From document to approved German.

Start with a clear original and the right curriculum context.

1. Open New document and choose a PDF or DOCX. Use PDFs up to 25 MB and 30 pages, and Word files up to 3 MB. Convert older .doc files first.
2. Select Turkish, Farsi, Russian, Arabic or English as the source language. German is always the target.
3. Enter the university, programme and curriculum year consistently. These are entered manually and determine which approved knowledge applies.
4. If the workspace owner has set a private pilot access code, enter it; access is then remembered on this browser for 30 days. Without a code, processing starts directly. Confirm that you are authorised to process the document, then click Extract source text. Keep the page open while processing.

Use PDF for scans, stamps, embedded images and page references. DOCX extraction includes text only. Keep the original file separately: Berlin saves extracted text, not the original file bytes.

Ready when: Extracted passages ready for source review.

First approval gate: confirm that the extracted text represents the original.

1. Compare every passage with the original. Check names, numbers, units, table relationships, footnotes and missing sections.
2. Correct extraction errors and click Save source correction for each changed passage.
3. Select the checkbox confirming that you checked the entire source, then click Confirm extracted source. Any later source edit requires confirmation again.

The original PDF preview is temporary and disappears on reload. Open your saved original alongside the app when needed. If a whole passage is missing, use a clearer or shorter input and repeat intake; the pilot has no add/split/merge passage controls. Source text is locked after translation.

Ready when: A human-confirmed source, ready for German translation.

Approved wording is searched before new translation is generated.

1. Click Prepare German translation after the source has been confirmed.
2. Berlin reuses an active approved translation when the source passage, language pair, university, programme and year match exactly and the entry passes the applicable checks.
3. GPT-6 Astra drafts the remaining passages using approved terminology, wording guidance and examples from the same context.
4. Inspect the origin shown for each passage: approved memory or a new AI draft. Both require expert review for this document.

Different hours, a different curriculum year or even a differently written university name can prevent an exact match. Similarity matching and automatic university/year recognition are not part of this pilot.

Ready when: A German draft with a visible origin for each passage.

Second approval gate: take responsibility for each German passage.

1. In Review, compare the source and German side by side. Correct meaning, terminology, wording, numbers and omissions.
2. Resolve the terminology and number warnings where they are real, then choose a Correction category; a false warning can be overridden by ticking the acknowledgement below the check list, and the override is recorded with the approval. Automatic checks help you inspect the text; they do not prove it is correct.
3. Select Save this passage for reuse only when the approved wording should be available to future matching documents.
4. Use Save draft to pause without approving. Use Approve & next or Approve passage when the passage is ready. Cmd/Ctrl + Enter also approves a passage.

With reuse selected, passage approval publishes the source/German pair to memory immediately, before whole-document approval. Saving a draft or choosing a correction category does not publish knowledge or create a general rule.

Ready when: An approved passage; reusable knowledge only when selected.

Third approval gate: check the translation as a complete document.

1. Approve every passage, resolve all remaining checks and save outstanding edits. The document will be Ready to approve.
2. Review completeness, sequence and consistency against the original, then click Approve document.
3. Inspect the approved preview and choose Download sample text to save the German translation as a TXT file.

Output is labelled as a synthetic sample. Formatted Word/PDF export, certification and client delivery are not implemented. Try next document creates another prepared Turkish sample; use New document to upload a new file.

Ready when: An approved sample translation with a plain-text download.

Translations, terminology and wording have different purposes.

1. Translations contains approved source/German passage pairs for exact reuse. Terminology holds preferred term equivalents. Wording holds writing conventions.
2. Open a document with the intended language and context, then choose Knowledge → Propose a rule. Select Terminology or Preferred wording and enter the rule.
3. Choose whether the rule applies to the current curriculum or all documents in the language pair. Click Save proposal, review it, then click Approve rule. Proposed rules are inactive.
4. After approving terminology, inspect affected documents and memory. Incompatible approved passages may be reopened and memory may be marked Needs review.

The learning loop uses approved memory and guidance; it does not retrain model weights. Wording rules guide the model and reviewer, but are not comprehensively enforced. Conflicting duplicate term replacements require a comparison workflow that is not yet implemented.

Ready when: Approved knowledge that informs later work in this workspace.

Corrections remain traceable, and unsuitable wording can be stopped from spreading.

1. Use Reopen for correction on an approved passage, edit it and approve it again. The complete document will also require a new approval.
2. If old wording must stop being reused immediately, separately open Knowledge → Translations → Withdraw from future reuse. Enter a reason and choose Withdraw entry.
3. Inspect Approval & version details and Activity for decisions, versions and available before/after text. A changed approved target for the same source/context creates a new memory version.

Reopening a passage does not automatically withdraw its previous memory entry. Withdrawal prevents future reuse; it does not rewrite existing documents. The pilot history uses a demonstration reviewer identity, not an authenticated employee signature.

Ready when: Reapproved work and a record of what changed.

Know the boundaries before inviting colleagues to the pilot.

1. The private pilot code is optional. When BERLIN_AI_ACCESS_CODE is set (at least 16 characters), enter it once per browser; access is remembered for 30 days, and changing the variable in Vercel and redeploying revokes all remembered access. When it is not set, AI processing is available to anyone who can open the site, so keep a spending limit on the provider key. Enter the code only in Berlin; the provider API key belongs in the server configuration, not in the app's pilot-code field.
2. Use the same browser to return to saved work. Each browser has a separate cookie-bound workspace. Another device or colleague does not see your knowledge, and clearing the cookie opens a fresh workspace.
3. After a timeout, reload and inspect saved state before retrying. Translation failure preserves saved source text. A job may have completed after the browser stopped waiting.
4. Send the owner the document type, selected language, step and error message when processing fails. Keep credentials out of screenshots and messages.

This pilot has no authenticated team accounts, assignments or shared company knowledge base. It allows 20 documents, 50 rules and 500 recorded events per workspace, plus a daily AI processing limit per workspace (10 by default; set BERLIN_AI_DAILY_LIMIT between 1 and 1000). Extraction and translation are separate attempts; failed attempts count. This is not a company-wide spending cap.

Ready when: A supervised pilot with clear recovery steps.