diff --git a/docs/step4.md b/docs/step4.md
new file mode 100644
index 0000000..20810d2
--- /dev/null
+++ b/docs/step4.md
@@ -0,0 +1,190 @@
+## Step 4: `services/upload.py` + `worker.py`
+
+## Objective
+
+Implement the MVP upload and background processing pipeline so the system can:
+
+1. Save uploaded files into `UPLOAD_DIR`
+2. Create `Document` + `Job(status="queued")`
+3. Process queued jobs in a worker loop:
+   - `queued -> processing`
+   - invoke Step 3 transcription
+   - persist `Transcript`
+   - finalize as `transcribed` or `failed`
+
+This covers MVP Feature 1 + 2 and supports REQ-1, REQ-2, REQ-3, REQ-4, REQ-6.
+
+---
+
+## Scope
+
+### In scope
+- `src/transcription/services/upload.py`
+- `src/transcription/worker.py`
+- Upload persistence logic + job creation
+- Worker polling and single-job processing lifecycle
+- Tests for upload + worker behavior (deterministic, no live API)
+
+### Out of scope
+- UI wiring (Step 5)
+- Job/transcript pages (Step 5)
+- External queue infrastructure
+- Async DB/session architecture changes (post-MVP)
+
+---
+
+## Planned Deliverables
+
+### Source files
+- `src/transcription/services/upload.py`
+- `src/transcription/worker.py`
+- `src/transcription/services/__init__.py` (export updates as needed)
+
+### Tests
+- `tests/services/test_upload.py`
+- `tests/services/test_worker.py` *(recommended for mirror consistency)*
+
+---
+
+## Design Decisions
+
+1. **Upload service owns file persistence + initial DB records**
+   - Save file first, then create `Document` + `Job(queued)`.
+
+2. **Worker owns job lifecycle transitions**
+   - Transition only in worker:
+     - `queued -> processing`
+     - `processing -> transcribed|failed`
+
+3. **Worker calls Step 3 service, not provider SDK directly**
+   - Use `transcribe_document_image(...)` from `services/transcription.py`.
+
+4. **Failure details are always persisted**
+   - On failure, persist transcript row with `text=None`, `error_detail=...`.
+
+5. **Worker loop is stoppable**
+   - Simple in-process loop with stop signal/event and poll interval.
+
+---
+
+## Task-by-Task Execution Checklist
+
+## Phase A — Upload service (`services/upload.py`)
+
+- [ ] Create `src/transcription/services/upload.py`
+- [ ] Add `UploadError` exception type
+- [ ] Add result dataclass (e.g., `UploadJobResult`) with:
+  - [ ] `document_id`
+  - [ ] `job_id`
+  - [ ] `stored_path`
+  - [ ] `original_filename`
+- [ ] Implement safe filename handling:
+  - [ ] basename normalization
+  - [ ] collision-safe naming (e.g., UUID prefix)
+  - [ ] reject path traversal patterns
+- [ ] Implement extension validation (`.jpg/.jpeg/.png/.tif/.tiff/.pdf`)
+- [ ] Implement empty-bytes validation
+- [ ] Ensure `UPLOAD_DIR` exists (`mkdir(parents=True, exist_ok=True)`)
+- [ ] Persist file bytes to disk
+- [ ] Insert `Document` and `Job(status=queued)` in one transaction
+- [ ] Return structured `UploadJobResult`
+- [ ] Add logging for upload success/failure (no sensitive payload logging)
+
+---
+
+## Phase B — Worker core (`worker.py`)
+
+- [ ] Create `src/transcription/worker.py`
+- [ ] Implement `process_next_queued_job(...) -> bool`
+  - [ ] fetch oldest queued job
+  - [ ] return `False` if none found
+  - [ ] set job to `processing` + update timestamp
+  - [ ] call `transcribe_document_image(image_path=...)`
+  - [ ] success path: persist transcript text, mark `transcribed`
+  - [ ] failure path: persist error detail, mark `failed`
+  - [ ] return `True` when a job is processed
+- [ ] Implement transcript upsert behavior (avoid duplicate unique `job_id` insert issues)
+- [ ] Add structured logs for lifecycle transitions and outcomes
+
+---
+
+## Phase C — Worker loop (`worker.py`)
+
+- [ ] Implement `run_worker_loop(...)`
+  - [ ] accepts stop signal/event
+  - [ ] configurable `poll_interval_seconds`
+  - [ ] repeatedly calls `process_next_queued_job`
+  - [ ] sleeps when no work
+  - [ ] exits cleanly on stop request
+- [ ] Keep implementation single-process/simple for MVP assumptions
+
+---
+
+## Phase D — Service exports
+
+- [ ] Update `src/transcription/services/__init__.py` exports
+  - [ ] include upload service symbols
+  - [ ] keep transcription exports intact
+
+---
+
+## Phase E — Tests (scaffold -> fill)
+
+### E1 Scaffold
+- [ ] Create `tests/services/test_upload.py` with class/method skeletons + one-line docstrings
+- [ ] Create `tests/services/test_worker.py` with class/method skeletons + one-line docstrings
+- [ ] Validate scaffold collection:
+  - [ ] `uv run pytest --collect-only -q`
+
+### E2 Fill
+
+#### Upload tests
+- [ ] `test_create_upload_job_writes_file_and_creates_records` (`integration`)
+- [ ] `test_create_upload_job_rejects_empty_bytes` (`unit`)
+- [ ] `test_create_upload_job_rejects_unsupported_extension` (`unit`)
+- [ ] `test_create_upload_job_uses_unique_stored_filename` (`integration`)
+- [ ] `test_create_upload_job_sets_job_status_queued` (`integration`)
+
+#### Worker tests
+- [ ] `test_process_next_queued_job_returns_false_when_queue_empty` (`integration`)
+- [ ] `test_process_next_queued_job_transitions_processing_to_transcribed` (`integration`)
+- [ ] `test_process_next_queued_job_persists_transcript_text_on_success` (`integration`)
+- [ ] `test_process_next_queued_job_sets_failed_and_error_detail_on_failure` (`integration`)
+- [ ] `test_process_next_queued_job_updates_existing_transcript_if_present` (`integration`)
+- [ ] `test_run_worker_loop_stops_when_stop_event_is_set` (`unit`)
+
+### Marker usage
+- [ ] Mark pure logic tests as `@pytest.mark.unit`
+- [ ] Mark DB/service orchestration tests as `@pytest.mark.integration`
+- [ ] Reuse existing `external` only for future live endpoint checks (not required in Step 4)
+
+---
+
+## Phase F — Verification
+
+- [ ] `uv run pytest --collect-only -q`
+- [ ] `uv run pytest tests/services/test_upload.py -q`
+- [ ] `uv run pytest tests/services/test_worker.py -q`
+- [ ] `uv run pytest -m "unit or integration" -q`
+- [ ] `uv run pytest -q`
+
+---
+
+## Implementation Guardrails
+
+- [ ] Do not couple worker directly to provider SDK internals
+- [ ] Do not swallow exceptions silently
+- [ ] Always persist terminal failure details
+- [ ] Keep status transitions explicit and timestamped
+- [ ] Keep default test suite deterministic and fast (no live network)
+
+---
+
+## Definition of Done (Step 4)
+
+- [ ] Upload service persists file + creates `Document` and queued `Job`
+- [ ] Worker processes queued jobs end-to-end via Step 3 service
+- [ ] Success writes transcript text and final `transcribed` status
+- [ ] Failure writes `error_detail` and final `failed` status
+- [ ] Queue-empty worker call returns cleanly without error
+- [ ] Full test suite passes with `uv run pytest -q`