fire shortcuts; focus returns to the same row
after a redraw. v1.38.0.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Rules sized on a real 204-page issue before writing them; each re-types
(not read aloud, reversible), none deletes. A run-on author line's sentence
becomes its own paragraph instead of being glued to "the next paragraph",
which was right in only 1 of 5 real cases. Runs on import and re-OCR, and on
demand with a layout backup. v1.37.0.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A hand-ordered page is locked: Fix reading order, type changes and newly drawn
boxes leave it alone until Automatic order. v1.36.0.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Columns are found from where column-width text starts, not from the whitespace
between them (loose OCR boxes shrank one gutter to 4-6 pt and a pull-quote
bridged it, collapsing page 38 into one interleaved column). Headlines get
their own band, pull-quotes are read at the end of the page. Measured over a
204-page issue: 89 sentence continuations kept vs 85.
Each review row gets a type picker (text kinds, page number, source, ad,
image) and its own delete button; text<->image conversions keep the region
in its article. Fix reading order re-reads existing issues without OCR and
leaves hand-ordered pages alone. The shared confirm dialog is now an
alertdialog for screen readers. v1.35.0.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The job checks the portrait crop, then the turnaround, story panel and
same-person against the sheet, re-renders on failure (up to 3 attempts) and
keeps the attempt with the fewest faults. ReActor transfers the profile's face
onto the turnaround and the largest face in the story panel. If the checker is
down, retries stop and the pair is marked unverified. The outcome is saved
with the pair and shown on the character page.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The model only reports observations; expectations come from the sheet and the
comparison is in code. Fail-safe: an unreachable checker or unreadable answer
is a failure. Calibrated 14/14 on panels with known defects using the already
loaded LLM; Ling 3.0 Flash VL (free) scored 8/14 with 6 unreadable answers.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A cold-booting model server would otherwise time out the first passage into tag guessing.
Documents that the custom attribution prompt applies to the classic engine only.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Pooled over three hand-corrected novels (~54k spoken words): agreed 82% right, disputed 48%,
Unknown ~0%, unverified ~76%. The old heuristic bands claimed 16% for lines that were 62% right
and tag/role signals did not discriminate in span casts; classic casts keep the old model.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Verify / recast now run span-natively (no text re-emission): a second differently-windowed
pass marks lines agreed/disputed and records the other answer, without changing any speaker
(~9 min vs ~49 min). Measured on the real pipeline, overruling disputed lines with a reasoning
adjudication made them worse (66.5% -> 61.4%), so that step was removed; agreed lines were 83%
right vs 48.5% for disputed ones, which is the signal used for confidence and review.
Auto-repair no longer runs after a span cast; segments carry a spanId.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
- fix: base-appearance matched pair lost its concept image + metadata on reload; visual save
wrote base concept art to the record root; one save path for base and forms now
- fix: extraction no longer lets the LLM author image_prompt (source of avatar/concept drift);
avatar-only generation uses the deterministic prompt and flags an existing pair
- fix: prompt writes and Physical/Clothing edits go to the active form; review endpoint checks
each form's own pair; fingerprint parity with the browser incl. visual_* overrides
- feat: appearance forms panel (add/rename/delete/show, per-form stale flag, generate one/all)
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Split the book into narration/dialogue spans from its own quotation marks and ask
the LLM only who speaks each dialogue span (POST /api/attribute-spans), so text can
no longer be fused, dropped or duplicated and a novel takes ~10 min instead of ~4 h.
Book 1 vs the hand-corrected reference: every-word-right 87.0% -> 92.3%, type
agreement 93.6% -> 97.4%; Books 2/3 (not tuned on) 90.8% / 89.9%.
- static/js/audiobook-spans.js (pure, unit-tested) + audiobook-cast-spans.js
- core/attribution_spans.py, route, tests; scripts/score_attribution.py
- default on; localStorage ab-span-mode=0 restores the classic engine
- architecture/audiobook-attribution.md, Manual, CHANGELOG 1.32.0, TODO
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Connect and "Use as LLM" on the LiteLLM Gateway card now resolve that
card's own saved key (engine_api_keys[name]) server-side in
routes/docker.py, ahead of the old global llm_api_key fallback. The
browser can no longer reliably forward a saved key at all -- it's
masked to "__configured__" in every /api/settings response the moment
it's saved -- so relying on a frontend-forwarded header alone left
Connect silently using a stale global key after LiteLLM's own key was
rotated, surfacing a misleading /api/version 404 instead of the real
401.
"Use as LLM" now goes through a dedicated /use-as-llm endpoint that
saves the exact credential Connect just verified (llm_url normalized
to .../v1, llm_api_key copied from the card), rather than a
separately-typed value that could drift again.
Also documents per-user provider-key storage as an explicit item for
the deferred multi-user/identity planning pass.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Confirmed live: re-ingesting c't KI Wissen 2026 with a local Qwen VL
engine, page 18's chart transcription degenerated into "das Produkt,
das Produkt, das Produkt, ..." repeated hundreds of times, filling the
whole response budget -- and was accepted as real region text, which
would have been narrated verbatim.
The codebase already had two narrow signatures for specific observed
failures (a formatting-instruction echo, a fabricated LaTeX loop), both
scoped to the plain-text fallback path only. Added a general detector
(_looks_like_repetition_loop): any short phrase containing a real word,
repeated 8+ times immediately back-to-back. Deliberately excludes
patterns with no letters -- a dot-leader contents line, a repeated
number sequence -- which are legitimate and must not be flagged.
Applied at both call sites: vision_ocr_page (the labelled-region path
most engines use) and _whole_page_region (the plain-text fallback).
Treated as a failed call in both, so vision_ocr_page_chain's existing
fallback (secondary engine, then tesseract) gets a real chance instead
of garbage being accepted as the page's transcription.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Real failures found by re-inspecting a fresh re-ingest of c't KI Wissen
2026 after the previous figure-boundary-snapping release: a bar chart
was still truncated to roughly a third of its real height, and a
chart's top edge cut into an unrelated paragraph above it.
Root causes, each independently confirmed against the real page render
before fixing:
- A bar chart's bars/gridlines/axis labels rarely reach the "solid
ink" density a photo does (measured 0.25-0.5 vs the 0.55 threshold),
so the existing pass anchored only on the one fully-solid element
(the zero-line) and never grew further. Added a second, looser pass
that grows an anchor -- or the seed itself -- through any real mark
at all, fenced by the same blocked cells.
- Anchor selection ("which solid component does the seed overlap
most") was seed-size-dependent: a small seed touching only the
zero-line anchored there, but re-running the SAME function on the
box it had just produced -- now large enough to also touch a dense
axis-label strip -- picked that instead and shrank an already-good
box back down. Fixed by dilating the solid mask and labelling once,
BEFORE any seed is considered, so which components count as one
figure is a property of the page alone.
- A merged bounding box can enclose real text sitting between two
components even though neither component's own cells touched it (a
rectangle union, not the components' actual shape). Added
_trim_blocked_edges() to pull each edge back to the nearest text
wall found searching outward from the box's own centre.
- Two adjacent figures fencing each other's growth made each one's
result depend on the other's CURRENT box -- for two dark photos
sharing a caption strip, this never settled, swinging by over 100pt
call after call. snap_visual_bboxes() now re-seeds every visual on a
page from every other's last result for a bounded number of rounds,
deciding same-graphic ambiguity once up front (fenced against the
original seeds, before any iteration) so two disputed seeds can't
mutually fence each other into a tidy-looking split instead of being
recognised as the same disputed claim.
- A result is now only accepted once it reproduces itself when fed
back through the same rules (IoU-checked), and the minimum
seed/graphic overlap for repositioning a badly-offset box was
lowered slightly -- it was narrowly rejecting some real cases.
Verified against the full 204-page/147-figure document: idempotency
holds except for one confirmed non-convergent pair, which now lands on
the same deterministic answer every run instead of continuing to
drift.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
A vision OCR engine grounds a figure far less precisely than it
transcribes one. Over all 138 figure regions of a 204-page issue the
transcribed contents were right while the boxes were routinely offset,
undersized, or sitting on the neighbouring text column -- page 23's card
grid was reported at x 0.24-0.54 where the graphic runs 0.41-0.95.
snap_visual_bboxes() reduces the page render to an ink-coverage grid,
seeds a connected-component search with the engine's box, and re-joins
tiles a gutter apart so a card grid comes back as one figure. Text
blocks are excluded from the mask, which is what stops a figure
swallowing the column beside it. A candidate must be solidly inked or
framed on nearly all of its rows and columns (admitting a white-interior
screenshot), may not shrink below 30% of the seed, and reverts when two
figures land on the same graphic; anything else keeps the box it had.
Runs before the figure crop in both OCR paths, so the images shown
during playback are framed correctly too. Replaces refine_picture_
bounds(), which only recognised coloured panels and had no text fence.
Also adds "Fix figure boundaries" in review and POST
/api/magazine/{id}/figures/refine behind it, so an issue ingested before
this pass can be corrected without re-running OCR. Verified live against
c't KI Wissen 2026: 78 of 138 boxes corrected, the rest declined.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Three research passes into what multi-user hosting would require found the
app has zero multi-tenancy today: one global settings/database/voice
library, no browser-UI authentication at all, global (not per-user)
concurrency limits, and no consent/audit mechanism for voice cloning.
The user wants a single shared, interconnected login/profile system across
all their self-hosted apps rather than a per-app login, extending the
GoTrue+Postgres+PostgREST stack OurFlashcards already runs on the NUC — so
this is recorded as a deferred, cross-repo initiative rather than
implemented now. One policy decision was made and is captured for future
implementation: voice cloning will require a consent attestation with an
audit trail once accounts exist.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Split-screen review: the rendered page (left) with a numbered box per
detected block, the extracted text (right) with matching numbers.
Clicking either side highlights the other. Checkboxes moved from a flat
list into this layout without changing their semantics (still gate
what gets narrated).
"OCR again": reocr_page() forces one page through the configured
vision engine again, replacing only that page's blocks -- deliberately
does NOT call build_articles(), so an already-approved article is not
silently re-grouped. A magazine-level option re-reads every page and
re-runs detection (with a confirmation, since that does reset article
status). extract_pages() gained force_vision_ocr to skip its quality
gate for this case.
Building the numbered-box review immediately did what it was built to
do -- surfaced a real bug on live production data (real magazine, real
scan, not a synthetic fixture): a photo caption at 49.9% of the page
width sat in the gutter between two genuinely separate columns and
merged them into one, so every block on the page came back column=0
and reading order fell back to pure top-to-bottom, interleaving the
two columns' sentences. Fixed two compounding causes: column boundary
detection now excludes blocks much wider than the page's median width,
and detect_columns' gutter threshold (3.5%) was itself narrower than
some real magazine gutters (~2.1% measured) -- lowered to 1.5%.
10 new/updated tests, including one built directly from the real
measured page-17 block coordinates that exposed the bug.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Magazine P1:
- Images shown automatically during playback (assign_images_to_articles
links each figure to the article whose column it falls in; the client
shows the last image reached in reading order as playback advances).
- "Play all articles" queues every approved article and advances
automatically at the end of each rather than stopping.
- Word-level highlighting, explicitly documented as an ESTIMATE (no
forced-alignment/timestamp source exists): each word's share of the
sentence's real measured audio duration is proportional to its
character length. A checkbox turns it off; sentence highlighting
(measured, not estimated) is unaffected.
- Driving mode: full-screen overlay, two oversized controls, reuses
whatever is already playing/queued.
- Sidebars/pull-out boxes excluded from narration automatically
(mark_sidebars: an anomalously narrow column vs. the page's regular
grid). Fixed a real defect found while building this: a full-width
headline was left in the column-width analysis and could bridge two
genuinely separate columns into one, hiding a real sidebar column
entirely -- caught by a failing test before the fix.
Second OCR engine + fallback ladder:
- KDL Frontier Parser nano wired in via vision_ocr_page_chain (ordered
engine list, first usable result wins, tesseract last resort).
Confirmed live: ~3x faster than Baidu (~12s vs ~40s/page) but returns
plain text with NO coordinate markers, unlike Baidu's labeled regions.
_whole_page_region treats the whole response as one ambiguous
full-page block (gated by ocr_quality_score) rather than discarding
usable text just because it lacks layout structure.
OCR engines page + a real UI bug:
- The OCR & Document Parsing page had no .s-engines-page[data-page="ocr"]
markup at all -- renderLocalContainers already supported an OCR grid
generically, but $('dc-grid-ocr') was always null, so the card grid
silently never rendered. Added an Active OCR Engine panel mirroring
Active Language Model exactly, plus the Local grid -- both configured
engines now appear as ready-made cards, no "Add Custom" dialog needed.
- Fixed the audiobook pipeline's "Previous/Next" nav bar bleeding into
every .section-head in the document (confirmed live on the Engines
page: a "Cast Audiobook" button and "STEP 1/8" hint that made no sense
there). Changed from a blocklist (naming each unrelated section as
noticed) to an allowlist of the pipeline's three actual destinations.
Verified live end-to-end: play-all queue building and manual advance,
word-span markup, driving-mode toggle, OCR page showing both engines as
local cards, zero stray nav bars on the Engines page, no console errors.
20 new backend tests (156 total in the magazine module).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
A page needing OCR (or whose already-extracted text -- native or
tesseract alike -- scores as mostly garbage via ocr_quality_score) is
now escalated to a connected vision OCR engine before falling back to
tesseract. Falls back automatically on any failure (unreachable engine,
timeout, unparseable output); never fatal to ingestion.
Verified live against ocr-baidu-unlimited-ocr on a real 28-page scanned
c't magazine excerpt (Epson ES-580W, no scanner OCR): every garbled
article title tesseract produced ("K | soll Arbeit abnehmen", ". |",
"@)X", "Owns mee") was replaced by the real headline ("1. Irrtum: KI
ist ein allgemeiner Jobkiller", etc.), and furniture detection jumped
from 44 to 151 blocks because the engine labels running headers/
footers/page numbers directly instead of needing fuzzy post-hoc
matching. 26 of 28 pages used the vision path; the remaining 2 fell
back to tesseract automatically.
Also fixes two real defects found investigating the same real-world
data, both regression-tested against the actual garbled strings
observed:
- Exact-match repeated-line detection (running header/footer) silently
failed once tesseract OCR'd the same line differently each time --
cluster_running_furniture fuzzy-matches near-edge blocks instead.
- ocr_page() built one Block per OCR text-LINE rather than per
paragraph, so a single anomalous line (a chart number, scan noise, a
wrapped headline continuation) could independently trigger title
classification -- now grouped by paragraph like the native PyMuPDF
path, using the paragraph's median line height rather than its total
bbox height for font size.
A vision-engine label (footer/title/page_number/caption) is trusted
outright and skips classify_blocks' font-size heuristics entirely,
since the engine gives no reliable font metric; an ambiguous "text"
region gets a synthesized estimate from its paragraph area and length
(estimate_font_size_from_area) and still goes through normal
classification.
Off unless vision_ocr_url/vision_ocr_model are configured
(vision_ocr_enabled setting, default on if reachable); per-request
override via the ingest endpoint body. Default engine URL uses
host.docker.internal, confirmed live -- the OCR container runs on its
own Docker network, not dgx_net, so a container-name default would not
resolve from the app container.
28 new tests (136 total in the module). ocr_quality_score is
deliberately coarse and documented as such: it catches text that is
mostly symbol noise, not a plausible single-word misread inside an
otherwise normal sentence (confirmed live: a misread "KI" -> "K |"
still scored 0.81) -- that class of error needs a dictionary or
language-model check, out of scope here.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Studio -> Magazines: library, contents tree grouped by section with
confidence/pages/listening-time/audio status, the Stage D review
screen (tick/untick layout blocks, correct title/author, approve or
reject), a reflowed article view with click-to-seek sentence
highlighting, a linked original-page view showing the source bounding
box, and a podcast panel (2-3 host pickers, script generation,
turn-by-turn playback).
Wired into the existing lazy-module system (module-manifest.json,
loader.js, nav.js) rather than a separate app. Reuses /api/tts-preview
for both narration and podcast turns -- no new audio pipeline.
Fixes found building this:
- The pane toggle used role="tab" without the full ARIA tabs pattern
(aria-selected, roving tabindex, arrow-key nav) -- fails WCAG 2.2 AA
and hid the buttons from role-based tooling. Replaced with a button
group + aria-pressed, matching the plain-click interaction it
actually has.
- Print layout cues ("(continued on page 8)") were being sent to TTS
verbatim -- narration would have spoken the parenthetical aloud.
Stripped from the narration copy only; layout.json keeps the
original for the review screen.
- has_podcast now comes from the articles listing so opening an
article without a generated script skips a doomed GET instead of a
routine 404 on every open.
Verified against the real API contract with Playwright (library,
review pre-selection, approval, sentence highlighting + PDF-page
linking, podcast approval gate, zero console errors across the flow)
and manually end-to-end against a real ingested fixture, screenshots
included. Full suite: 94 Python tests, 33/34 desktop Playwright tests
(one unrelated pre-existing flaky DOM-count assertion).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Magazines break the audiobook pipeline's assumptions: several unrelated
articles share a page, columns run in parallel, adverts and sidebars
interrupt body copy, and a feature routinely jumps twenty pages to its
tail. Narrating page-by-page walks out of an article mid-sentence and
into an advert, so the canonical object is the article, not the page.
core/magazine.py reconstructs articles from a flat PDF: PyMuPDF extracts
blocks with bboxes and font metrics, tesseract supplies a TSV fallback so
scanned pages join the same column grouping, blocks are classified and
ordered column-major, and articles are anchored by the parsed table of
contents with "continued on page N" tails linked back.
routes/magazine.py adds 13 endpoints. Ingestion is a restart-safe
background job. Nothing is narrated until a human approves the
boundaries; podcast generation refuses unapproved articles.
Podcast mode renders an approved article as a 2-3 host discussion using
the configured LLM and voices the user picks, alongside narration rather
than instead of it.
Verified end-to-end over HTTP on a synthetic 8-page issue: the feature
came back as pages [3, 8] with its jump tail absorbed and the advert page
excluded, and an image-only rendering recovered all 47 blocks via OCR.
44 new tests; full suite 90 passing.
The Studio UI is not built — this is reachable only via the API.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
