diff --git a/README.md b/README.md
index 5166af2..3db8ee2 100644
--- a/README.md
+++ b/README.md
@@ -56,6 +56,8 @@ Use the editor for your own SQL, inspect a query plan, revisit query history or
**Use Tab to accept a suggestion that follows your current schema.** Choose word completion, the next word, the next three words or the rest of a line. Suggestions use local SQL vocabulary, table names and columns. This editor assistance works without an AI account or model call; asking the tutor to generate SQL is a separate feature.
+Longer completion modes extend a suggestion only when the continuation is unambiguous. Choosing the next three words does not force three guessed words: the editor may offer a shorter completion when several tables, columns or SQL clauses would be valid. Accepting a suggestion preserves the SQL after the cursor. See the [editor architecture](docs/en/architecture.md) for how local workers handle completion and highlighting.
+
## Learn a concept, then try it

diff --git a/README.tr.md b/README.tr.md
index 2f2a6fd..245ab1b 100644
--- a/README.tr.md
+++ b/README.tr.md
@@ -56,6 +56,8 @@ Kendi SQL’ini yaz, sorgu planını incele, geçmiş sorgulara dön veya sohbet
**Mevcut şemanı izleyen öneriyi Tab ile kabul et.** Kelime, sıradaki kelime, sonraki üç kelime veya satırın kalanı için tamamlama seçebilirsin. Öneriler yerel SQL sözlüğünü, tablo ve sütun adlarını kullanır. Editörün bu yardımı AI hesabı veya model çağrısı gerektirmez; öğretmenden sorgu yazmasını istemek ayrı bir özelliktir.
+Uzun tamamlama modları, devamı belirsiz olmadığında öneriyi genişletir. Sonraki üç kelimeyi seçmek, her durumda üç kelime önerileceği anlamına gelmez: birden fazla tablo, sütun veya SQL ifadesi uygun olduğunda öneri daha kısa kalabilir. Öneriyi kabul etmek, imleçten sonraki SQL'i korur. Yerel iş parçacıklarının tamamlamayı ve renklendirmeyi nasıl işlediğini [editör mimarisinde](docs/tr/architecture.md) inceleyebilirsin.
+
## Bir kavram öğren, sonra dene

diff --git a/docs/architecture/architecture.json b/docs/architecture/architecture.json
index c1af9ac..27669f5 100644
--- a/docs/architecture/architecture.json
+++ b/docs/architecture/architecture.json
@@ -1159,7 +1159,7 @@
"references": {
".github/SECURITY.md": "470353c7e0447bd45fea3e4040507f03253c3cc762bac12ed424e9f6f93d648d",
"LICENSE": "c03af9304b8b44d8f65921a325a7f2077ebb89e5c64fc309106b3c547ef02d09",
- "README.md": "74e2b60b292faa39f6624a22429f01d9e234a993e9056ad889ecfb320e20281c",
+ "README.md": "245c647ac23f66cb8b28ded1885a4777dbf0ceeee47e83ec59b82a46582a1ec1",
"app/ai-rate-limit.js": "fdf24873dfb664444f28728abe483a3f363e3aa61d16201a2cb1429ef67a2ff8",
"app/chatgpt-account.js": "1ea42388b92788b93804a9f0b6fc3441039ce26af2b837e9e5e732c160a99f32",
"app/database-inspection/runner.cjs": "32265046ec123e4248845436a523f67ff069bb7ebd8bb83102fc43e249d11d20",
diff --git a/docs/architecture/index.html b/docs/architecture/index.html
index 1dddae0..f4c8cf8 100644
--- a/docs/architecture/index.html
+++ b/docs/architecture/index.html
@@ -1,6 +1,6 @@
PractiSQL · How it works
PractiSQL
How the application works
Local SQL · AI tutor · Actual results
From a question to SQL. From SQL to learning.
The model interprets intent. The application supplies scoped context, runs eligible SQL in local SQLite and brings actual results back into the conversation. Choose a workflow and component below to explore its responsibilities.
01 / 08
Main flow
The model selects the SQL approach and teaching explanation. The application harness prepares context, applies eligible output, checks specific violations and carries actual results into the next turn.
The current request, history and database evidence are labelled separately. The schema is not a copy of every row; missing record information can be retrieved with a tool.
These are the two read-only tools in normal SQL chat. Final SQL follows a separate execution path. Data-project discovery and review tools appear in the project view.
Successful execution does not prove that the business request or explanation is entirely correct. Specific violations are checked; not every semantic or teaching error can be detected.
The editor, chat, exams, projects and lesson PDFs have separate contracts. Lesson PDFs use their own local fixtures, rather than reading a personal workspace.
An optional, separate workflow. Source, SQL and numeric checks provide different evidence from AI judgments about business meaning. Numeric success does not prove a flawless interpretation or assignment.
This map explains component responsibilities. Successful SQL execution alone does not prove that every user intention or AI explanation is correct. Signing infrastructure is prepared; the current local installer is unsigned.