ئەو بڕگەیە بدۆزەرەوە کە پێویستتە لە کۆمەڵێک بەڵگەنامەی یاسایی عەرەبیی سکانکراودا، بەبێ متمانەکردن بە وەڵامی ئامێرەکە

چی بەسەر زانیارییەکانتدا دێت

تێچوو

پێش ئەوەی دەست پێ بکەیت

• A laptop (Windows or Mac). The offline lane genuinely needs a laptop. It cannot be done on a phone. About 4 GB of free disk space for the software and your PDFs.
• A laptop whose disk you are willing and able to encrypt. This is Step 2 and it is not optional. If you cannot turn on BitLocker or FileVault on this machine, do not build a searchable archive of sensitive documents on it.
• Your documents as PDF or image files in one folder. If they are still paper, scan them at 300 dpi in black and white or greyscale, not colour, and not at 150 dpi. Scan quality sets the ceiling on everything that follows.
• Roughly 90-120 minutes the first time, most of it waiting for downloads. Steps 1-4 and Steps 25-26 are paper-and-thinking work and take about 20 minutes on their own. You can do those before you install anything. After the first time, running the pipeline on a new pile takes about 10 minutes of your attention.
• Internet for the installation downloads only, and enough of it: about 250 MB on Windows, about 1.5 GB on Mac. On a metered or intermittent connection, do the installs where the connection is good. After that the offline lane runs with the Wi-Fi switched off.
• A free Google account, ONLY if you intend to use the online lane for already-public documents. Not needed for the offline lane. Be warned that creating one from Iraq often requires SMS verification, see the Iraq payment and access note.
• One colleague who reads the language of the documents fluently and has agreed to do the verification in Step 25. This is a real prerequisite, not a nice-to-have. Without it, do not run this recipe.

هەنگاوەکان

1. هەنگاو 1 / 26هەموو سیستەمەکان

   Before touching any software, sort your pile into two folders. Name them exactly as below. Everything else in this recipe branches off this one decision.

   ئەمە بە دروستی کۆپی بکە

   Folder 1 name: PUBLIC-documents
   Folder 2 name: SENSITIVE-documents

   دڵنیا ببەرەوە کە کاری کردووە

   Every document is now inside one of the two folders and none are left loose. Count them: the two folders together must add up to the number of documents you started with. If you find yourself with a third pile of 'not sure', that pile is SENSITIVE.

   تێبینی

   PUBLIC means the document is already published to the world: an official gazette issue, a published judgment, a bill, a ministry circular, a public tender notice. SENSITIVE means a real person can be identified: client files, witness statements, complaints, internal memos, anything with a name, phone number, address or case number belonging to a living individual.

   THE HARD CASE, WHICH THE PREVIOUS VERSION OF THIS RECIPE GOT WRONG. Public status attaches to the DOCUMENT, not to the safety of the people named in it. Published judgments routinely name parties. Iraqi gazettes publish named appointments, dismissals, and in some periods citizenship and property decisions naming individuals. So: if the document is public but a named person could be harmed by your interest in them becoming visible, a detainee, an asylum seeker, a dismissed official, a witness, anyone whose situation is still live. It goes in SENSITIVE, even though anyone can buy a copy at a kiosk. Uploading it would not leak the document. It would leak that you are looking at that person.

   If you hesitate for more than five seconds about which folder a document belongs in, it goes in SENSITIVE. 86% of the organisations surveyed named information leaving the organisation as their top fear. This folder is where you act on that fear instead of worrying about it.

2. هەنگاو 2 / 26هەموو سیستەمەکان

   Turn on full-disk encryption BEFORE you install anything or copy any document. This recipe is about to turn a pile of unreadable scans into a fast, searchable, full-text archive. That is exactly as useful to whoever takes the laptop as it is to you.

   ئەمە بە دروستی کۆپی بکە

   WINDOWS:
     Settings > Privacy & security > Device encryption  ->  switch it On
     If there is no 'Device encryption' entry: press Start, type  Manage BitLocker  and turn on BitLocker for drive C:
   
   MAC:
     System Settings > Privacy & Security > FileVault  ->  Turn On
   
   Both: set a strong login password at the same time. Encryption does nothing if the machine is left logged in and unlocked.

   دڵنیا ببەرەوە کە کاری کردووە

   Windows: the Device encryption switch reads On, OR the Manage BitLocker panel shows 'BitLocker on' for drive C:. Mac: the FileVault panel says FileVault is turned on for your disk. Encryption can take an hour or more to finish in the background after you switch it on. You can carry on with the recipe, but do not copy sensitive documents onto the machine until the panel stops saying it is still encrypting. If neither setting exists on your edition of Windows, stop and read the note before you put any sensitive document on this machine.

   تێبینی

   Why this is Step 2 and not an appendix. For Iraqi documenters and journalists the dominant threat is usually not a cloud leak. It is a laptop taken at a checkpoint, at a border, or in an office raid. Everything this recipe creates is plain readable text at rest: the text layer inside each searchable PDF, the .txt transcripts from Step 20, and the storage folder of the optional tool in Step 23. Without encryption you have made your archive more useful to an adversary, not less.

   RECOVERY KEY WARNING, which matters for this audience. Windows backs up the BitLocker recovery key to your Microsoft account by default; macOS offers to escrow the FileVault key with iCloud. That means Microsoft or Apple can be compelled to hand over the key. When you are offered the choice, pick the local option: on Mac choose 'Create a local recovery key' rather than the iCloud option, and write it on paper stored somewhere other than the laptop bag. On Windows you can print or save the key to a file on a USB stick. If you lose the key and the password, the data is gone permanently. That is the whole point of it, so store the paper carefully.

   If your edition of Windows offers neither Device encryption nor BitLocker, use VeraCrypt (free, open source, veracrypt.io, version 1.26.29, released 9 June 2026, verified 30 Aug 2026) to make an encrypted container file and keep the work folder from Step 3 inside it. That is more clicking, and it is better than nothing by a very large margin.

3. هەنگاو 3 / 26هەموو سیستەمەکان

   Make a plain local work folder that is NOT inside OneDrive or iCloud, and put your documents there. Do not use the Desktop. Do not use Documents.

   ئەمە بە دروستی کۆپی بکە

   WINDOWS (File Explorer):
     Open  This PC > Local Disk (C:)  and create a new folder named:  ocr-work
     Then inside it create another folder named:  docs
     The address bar must read exactly:   C:\ocr-work\docs
   
   MAC (Finder):
     Press Shift + Command + H to open your home folder, create a new folder named:  ocr-work
     Then inside it create another folder named:  docs
     The path must be:   /Users/YOURNAME/ocr-work/docs
   
   Now copy the PDFs you want to process into that docs folder.

   دڵنیا ببەرەوە کە کاری کردووە

   WINDOWS: click into the folder and read the address bar. It must say C:\ocr-work\docs. If it says anything containing OneDrive, you are inside a cloud folder, delete it and create it again at the root of the C: drive. Second check: right-click the ocr-work folder. If you see a green tick, a cloud icon, or a 'Free up space' / 'Always keep on this device' option in the menu, it is being synced to Microsoft. Move it.

   MAC: the folder must sit directly in your home folder alongside Desktop, Documents and Downloads, not inside them. To confirm iCloud is not involved, open System Settings > [your name] > iCloud > Drive and look at 'Desktop & Documents Folders'. If that is on, those two folders are cloud folders; ~/ocr-work is not, which is why we use it.

   تێبینی

   This is the single most dangerous default in the whole exercise, and the previous version of this recipe walked straight into it by telling readers to put test.pdf and a docs folder on the Desktop.

   On most Windows 10 and 11 machines signed into a Microsoft account, a feature called OneDrive Known Folder Move is on by default and your Desktop IS a OneDrive folder. Copying a witness statement onto it uploads that witness statement to Microsoft. There is no prompt, no warning and no error message; the file simply appears in OneDrive later. The same applies on macOS when iCloud 'Desktop & Documents Folders' sync is enabled.

   So the recipe that was written to stop sensitive material leaving the organisation was, in its own instructions, causing sensitive material to leave the organisation. Making one folder in the right place costs thirty seconds and closes it.

4. هەنگاو 4 / 26هەموو سیستەمەکان

   Find out what kind of PDFs you actually have. Open one PDF and try to select a line of text with your mouse, as if you were going to copy it. Then paste what you selected into Notepad (Windows) or TextEdit (Mac).

   ئەمە بە دروستی کۆپی بکە

   Open one PDF  ->  drag the mouse across a line of text  ->  Ctrl+C (Windows) or Command+C (Mac)  ->  paste into Notepad / TextEdit

   دڵنیا ببەرەوە کە کاری کردووە

   You are now in exactly one of three cases, and you need to write down which one each file is in:

   (a) The text highlighted letter by letter and pasted correctly. This PDF already has real text. It needs no OCR at all, go straight to Step 22 to search it, or Step 9 if it is public and you want to ask questions about it.

   (b) Nothing highlighted; your mouse just drew a blue rectangle over a picture. It is a scan, a photograph of words. No search tool on earth can find anything in it until a text layer is added. That is what Steps 18 to 21 do.

   (c) Text highlighted, but what you pasted is garbage, random letters, boxes, or Latin characters where Arabic should be. This PDF has a BAD text layer from earlier, poor OCR. This is the case people miss, and Step 18 gives you a different command for it.

   تێبینی

   Most Iraqi official gazette PDFs are case (b). Old digitisation projects sometimes produce case (c), and case (c) is dangerous because a search over it silently returns nothing while the tool reports success.

   If your documents are public and you want the better-quality online route, the online lane is Steps 5 to 10. If they are sensitive, skip Steps 5 to 10 entirely and go to Step 11.

5. هەنگاو 5 / 26هەموو سیستەمەکان

   ONLINE LANE, PUBLIC DOCUMENTS ONLY. If your documents are in the SENSITIVE folder, skip Steps 5 to 10 completely and go to Step 11. If they are public, open Google AI Studio in your browser and sign in with a free Google account.

   ئەمە بە دروستی کۆپی بکە

   https://aistudio.google.com

   دڵنیا ببەرەوە کە کاری کردووە

   You can see a chat box with a model name at the top of the page, and at no point were you asked for a credit card. If Google demanded an SMS verification code and you could not receive it, stop here: the online lane is closed to you, and the offline lane from Step 11 does everything you genuinely need. Do not try to work around a verification wall with a borrowed number, an account you do not control is a worse place for your research history than no account.

   تێبینی

   Verified 30 Aug 2026: Google's own available-regions page lists Iraq explicitly, immediately before Ireland, so this works from Erbil or Baghdad without a VPN, and Google's pricing page states verbatim 'Google AI Studio usage is free of charge in all available regions'. No credit card is requested, which sidesteps the Iraqi-card rejection problem entirely.

   The price you pay instead is stated by Google on the same pricing page, as a row in its own comparison table: free tier, 'Content used to improve our products'; paid tier, 'Content not used to improve our products'. For a document that is already published, that costs you nothing.

   But be clear about the second thing you are spending. Uploading a public document leaks nothing about the document, the world already has it. It does reveal what you are investigating: which gazette issue, which decree number, which case, on which date, tied to your account and your IP address. For a human rights organisation the pattern of interest is often more sensitive than any single document in it. Decide that you are comfortable with that before you upload, not after.

6. هەنگاو 6 / 26هەموو سیستەمەکان

   Upload one public PDF using the paperclip or 'Insert' button, then paste this exact prompt and send it. Do one document first, not the whole pile.

   ئەمە بە دروستی کۆپی بکە

   Transcribe this document verbatim into Arabic text. Do not summarise, do not correct, do not modernise the wording. Preserve the original line and paragraph structure. Mark the start of every page as [صفحة 1], [صفحة 2] and so on. Where the scan is unreadable, write [غير مقروء] instead of guessing the word. Do not invent any text that you cannot actually see.

   دڵنیا ببەرەوە کە کاری کردووە

   The reply is Arabic text that visibly corresponds to your page, it contains at least a [صفحة 1] marker, and it ends at the end of a sentence rather than stopping mid-word. If it stopped mid-word or mid-sentence, it ran out of room, go straight to Step 7, which you must do anyway. Save the result into a plain text file inside your ocr-work folder, next to the PDF.

   تێبینی

   The two instructions that matter most are the last two. A model asked to transcribe a smudged line will happily invent a plausible legal phrase to fill the gap, and you will never spot it because it reads perfectly. Forcing it to write [غير مقروء] converts an invisible fabrication into a visible hole you can go and check with your own eyes.

7. هەنگاو 7 / 26هەموو سیستەمەکان

   Count the pages. Compare the number of page markers in the transcript against the number of pages in the actual PDF. Do this before you read a single word of the transcript.

   ئەمە بە دروستی کۆپی بکە

   1. Open the PDF and note its total page count, e.g. 40 pages.
   2. In the transcript, find the LAST page marker:  [صفحة 40]  ?
   3. The last marker number must equal the PDF's page count.
   4. Also check that no marker numbers are missing in the middle: 1, 2, 3 ... with no gaps.

   دڵنیا ببەرەوە کە کاری کردووە

   Last marker number equals the PDF page count, and there are no gaps in the sequence. If the PDF has 40 pages and the last marker is [صفحة 12], then 28 pages are missing and nothing told you. If markers jump from [صفحة 5] to [صفحة 9], four pages were skipped.

   تێبینی

   This step is new, and it exists because the previous version's spot-check could not catch the failure it most needed to catch.

   Every model has a maximum reply length. Give it a forty-page gazette and it will transcribe until it hits that limit and then stop. The stopping point looks like a perfectly clean ending, the last paragraph is well-formed, the Arabic is fluent, nothing is marked. If your only check is to inspect 'the start, the middle and the end' of the transcript, you inspect the end of the transcript, which is fine, and never discover that it is not the end of the document.

   A missing second half is a safety problem, not a quality problem: it is how an organisation concludes that a gazette contains no reference to something when in fact the model never read that part.

   THE FIX when it truncates: do it in chunks. Re-send the Step 6 prompt with the extra line 'Transcribe pages 11 to 20 only.' and repeat until the page markers add up. Ten pages at a time is a reasonable starting size. Then join the chunks and run this count again on the joined file.

8. هەنگاو 8 / 26هەموو سیستەمەکان

   Spot-check the transcription before you trust any of it. Pick three random places in the transcript, near the start, the middle and the end, and find the same passage on the actual scanned page with your own eyes.

   دڵنیا ببەرەوە کە کاری کردووە

   For each of the three passages you can point at the same words on the scanned image. If you cannot find one of them anywhere on the page it claims to come from, you have found a fabrication and the transcript is not usable for legal work.

   تێبینی

   You are checking for two different failures. Small character errors (hamza, taa marbuta) are normal and harmless for searching. Whole invented sentences, a paragraph that appears in the transcript but not on the page, or a page that quietly got skipped are disqualifying, if you find one, fall back to the offline lane and read the pages yourself. Do this check every time. It takes four minutes and it is the difference between a research tool and a liability.

9. هەنگاو 9 / 26هەموو سیستەمەکان

   To work across a whole set of public documents rather than one, open Gemini Notebook, create a new notebook, and upload your public PDFs as sources.

   ئەمە بە دروستی کۆپی بکە

   https://notebook.google.com

   دڵنیا ببەرەوە کە کاری کردووە

   The notebook lists your uploaded files as sources, and when you ask a question the answer carries clickable citations that jump into the source document. If a file uploaded but the notebook cannot quote anything at all from it, that file is an image-only scan and the notebook is not reading it, see the note.

   تێبینی

   Use the address exactly as written above. The previous version of this recipe had a misspelled domain in this field, which for people in this line of work is a security problem and not a typo: a mistyped address is how a lookalike site gets your Google password.

   Name and address, verified 30 Aug 2026: the product Google used to call NotebookLM is now called Gemini Notebook in Google's own support documentation, and notebooklm.google.com issues a 301 redirect to notebook.google.com.

   Verified free-plan limits from Google's support page, 30 Aug 2026: 100 notebooks, up to 50 sources per notebook, 500,000 words per source, 50 chat queries per day, 3 audio generations per day. Fifty sources is enough for a year of one gazette or one case file's worth of public filings; it is not enough for a whole archive, so split by year or by topic into separate notebooks.

   Its advantage for legal work is specific and real: answers carry clickable citations back into the source document, so verification is one click rather than a hunt. Its limit is equally specific: it cannot read an image-only scan any better than anything else. If your PDF is a pure scan, either upload the transcript text file from Step 6, or OCR the file first using Steps 11 to 21 and upload the searchable- version.

10. هەنگاو 10 / 26هەموو سیستەمەکان

    Ask your question, but require a citation in the question itself. Paste this template and fill in your own question where marked.

    ئەمە بە دروستی کۆپی بکە

    Answer this question using only the documents I have given you: [YOUR QUESTION HERE].
    
    For every single statement in your answer, give me: the document name, the page number, and the exact sentence in the original Arabic that supports it. If the documents do not answer the question, say clearly "غير موجود في المستندات" and stop. Do not use any knowledge from outside these documents. Do not tell me what the law probably says.

    دڵنیا ببەرەوە کە کاری کردووە

    Every sentence in the answer carries all three things: a document name, a page number, and a quoted Arabic sentence. If any sentence in the answer has no citation attached, that sentence is not usable, ask again, or treat it as absent. An answer where the model says غير موجود في المستندات is a good answer, not a failure.

    تێبینی

    'Do not tell me what the law probably says' is doing heavy lifting. Left to itself, a model will smoothly blend what it read in your gazette with what it half-remembers about Iraqi law generally, and the join is invisible. Demanding the exact original sentence for every claim gives you something checkable: a fabricated citation usually cannot produce a verbatim Arabic sentence that actually appears on the page you open. Step 25 is where you actually open it.

11. هەنگاو 11 / 26هەموو سیستەمەکان

    OFFLINE LANE, START HERE FOR SENSITIVE DOCUMENTS. This lane is where you go when the documents name a living person. Read this routing box carefully; it is the one place people get lost.

    ئەمە بە دروستی کۆپی بکە

    IF YOU ARE ON WINDOWS  ->  your next step is Step 12.
    IF YOU ARE ON A MAC     ->  your next step is Step 15.
    
    Both platforms meet again at Step 18.
    
    Windows readers: do not run any command beginning with 'brew'. Those are Mac-only and will fail.
    Mac readers: do not run any command beginning with 'winget' or 'py'. Those are Windows-only and will fail.

    دڵنیا ببەرەوە کە کاری کردووە

    You know which single number you are going to next, and it is 12 if you are on Windows. The previous version of this recipe sent Windows readers to the macOS Homebrew installer at this exact point, which meant pasting a Unix command into PowerShell and getting an incomprehensible error. If you find yourself looking at a command that starts with /bin/bash on a Windows machine, you are in the wrong section.

    تێبینی

    Download volume before you begin, so you can choose where to do it. Windows: about 250 MB in total. Mac: about 1.5 GB, and more if Apple's command line developer tools are not already installed. On a metered or intermittent connection, do the installs somewhere with a good connection and then take the laptop away. Everything after Step 18 works with the Wi-Fi switched off, and Step 18's check asks you to prove that.

12. هەنگاو 12 / 26ویندۆز

    WINDOWS ONLY. Install Tesseract, the OCR engine, together with the Arabic language data. Open your browser, go to the address below, and download the 64-bit Windows installer (the file ending in .exe).

    ئەمە بە دروستی کۆپی بکە

    https://github.com/UB-Mannheim/tesseract/wiki
    
    The current file is named like:  tesseract-ocr-w64-setup-5.5.3.20260724.exe
    (the version number will be higher by the time you read this, take the newest 64-bit one)

    دڵنیا ببەرەوە کە کاری کردووە

    After the installer finishes, open File Explorer and confirm that the folder C:\Program Files\Tesseract-OCR exists and contains tesseract.exe. The full proof that Arabic installed comes in Step 14, do not skip it.

    تێبینی

    CRITICAL, and the single most common point of failure: during the installer, when you reach the screen listing components, expand the section called 'Additional language data (download)' and TICK 'Arabic'. If you click Next past this screen without ticking Arabic, Tesseract installs English-only, and every Arabic OCR command later will fail with a message about a missing language.

    Also, on the installer screen that offers it, choose the option to add Tesseract to the system PATH (wording varies by version; it may appear as 'Add to PATH' or as an install-for-all-users choice). If you are not offered it, do not worry, Step 14 shows you how to check and how to fix it.

    IF YOU ALREADY MADE THE ARABIC MISTAKE, do not reinstall. Download the Arabic model by hand from the link below and put the file into C:\Program Files\Tesseract-OCR\tessdata (Windows will ask for administrator permission to write there, which is normal):

    https://github.com/tesseract-ocr/tessdata_best/raw/main/ara.traineddata

    That file is 12.6 MB, verified present in the tessdata_best repository on 30 Aug 2026. The tessdata_best models are the slowest but the most accurate, which is what you want on damaged gazette scans.

13. هەنگاو 13 / 26ویندۆز

    WINDOWS ONLY. Install Python, then Ghostscript by hand, then OCRmyPDF. Do the three parts in this order. Use a NORMAL PowerShell window, not 'Run as administrator'.

    ئەمە بە دروستی کۆپی بکە

    PART 1, Python.
      Click Start, type  PowerShell , open 'Windows PowerShell' NORMALLY (do NOT run as administrator).
      Paste and press Enter:
    
      winget install -e --id Python.Python.3.12
    
    PART 2, Ghostscript. This one is NOT available through winget. Download it in your browser:
    
      https://ghostscript.com/releases/gsdnld.html
    
      Choose the 'Ghostscript AGPL Release' for Windows (64 bit). The file is named like  gs10071w64.exe  (64.9 MB).
      Run it and accept the default install folder, which will look like  C:\Program Files\gs\gs10.07.1
    
    PART 3: OCRmyPDF.
      CLOSE every PowerShell window. Open a NEW normal PowerShell window. Then paste:
    
      py -m pip install ocrmypdf

    دڵنیا ببەرەوە کە کاری کردووە

    Part 1 ends with 'Successfully installed'. Part 3 ends with a line beginning 'Successfully installed ocrmypdf-' followed by a version number (17.11.0 or higher). If Part 3 says 'py : The term "py" is not recognized', Python either did not install or you did not open a fresh window, close every PowerShell window, open exactly one new one, and run Part 3 again. Do not proceed to Step 14 until you have seen the 'Successfully installed ocrmypdf-' line.

    تێبینی

    WHY GHOSTSCRIPT IS DOWNLOADED BY HAND. The previous version of this recipe told you to run `winget install -e --id ArtifexSoftware.GhostScript` and attributed that command to OCRmyPDF's official documentation. Both halves were wrong. That package was removed from the winget catalogue on 10 October 2023 and the command now fails with 'No package found matching input criteria', verified 30 Aug 2026 against Microsoft's winget-pkgs repository, where the ArtifexSoftware folder now contains only 'mutool'. And OCRmyPDF's own installation page says the opposite of what was claimed: 'You will need to install Ghostscript manually, since it does not support automated installs anymore.' Ghostscript is a hard dependency, so on the old instructions every later OCR command failed too. Third-party winget mirrors still list the dead package, which is how the error survived a casual check.

    WHY A NORMAL WINDOW, NOT ADMINISTRATOR. winget installs Python for your user account. An administrator PowerShell runs as a different account and may not see it, so `py -m pip` fails in a way that is very hard to diagnose. Use a normal window for all three parts.

    WHY `py -m pip` AND NOT BARE `pip`. Bare `pip` is frequently not on PATH after a fresh Windows Python install. `py` is the Python launcher that Windows installs reliably, and `py -m pip` is the form OCRmyPDF's documentation recommends.

    Verified 30 Aug 2026: ocrmypdf 17.11.0 published to PyPI on 28 Aug 2026, requires Python 3.11 or newer, so the 3.12 above is correct. Ghostscript 10.07.1 released 19 May 2026; the Windows 64-bit installer on its release page is gs10071w64.exe at 64.9 MB. The winget identifier Python.Python.3.12 was confirmed to still exist in the winget catalogue.

14. هەنگاو 14 / 26ویندۆز

    WINDOWS ONLY. Prove that all three tools can actually be found before you try to OCR anything. Open a new normal PowerShell window and run these four commands one at a time.

    ئەمە بە دروستی کۆپی بکە

    tesseract --version
    tesseract --list-langs
    gswin64c --version
    ocrmypdf --version

    دڵنیا ببەرەوە کە کاری کردووە

    All four must print something, and one specific thing must appear:

    1. `tesseract --version` prints a version line such as 'tesseract v5.5.3...'. 2. `tesseract --list-langs` prints a header line like: List of available languages in "C:\Program Files\Tesseract-OCR/tessdata/" (N): followed by one three-letter code per line. THE LINE `ara` MUST BE IN THAT LIST. If it is not, Arabic did not install, go back to Step 12 and use the manual ara.traineddata download. 3. `gswin64c --version` prints a number like 10.07.1. 4. `ocrmypdf --version` prints a number like 17.11.0.

    If any of them says 'The term ... is not recognized as the name of a cmdlet', that program is installed but Windows cannot find it. See the note for the fix. Do not continue until all four print.

    تێبینی

    This step exists because a missing PATH entry is the single most common Windows problem with this toolchain, and it produces error messages later that mean nothing to a non-technical reader.

    TO ADD SOMETHING TO PATH: Press Start, type environment variables , open 'Edit the system environment variables'. Click the 'Environment Variables...' button. In the lower box ('System variables'), select the row named 'Path' and click 'Edit...'. Click 'New' and type the folder that contains the missing program: for Ghostscript: C:\Program Files\gs\gs10.07.1\bin for Tesseract: C:\Program Files\Tesseract-OCR (If your Ghostscript version number is different, use the folder name you actually see in C:\Program Files\gs.) Click OK on all three windows. CLOSE PowerShell completely and open a new window, then run the four commands again.

    The manual Ghostscript installer definitely does not add itself to PATH, so if exactly one of the four fails it is usually gswin64c.

15. هەنگاو 15 / 26ماک

    MAC ONLY. Open the Terminal application: press Command and the space bar together, type the word Terminal, and press Enter. Then install Homebrew, the tool that installs the other tools. Copy this entire line, paste it, and press Enter. It will ask for your Mac login password, type it and press Enter. The password will NOT appear on screen as you type; that is normal, keep typing.

    ئەمە بە دروستی کۆپی بکە

    /bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"

    دڵنیا ببەرەوە کە کاری کردووە

    When it finishes, type brew --version and press Enter. It must print something like 'Homebrew 4.x.x'. If it says 'command not found: brew', the installer printed two lines beginning with (echo near the end and told you to run them, scroll up, copy those two lines exactly as printed, paste them, press Enter, then close the Terminal window, open a fresh one, and try brew --version again.

    تێبینی

    If you have never opened this window before: it is a place where you type an instruction and press Enter, and the computer does it. It will print a lot of text while it works. That text is progress, not errors. You do not need to read it. When it stops and shows your username with a $ or % sign again, it has finished and is waiting for the next instruction.

    Homebrew takes 5 to 15 minutes and downloads several hundred megabytes; it may also install Apple's command line developer tools, which is a further large download. Skip this step entirely if you already have Homebrew.

16. هەنگاو 16 / 26ماک

    MAC ONLY. Install the OCR software and the Arabic language data. Paste this one line and press Enter, then wait.

    ئەمە بە دروستی کۆپی بکە

    brew install ocrmypdf tesseract-lang unpaper ghostscript

    دڵنیا ببەرەوە کە کاری کردووە

    The command returns you to the prompt without a line beginning 'Error:'. The real proof is Step 17, which you must do next.

    تێبینی

    This installs four things at once: OCRmyPDF (which adds a searchable text layer to a scanned PDF while leaving the scan looking identical), Tesseract's full language pack including Arabic, unpaper (which cleans up speckles and shadows on bad photocopies), and Ghostscript (which OCRmyPDF needs internally).

    Verified 30 Aug 2026 against Homebrew's own formula listing: all four formulae exist and are current, ocrmypdf 17.11.0, tesseract-lang 4.1.0, unpaper 7.0.0, ghostscript 10.07.1. This matches OCRmyPDF's official macOS installation instructions, which give `brew install ocrmypdf` plus `brew install tesseract-lang` for non-English languages.

    Expect 10-20 minutes. Measured on the machine used to write this recipe: tesseract-lang alone occupies 654 MB installed and Ghostscript 258 MB, so allow roughly 1 GB of disk on top of Homebrew itself. Everything installed here is free, open-source, and works with the internet switched off from now on.

17. هەنگاو 17 / 26ماک

    MAC ONLY. Prove the tools work and that Arabic is really there. Run these two commands.

    ئەمە بە دروستی کۆپی بکە

    ocrmypdf --version
    tesseract --list-langs

    دڵنیا ببەرەوە کە کاری کردووە

    `ocrmypdf --version` prints a version number such as 17.11.0. `tesseract --list-langs` prints a header line then a long list of three-letter language codes, and THE LINE `ara` MUST BE IN IT. On the machine used to write this recipe the header read: List of available languages in "/opt/homebrew/share/tessdata/" (163): and ara was present. If ara is missing, run brew install tesseract-lang again. If ocrmypdf prints a long red error about a library that could not be loaded rather than a version number, run brew reinstall ocrmypdf and try again.

    تێبینی

    While that list is on your screen, look for two other codes, because this is the Kurdish problem made visible in one command rather than argued about.

    Search the list for `ckb` (Sorani, Central Kurdish). It is not there. Search for `kur`. It is not there either. Search for `kmr`. That one IS there, and it is Kurmanji, Northern Kurdish, written in Latin script, which is not what a Sorani document is written in.

    That is a complete Tesseract language installation, 163 languages, with no Sorani model in it. On the machine used to write this recipe, on 30 Aug 2026, that is exactly what it printed. Read the Kurdish Sorani section of this recipe before you spend an afternoon on Sorani material.

18. هەنگاو 18 / 26هەموو سیستەمەکان

    BOTH PLATFORMS MEET HERE. Test on ONE bad page before committing to the whole pile. Copy a single difficult PDF into your work folder, rename it test.pdf, then run the command that matches your platform AND the case you found in Step 4.

    ئەمە بە دروستی کۆپی بکە

    WINDOWS (PowerShell), for case (b), a page that is a pure picture:
    cd C:\ocr-work
    ocrmypdf -l ara --skip-text --deskew --rotate-pages --sidecar test.txt test.pdf test-searchable.pdf
    
    MAC (Terminal), same case (b):
    cd ~/ocr-work
    ocrmypdf -l ara --skip-text --deskew --clean --rotate-pages --sidecar test.txt test.pdf test-searchable.pdf
    
    BOTH PLATFORMS: for case (c), a PDF that already has a BAD, garbled text layer.
    Note that --deskew and --clean must be left OUT of this one:
    ocrmypdf -l ara --redo-ocr --sidecar test.txt test.pdf test-searchable.pdf

    دڵنیا ببەرەوە کە کاری کردووە

    Three things must ALL be true. Check all three; the first one on its own is not enough.

    1. The command finished with lines about 'Postprocessing' and an 'Output file is a PDF/A' message, and gave you the prompt back with no red error. 2. A file named test-searchable.pdf now exists in the folder. 3. A file named test.txt now exists AND IS NOT EMPTY. Open it in Notepad or TextEdit. You must see Arabic words in it that you can recognise from the page.

    An empty or one-line test.txt after a run that reported success is the failure mode that matters most here: OCR found nothing, and nothing told you. It usually means the page is too degraded, the language code is wrong for what is on the page, or the page was blank. Try -l ara+eng, try a different page, and if five of your worst pages all come out empty, this archive needs better scans before it needs better software.

    تێبینی

    WHAT EACH PART DOES. -l ara, the language is Arabic. Use -l ara+eng if pages mix Arabic and English, which Iraqi official documents often do. --skip-text, OCR pages that are pictures, and leave completely alone any page that already has real text. --deskew, straightens a crooked scan. --clean, removes photocopy speckle. Mac only: it needs unpaper, which is awkward on Windows, so the Windows line leaves it out. --deskew and --rotate-pages still do most of the useful work. --rotate-pages, fixes pages scanned upside down or sideways. --sidecar test.txt, also writes the recognised text into a plain text file. This is what makes Step 22's search possible without installing anything else, and it is what makes the empty-output check above possible at all.

    WHY --skip-text AND NOT --force-ocr. The previous version of this recipe used --force-ocr on everything. --force-ocr rasterises every page and throws away any existing text layer, replacing correct embedded text with OCR guesses, so a born-digital PDF that happened to be sitting in your folder came out worse than it went in, silently. --skip-text never touches a page that already has text. This was tested while writing this recipe: running the --skip-text command on a PDF that already had a text layer printed 'skipping all processing on this page' and left it alone.

    WHY --redo-ocr HAS NO --deskew. If you try to combine them, OCRmyPDF refuses to start and prints: '--redo-ocr (or --mode redo) is not currently compatible with --deskew, --clean-final, and --remove-background'. That exact message was reproduced on OCRmyPDF 17.10.0 on 30 Aug 2026 while writing this recipe. OCRmyPDF's documentation explains why the trade is acceptable: redo mode 'will replace OCR without rasterizing', so it repairs a bad text layer without degrading the page image.

    If --redo-ocr does not fix a case (c) file, --force-ocr is the last resort for that one file, named deliberately, used deliberately, never applied to a whole folder.

    Pick your genuinely worst page for this test, not your best one. Ten minutes here tells you whether the whole approach will work on your archive.

19. هەنگاو 19 / 26هەموو سیستەمەکان

    Confirm the text layer actually works, with your own eyes, before doing anything else. Open test-searchable.pdf. Look at the scanned image and pick a clear Arabic word you can see, ideally in the middle of a page. Now press Ctrl+F (Windows) or Command+F (Mac) and type that same word.

    دڵنیا ببەرەوە کە کاری کردووە

    The word is found and highlighted on the page where you saw it. If nothing is found: on Windows, check that Arabic really installed by re-running `tesseract --list-langs` from Step 14 and confirming `ara` is in the list; on Mac, re-run the same check from Step 17. If ara is present and the search still fails, the page was too degraded for OCR, see Step 18's note.

    تێبینی

    This is the verification step people skip, and skipping it is how organisations discover, after processing 400 documents, that none of them are searchable.

    One important quirk that alarms people unnecessarily: Arabic text copied out of a PDF often looks scrambled or reversed when pasted into a program that does not handle right-to-left text properly. That is a display problem in the other program, not an OCR failure. Searching still works correctly.

20. هەنگاو 20 / 26هەموو سیستەمەکان

    Process the whole folder at once. Your PDFs should already be in the docs folder you made in Step 3. Run the block for your platform.

    ئەمە بە دروستی کۆپی بکە

    WINDOWS (PowerShell), paste all four lines together:
    cd C:\ocr-work\docs
    $files = @(Get-ChildItem -File -Filter *.pdf | Where-Object { $_.Name -notlike "searchable-*" })
    "Found $($files.Count) PDFs to process"
    foreach ($f in $files) { ocrmypdf -l ara --skip-text --deskew --rotate-pages --sidecar "$($f.BaseName).txt" $f.Name "searchable-$($f.Name)" }
    
    MAC (Terminal), paste both lines together:
    cd ~/ocr-work/docs
    for f in *.pdf; do case "$f" in searchable-*) continue;; esac; ocrmypdf -l ara --skip-text --deskew --clean --rotate-pages --sidecar "${f%.pdf}.txt" "$f" "searchable-$f"; done

    دڵنیا ببەرەوە کە کاری کردووە

    On Windows the second line prints 'Found N PDFs to process' before anything else happens, check that N matches the number of original PDFs you put in the folder. Then Step 21 does the real verification: do not skip it.

    تێبینی

    This creates two new files next to each original, searchable-something.pdf and something.txt, and never modifies your originals. If it goes wrong you have lost nothing.

    Budget roughly 20-60 seconds per page on an ordinary laptop, so a 500-page gazette is a lunch break and a folder of 100 documents is something you start before going home. The laptop can be offline the entire time; if you want to prove that to yourself, turn off the Wi-Fi first.

    If one file fails, the loop carries on to the next. Step 21 finds the ones that failed.

    TWO THINGS THE PREVIOUS VERSION GOT WRONG HERE. First, it used --force-ocr, which destroys good embedded text (see Step 18's note). Second, its Windows loop streamed the file list, so files created during the run got picked up and processed again, producing searchable-searchable-something.pdf and doubling the runtime. The @( ) wrapper and the -notlike "searchable-*" filter above exist to stop exactly that.

    HONESTY ABOUT TESTING: the Mac line was run start to finish on a real folder while writing this recipe, on 30 Aug 2026, and produced one searchable- PDF and one non-empty .txt per original. The PowerShell block could not be executed on the machine used to write this recipe, because that machine is a Mac. Have one Windows colleague run it on a real Windows laptop before you teach it to a room.

21. هەنگاو 21 / 26هەموو سیستەمەکان

    Verify the batch. This is the step that catches the documents that silently produced nothing. Run the block for your platform, in the same docs folder.

    ئەمە بە دروستی کۆپی بکە

    WINDOWS (PowerShell):
    (Get-ChildItem -File -Filter *.pdf | Where-Object {$_.Name -notlike "searchable-*"}).Count
    (Get-ChildItem -File -Filter searchable-*.pdf).Count
    Get-ChildItem -File -Filter *.txt | Where-Object { $_.Length -lt 100 } | Select-Object Name, Length
    
    MAC (Terminal):
    ls *.pdf | grep -vc '^searchable-'
    ls searchable-*.pdf | wc -l
    find . -name "*.txt" -size -100c -print

    دڵنیا ببەرەوە کە کاری کردووە

    Two conditions, both required.

    1. The first two numbers must be EQUAL. If you had 40 originals you must have 40 searchable- files. If the second number is smaller, some files failed, find which original has no searchable- twin, run the Step 18 command on that one file alone, and read the error it prints.

    2. The third command must print NOTHING AT ALL. Every file it prints has a text transcript under 100 characters, meaning OCR found essentially nothing on that document. Those documents are NOT searchable, and a search across the folder will silently skip them. Write their names down and either re-scan them at better quality or plan to read them by hand.

    On the test run for this recipe both counts were 2 and the third command printed nothing.

    تێبینی

    This step is the answer to the worst failure mode in the whole recipe: a tool that reports success and produces an empty result. Everything upstream of here can look fine while a quarter of your archive is invisible to search. Ten seconds of counting closes that gap.

    Do not tell a colleague 'the archive is searchable' until you have run this and both conditions held.

22. هەنگاو 22 / 26هەموو سیستەمەکان

    Search across everything. There are two ways and you will use both: one document at a time inside the PDF, and all documents at once across the .txt transcripts.

    ئەمە بە دروستی کۆپی بکە

    ONE DOCUMENT AT A TIME, works everywhere, needs no extra software:
      Open any searchable-....pdf in any PDF reader and press Ctrl+F (Windows) or Command+F (Mac).
    
    ACROSS THE WHOLE FOLDER, search the .txt transcripts:
    
    WINDOWS (PowerShell), in C:\ocr-work\docs :
    Select-String -Encoding utf8 -Path *.txt -Pattern "4712"
    
    MAC (Terminal), in ~/ocr-work/docs :
    grep -in "4712" *.txt
    
    Replace 4712 with what you are actually looking for.

    دڵنیا ببەرەوە کە کاری کردووە

    Test it with something you KNOW is there. Pick a decree number, a year or an article number that you can see with your own eyes on one of the scanned pages, and search for that. The command must print the filename and the matching line. On the test run for this recipe, searching for a number printed the filename, the line number and the whole line for each matching document.

    If you get no output at all for a term you can see on the page: either the OCR missed it on that document (open that document's .txt file and look), or you searched a longer phrase than the OCR got right. Try a shorter word. Never conclude 'the archive does not contain X' from a single failed search.

    تێبینی

    ARABIC SEARCH ADVICE THAT WILL SAVE YOU AN HOUR. Search SHORT words. OCR gets long inflected phrases slightly wrong often enough that exact-phrase searching fails, while a short root word like قانون or مادة or قرار still hits. Better still, search for a distinctive NUMBER, an article number, a decree number, a year, because digits survive bad OCR far better than Arabic letters do.

    Searching by number has a second benefit on Windows: typing Arabic into a PowerShell window can be awkward, and the console may render right-to-left text badly even when the match is correct. If you need to search an Arabic word, either paste it in from Notepad, or use Ctrl+F inside the searchable PDF, which handles right-to-left properly.

    WHY THERE IS NO SEARCH APP TO INSTALL HERE. The previous version of this recipe recommended Recoll at this point. It has been removed. Its own Windows page asks for a payment contribution to download the Windows build (the author will send it free if you email him and ask, but that means emailing a stranger from your work address before you can search your own archive), and on Apple Silicon Macs, every Mac sold since 2020. Its download is unsigned, and its own macOS page states that 'The ARM-based Mac computers will absolutely refuse to run unsigned code', requiring you to code-sign the app yourself from the Terminal. Both verified 30 Aug 2026. For a room with no engineer and a majority on Windows, that is a hard stop dressed up as a one-liner. Searching the .txt files costs nothing and works identically on both platforms.

    Before you finish the day, read Step 24. Those .txt files are readable copies of your documents.

23. هەنگاو 23 / 26هەموو سیستەمەکان

    OPTIONAL, offline question-answering. Only do this if your laptop has 16 GB of RAM. Check first; this is a go/no-go, not a suggestion.

    ئەمە بە دروستی کۆپی بکە

    STEP A, check your RAM before downloading anything:
      WINDOWS:  Settings > System > About  ->  read 'Installed RAM'
      MAC:      Apple menu >  About This Mac  ->  read 'Memory'
    
      If it says 8 GB:  STOP. Skip this step entirely. Steps 20-22 already do the useful work.
      If it says 16 GB or more:  continue.
    
    STEP B, download AnythingLLM Desktop:
      https://anythingllm.com/desktop
    
    STEP C, BEFORE you add a single document, open the app and go to:
      sidebar > Privacy  ->  turn telemetry OFF

    دڵنیا ببەرەوە کە کاری کردووە

    The Privacy panel shows telemetry disabled, and you can see that BEFORE you have added any document. If you added documents first, remove them, turn telemetry off, and add them again. Then, to confirm it really is running locally: disconnect the Wi-Fi and ask it a question about a document you added. If it answers, the model is on your machine. If it errors, it is configured to use a cloud provider and you must fix that before putting anything sensitive into it.

    تێبینی

    WHY THE RAM CHECK IS A GATE. AnythingLLM's own system requirements page gives the recommended specification as 16 GB of RAM and an 8-core CPU. The previous version of this recipe stated that correctly and then added 'it will run on less but slowly', which buries the decision. For an audience where organisations are already blocked by hardware and are working on old 8 GB laptops, a 16 GB recommendation is closer to a prerequisite than a footnote. On 8 GB it will either swap continuously or fail, after a multi-gigabyte download you may have paid for by the megabyte.

    WHY THE TELEMETRY STEP IS MANDATORY. AnythingLLM's download page says: 'Your models, documents, and chat history stay on your machine. Nothing phones home.' Its own project README says something different about metadata: telemetry is a thing you 'opt out' of, 'Set DISABLE_TELEMETRY in your server or docker .env settings to "true" to opt out of telemetry. You can also do this in-app by going to the sidebar > Privacy and disabling telemetry.' Both quotes verified 30 Aug 2026.

    What is actually sent is metadata, not content: the installation type, the fact that 'a document is added or removed. No information about the document', the vector database type, the LLM provider and model tag, and the fact that 'a chat is sent'. The README states 'No IP or other identifying information is collected'. So your documents are not leaving. But 'nothing phones home' is not true out of the box, and for a room where 86% fear information leaving the organisation, the correct response is an instruction, not an omission.

    WHAT YOU ACTUALLY GET. A small model running on a laptop writes considerably weaker Arabic than a cloud model and will miss nuance in legal language. Treat its answers strictly as a pointer to which document and page to open, which is genuinely useful, and never as a reading of the law. Everything in Step 25 still applies to every sentence it produces.

24. هەنگاو 24 / 26هەموو سیستەمەکان

    Know what is now sitting on your disk in plain readable text, and delete it when the project ends. Do this before you close the laptop for the day.

    ئەمە بە دروستی کۆپی بکە

    THESE ARE ALL READABLE TEXT COPIES OF YOUR DOCUMENTS, on this laptop:
      WINDOWS:  C:\ocr-work\docs\*.txt              the transcripts from Step 20
                C:\ocr-work\docs\searchable-*.pdf   the text layer inside each PDF
      MAC:      ~/ocr-work/docs/*.txt
                ~/ocr-work/docs/searchable-*.pdf
      Plus AnythingLLM's storage folder, if you did Step 23.
    
    WHEN A PROJECT ENDS, delete the transcripts:
      WINDOWS (PowerShell):  Remove-Item C:\ocr-work\docs\*.txt
      MAC (Terminal):        rm ~/ocr-work/docs/*.txt

    دڵنیا ببەرەوە کە کاری کردووە

    After deleting, run the Step 22 folder-search command again for a word you know was in those documents. It must now return nothing at all, on Windows, Select-String will report that it cannot find the path because no .txt files remain; on Mac, grep will say 'No such file or directory'. That is the correct result. If it still prints matches, the files are still there.

    If the project is completely finished and you only need the original scans, delete the searchable- PDFs too. They carry the same text inside them.

    تێبینی

    Deleting a file does not wipe it from the disk; it removes the pointer. On an encrypted disk that is acceptable, because the whole disk is unreadable without your password. On an unencrypted disk it is not. That is the whole reason full-disk encryption is Step 2 and not an afterthought.

    Think about it this way: before this recipe, a seized laptop gave someone a folder of unreadable scans they would have to sit and read. After it, the same laptop gives them a searchable, indexed archive with a text file per document. The capability you built for yourself is the same capability you built for whoever takes the machine. Encrypt the disk, keep the work folder out of OneDrive and iCloud, and clear the transcripts when you are done.

25. هەنگاو 25 / 26هەموو سیستەمەکان

    MANDATORY VERIFICATION PROTOCOL. Before any AI-derived statement leaves your organisation, in a filing, a report, a press release, or advice to a client, run every single point below. Print this and put it on the wall.

    ئەمە بە دروستی کۆپی بکە

    1. SOURCE: every claim carries a document name AND a page number. No page number means the claim does not exist yet.
    2. OPEN IT: a human opens that exact page and reads the sentence with their own eyes. Not the transcript. The original scan.
    3. THE INSTRUMENT IS REAL: confirm the gazette issue number, date, decree or article number against the printed gazette or the official register, never against the AI, and never against a second AI.
    4. VERBATIM MATCH: the Arabic sentence the model quoted must appear on that page. Approximately is a failure.
    5. COMPLETENESS: if the claim rests on a transcript, confirm the transcript covers the whole document, page markers counted against the PDF's page count, as in Step 7. An answer drawn from half a gazette is wrong in a way no citation check will reveal.
    6. ASK TWICE: put the same question in a fresh, empty chat. If the two answers disagree on any citation, discard both.
    7. LANGUAGE CHECK: a fluent Arabic or Kurdish reader confirms the meaning, not just the presence, of the passage.
    8. SIGN IT: write in the file who verified each citation and on what date.
    9. LOG THE MISSES: when this protocol catches a fabricated citation, decree number or quotation, write it down, date, tool, the exact prompt used, and what it invented. Keep that log.
    10. THE RULE: nothing unopened gets filed. No exceptions, no deadlines, no "it looked right".

    دڵنیا ببەرەوە کە کاری کردووە

    Take one real AI-produced claim from your own work this week and run all ten points on it, timing yourself. If it takes more than about ten minutes per claim, your prompts are not demanding enough citation detail, go back to the Step 10 template. If any point cannot be completed, the claim does not leave the organisation.

    تێبینی

    Why this is not optional. In Mata v. Avianca, a New York lawyer filed a brief containing case citations that ChatGPT had invented. On 22 June 2023 Judge P. Kevin Castel sanctioned the lawyers USD 5,000, finding they had acted with subjective bad faith. That is a verified, checkable case, and it is the one this recipe stakes its argument on.

    A statement about later trends has been REMOVED from this recipe. The previous version cited an academic database of AI hallucination cases in court, attributed it to named institutions, and gave a count of over 1,300 instances. That database could not be opened during this revision, the site returned HTTP 403, so neither the attribution nor the figure could be confirmed, and an unverifiable citation has no place in the paragraph arguing that unverifiable citations destroy cases. If you want a current figure, find the source and read it yourself before you quote it to a room.

    What does not need a citation is the mechanism. Every lawyer in every one of these cases believed the citation looked correct. It always looks correct. That is the entire failure mode. A fabricated Iraqi decree number is indistinguishable from a real one until someone opens the gazette. The model is a finding aid that tells you where to look. It is never the finding.

    Point 9 is new. A caught fabrication is the single most useful artefact your organisation will produce for its own policy: it is local, specific, undeniable evidence for the colleague who thinks this caution is excessive. The previous version let those near-misses evaporate.

26. هەنگاو 26 / 26هەموو سیستەمەکان

    Write your organisation's one-page rule and give it to everyone, including interns and volunteers. Four sentences is enough to start.

    ئەمە بە دروستی کۆپی بکە

    1. Documents naming a living person never go into any online AI tool, free or paid, ever, including a public document that names someone whose situation is still live.
    2. No AI-produced citation, quote or legal conclusion leaves this organisation until a named human has opened the source page and signed for it.
    3. Public official documents may be used with cloud tools; when in doubt, a document is not public.
    4. Every time we catch the tool inventing something, we write it down, date, tool, prompt, what it invented, and we keep the list.

    دڵنیا ببەرەوە کە کاری کردووە

    Print it, tape it above the scanner, and ask one colleague who was not in the room to read it and tell you which folder a specific real document from your pile belongs in. If they hesitate or get it wrong, the wording is not clear enough yet, fix the wording, not the colleague.

    تێبینی

    71% of the organisations surveyed have no written AI policy, and 43% have already put sensitive material into an AI tool, usually because nobody had ever said out loud where the line was. These four sentences, printed and taped above the scanner, close most of that gap in an afternoon. Add to it later; do not wait for a perfect policy.

    This recipe contains that rule in full, inline, on purpose. Do not send anyone to a policy-builder page or an online template to get it, write these four sentences on paper today.