`w:bidiVisual` was appended to the end of `w:tblPr` and `w:bidi` to the end of
`w:sectPr`. Both have a fixed position in the OOXML schema, and out of place
they are dropped silently - so every table in the Arabic bag rendered left to
right with «م» on the wrong side, while the code said otherwise.
Both are inserted in schema order now, the styles carry the direction too, and
the index columns are pinned in `tblGrid` with a fixed layout so «م» stays
narrow on the right and «الصفحة» narrow on the left. The appendix renderer had
the same two lines and the same fix.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Five types where there was one, and a case study that runs
`case_study_pipeline.ipynb` is the fourth notebook to come across and the
first to arrive without a new class of defect: eighteen modules, every prompt
carried over character for character, and the three faults the earlier ports
each contributed a bug to - an agent factory passed instead of called, a model
that comes back from the store as a dict, and an int key that comes back as a
string - all covered by tests before the first run rather than after it. Five
dry tests exercise the JSON round trip, the int-key restore, the review
corrections and both exporters without opening a socket.
The instructional-design type was four independent producers behind one
checkbox list. The notebook's own comment says why they must not share a crew;
nothing in them shares anything else either, so someone who wanted a pre/post
test was picking one box out of four on a form that asked for all of it. Each
is a type of its own now - «أنشطة تدريبية», «أدوات تطبيقية», «اختبار قبلي
وبعدي», «دراسة حالة تحليل…
Teach it before you test it: the interactive infographic's forty findings
The first interactive infographic that ever finished, reviewed. The structure
was right - covers, zoom, intro, main screen, three elements, an interaction
of a different kind after each, closing, thank-you, the freeze, the feedback
boxes, the check button, the unrecorded X. What went wrong was the teaching
and the timing.
**The run's own stored outputs settle three open questions and correct one
root cause.** `include_unit_cover` was true, so the invented programme and
unit lines came from a cover that was asked for. The brief carries no
`next_context`, so a closing with no bridge is right. `tips: []` and
`tips_nature: "none"`, so no tips scene is right. And `technical_terms: []` -
the analyst found **none**; «Faith / Worship / Standard / Hotspot» was the
*writer's* invention, so the fix is the writer's constraint and an example
taken out of the skeleton, not only a tighter analyst definition.
**One hole made findings 1, 2, 15, 16, 20 and 21.** The analysis returned ten
flat blocks a…
One script in the file, and an eight-step run that finishes
**The export that killed a finished run.** Eight steps, 2h32m, $0.3735, seven
of them succeeded - and the last one raised `ValidationError: ReviewOut
overall Field required`. `step_review` returns a report, `{"rounds",
"lint_after", "warnings"}`, and the export was validating it as `ReviewOut`,
which is the reviewer agent's own output schema and a different thing.
`assemble` only stores it, so there was nothing to validate for.
**Two scripts in one file.** The interactive-motion run hit its duration for
the first time - 597 words of 600 - and then handed the designer and the
narrator different text. The VO table said «يمثل المعيار… ركيزة أساسية» and
the motion breakdown said «يُعد المعيار… الركيزة الأساسية»; the interaction
said «تأثر الأب» and the motion said «أبدى الطفل دهشته». Same scene.
The cause is in the review: a scene whose re-direction is refused keeps the
motion written for the *previous* script, and `b.vo_line = line` never runs on
it because there is no new object to run…
Stop putting answers in the skeletons, and let the reviewer see the source
The review of «ركائز البناء الإيماني في الأسرة»: eighteen findings on one
produced element, applied.
**The cost, first.** Run e5e66982 cost **$0.4204** - the most expensive run of
any pipeline here, four times the video's for less output. Four of its eight
steps are on `reasoning`, which was still `gemini-2.5-pro`; the review alone
took 4m52s of an 11-minute run. What had already been done to cut calls was
real - `one_call_then_fill` on the interactions and their descriptions,
`triage` so a style note is written down instead of paid for again, `repair`
for what is arithmetic - and none of it touched the model tier, which is the
whole difference. `reasoning` moves to flash, `visual` moves up to it, and
`_NOTEBOOK_MODELS` keeps the record. The defects below are copied examples and
missing context; none of them is a frontier model's job.
**One mistake made three of the four worst findings.** A real value sitting in
a skeleton where a description belonged. The `"..."` entries beside them…
Refuse to draw God, count the path a learner takes, and drop the gate
The fifth review of a delivered file, applied.
**The one that cannot ship.** The file drew «أيقونة يد كبيرة تمتد من الأعلى
(رمز لقدرة الله)» in its closing scene and «يد تشير للأعلى (رمز للخالق)» on an
option card. `lint_divine_symbol` refuses those - the only check in the run
that raises rather than reports - in the motion, the interactions and the
covers, and the rule with its rejected example is in all four checklists. It
is deliberately narrow: «ابتغاء وجه الله» and «يد الأب تمتد من الأعلى» pass,
because a rule this severe costs a call every time it is wrong.
`lint_worship_negative` reports the other half - «لوحة أسماء الله الحسنى
وطفلان ينظران بملل» - where the wrong option became a picture of worship as
dull. The wrong option is the lesser answer, never the worse one.
**Why six minutes delivered 3.8.** Two arithmetic errors, both in the same
direction. An interaction was billed for `feedback_correct` *and*
`feedback_incorrect_final`, which are alternatives - about 44 words, 26
…
Apply the third review, and give the unit cover a real name
**The unit.** «— يُستكمل —» came from me, and from two decisions colliding:
the form stopped asking for the project, the course, the unit and the element
number, and then the review asked for the missing rows to be shown. A
placeholder on a field nobody is ever asked about is a blank that can never
be filled. The form asks for the unit now, optionally; left empty, the
designer names the subject area from the content itself - every video here
stands on its own, so it is a subject and never «الوحدة الثالثة». The rows
the form does not ask about are simply not part of the file.
**The duration, again.** `FEEDBACK_FLOOR` was 20 while `validate_interaction`
refuses anything under 35, so every interaction was flagged for obeying the
budget - the same class of own-goal as the description length. The floor is
the standard's own floor now, `MIN_SCENES` is five because an outcome follows
every interaction, and the shortest simulation with the mandatory two
interactions is five and a half minutes…
Apply the second review: the standard wins, and the budget says what it costs
Every point from the review of «غرس العقيدة», plus the two defects the
previous round introduced.
**The duration, decided.** The team's answer is that an interaction's text
counts against the running time - the last feedback block included - so
`INTERACTION_TIME_WEIGHT` is 1 and `spoken_words` counts the block the
notebook's own word count leaves out. That makes the arithmetic unforgiving
and honest: a simulation with the mandatory two interactions needs about four
and a half minutes, and `shortest_minutes` says so before anything is paid
for, rounded up rather than to the nearest - rounding to the nearest had it
recommend four minutes for a brief four minutes misses by two words.
The budget also stopped stepping. A threshold at three minutes handed a
3.5-minute video the generous feedback budget and made it come out *less*
feasible than a 3-minute one; the feedback cap is now derived from what is
actually left, so a longer video is never harder to fit. And the caps are per
part - si…
Make the motion video meet its brief, and stop paying twice for it
Acting on the review of «حين تتكلم السماء» and on what the run's own telemetry
said. Five defects, all traced to a cause in the code.
**The duration.** Two minutes asked for, four delivered. `step_scripts`
computed `max(50% x target, target - interaction_words)`, which handed the
writer a hundred words for a video whose two interactions had already spent
237 of its 200 - and the floor is what made that look like a target instead
of an impossible brief. `budget.py` does the arithmetic once: what an
interaction really costs (measured at 110-120 spoken words, not the 90 the
notebook assumed), how much of that counts against the running time (half,
because the video pauses and the learner reads - a named constant, not a
formula), how many scenes fit at forty-five words each, and how long a
feedback block may be in a short video. When a brief does not fit it says so,
with the two ways out: longer, or fewer interactions. The numbers reach the
design, the writer and the interaction designer …
Ask the two infographics for a handful of things, and each agent once
The same three changes the motion video pipeline just had, applied to both
infographics. No prompt either notebook wrote is altered; what is added is
appended and named.
The interactive form asked for nineteen fields and now asks for five, of
which one is required: the content, the voice-over language, the audience,
whether to add a unit cover, and notes. The static one asked for eighteen and
now asks for four. Every field that left the form is still handed to the
agents, with its own default - which is the notebook's value - so the screen
pattern, the element count, the attempts policy, the character and the visual
identity are unchanged, they are simply no longer a blank box inviting a
worse answer than the standard.
Both forms replace two controls with one. The notebook split the language
decision into a language and, if English, a variant, so the form showed a
second control that meant nothing half the time. It is now one choice -
Arabic, British English, American English, or aut…
`infographic_pipeline.ipynb`, third of three: seven agents that turn raw
content into a complete interactive element - one of four screen patterns, a
zoom-in opening, an introduction and tips, the main screen, every element with
its definition and points, interactions with feedback, a storyboard for every
scene, and the two covers. Text only, and the scene description is written for
the graphics team in Arabic.
Every standard, every pattern brief, every agent and every prompt is the
notebook's to the character - including cell [10], the largest in the file,
which holds every measurable constraint and decides what an agent is asked to
do again. `test_interactive_infographic_fidelity.py` proves it in forty-two
checks.
This notebook writes its standards as `"""\` with a line continuation and keeps
several of its tables as tuples, so the fidelity test compares the *evaluated*
value rather than the source text - stricter, and honest about what the agent
actually reads.
The platform's par…
`static_infographic_crewai_pipeline.ipynb`: seven agents that turn raw subject
content into a complete infographic element - a layout profile chosen from
three, a definition box and stations with their points, interactions with their
feedback, a visual description for every station and every interaction, and the
two covers.
Every standard, every layout brief, every agent's role and backstory and every
prompt is the notebook's to the character. `test_static_infographic_fidelity.py`
reads the .ipynb and proves it in thirty-eight checks, including one that walks
every Arabic line the notebook puts in front of a model and finds it in the
ported module.
Three of those checks were wrong about the notebook before they were right
about the port, which is the useful kind of failure: `STD_LINK_PHRASES` is
assembled from a list rather than written out, `STD_GENERAL_DESIGN_NOTES` is a
list, and `STD_ALL_BODY` joins a chosen six rather than all of them - the
title-design and storyboard standards …
`motion_video_multiagent_v2.ipynb`: seven agents that turn raw teaching content
into a production-ready script - voice-over, an executable scene description
for the graphics team, the interactions, and the two covers every video opens
on - for three kinds of video, each with its own structure.
Every prompt, every standard, every agent's role and backstory, every JSON
skeleton and every lint pattern is the notebook's to the character.
`test_motion_video_fidelity.py` reads the .ipynb and proves it: forty-two
checks, including one that walks every Arabic line the notebook puts in front
of a model and finds it in the ported module.
What the platform adds is the orchestration. Each of the seven becomes a step,
so a failure costs that stage rather than the script and a resumed run picks up
where it stopped. The run stops once, at the instructional design - the title,
the axes, the objectives, the visual system and the scene plan - because the
six steps after it are all written against that…
Give the blueprint file a template rather than a dump
Read again and cut again. Gone: the audience, the description, the learning
outcome. The file names itself off what it turns out to hold - «التصور الفني
لكتاب», «التصور الفني لحقيبة تدريبية» - rather than borrowing the gate's
heading, which names a screen and not a document.
And it is laid out now instead of printed. One rule decides every line, from
what the row holds rather than from how deep it sits: a row that holds other
rows is a heading, a row that holds none is a bullet. That gives a book
chapters over numbered محاور over bulleted مواضيع, and a bag days over sessions
over مواضيع over bulleted محاور, with no per-task-type layout anywhere.
A chapter is 17pt in ultramarine with a rule under it and no number in front,
because «الفصل الأول» already says which one it is. Its children are numbered
and smaller, because «الفرق بين الأتمتة والمساعدة» does not. Their children are
bullets, because a list of leaves reads as a list. Each level sits further in
than the one above.
The name …
The first draft printed everything the run was holding: ninety-one paragraphs
for a two-chapter book, most of them the agent's commentary on its own
headings. Read and cut down to what was asked for.
Gone: the count over a list the reader can count («المحاور (4)» over four
محاور), the number in front of «الفصل الأول» which already says which one it
is, the overview, the purpose, the page estimate and the difficulty. Gone too
the header the agent wrote for itself - the repeated title line, the page
total, the editorial notes. What is left is the shape, the name, the audience,
and the learning outcome, which is the one line a chapter owes that its title
does not already say and the one that keeps coming back empty. Thirty
paragraphs.
Running it over a real bag found three more, all the same kind. A day carries
only `day_number`, so it printed as «عنصر 1» - it is now «اليوم الأول», with
the digits taking over once the words run out. `cover_nature` sits on every
single محور and is the ba…
Keep the author's header at the gate, and let the gate add and export
Four things wrong at the review gate, three of them one cause.
The structure was protected from the agent's edits and the header was not.
Every header field goes into the outline prompt and comes back out of it, and
`restore` returned `{**outline, "chapters": restored}` - so the title, the
audience, the description and the page count stayed whatever the agent sent
back. An audience written carefully by the designer returns paraphrased and
shorter; a page count typed into a form returns as the agent's number, and it
is the number every later stage divides its budget by. Restored now, with one
exception that is the rule rather than an exception to it: a field the author
left blank has no author's version to put back, and a description written for
a book that had none is the agent doing its job.
The fourth: `Chapter.outcome` is required in the schema with a floor of twelve
characters and the prompt prints it in the JSON it asks for, and on a real
two-chapter run both chapters came back …
Put the bag's shape guarantee where it can actually work
Asked whether the last change covered both pipelines, and it did not. The book
half was right. The bag half was wrong twice, and both faults read exactly
like success.
The repair sat in `BlueprintStep.run`, after `_design` returns - and both of
the things it protects live inside `_design`. `_honour_supplied` does
`dict(topic)` and raises «dictionary update sequence element #0 has length 1»;
`drop_activity_topics` does `topic.get("topic")` and raises «'str' object has
no attribute 'get'». Both consume the draw before the old placement ran. A
guarantee placed after the code it protects is not one. It now runs on each
route the moment the draw exists: after `_enrich` on the supplied route, and
after the days check on the route that draws from scratch.
And `BAG_DAY` begins at `sessions` while a blueprint begins at `days`, so
handed a blueprint it found no key it recognised, repaired nothing and
reported nothing - indistinguishable from a tree that was already sound. Hence
`BAG_BLUEPRINT`…
Guarantee the shape of what a structuring agent returns
Both pipelines already cut a supplied design into its own parts and ask an
agent about one part at a time - chapters for a book, days for a bag. That is
what keeps a long design from coming back truncated and it works: Aya's
eight-chapter design was split correctly, and `book.units` held every chapter
and every topic exactly as she typed them.
What neither did was check the shape of a part before accepting it. The test
was "did anything come back", and a list with something in it passes that
however wrong the something is.
Her seventh chapter was four bare headings - «على مستوى الفرد:» with nothing
written under any of them. The outline agent, having nothing to structure,
invented sub-points for one of them and returned them as plain strings
carrying the source document's own Word bullet glyph:
"subsections": ["o\tدعم المحتوى العربي.", …]
Nothing noticed. Six steps and twenty-one minutes later the planner did
`sub.get("title")` on a string. One malformed item out of forty-one s…
Make the page count binding without touching the محاور
The planner no longer receives the evidence it never reads, the quality
pass is off on its own measurements, the outline is held to the number of
مواضيع the page count pays for, a محور is written in one call instead of
one per موضوع, and the writer's model comes from settings.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
ContiAI platform: API and queue worker in one container (part 2)