diff --git a/docs/plans/2026-02-16-spot-redaction-design.md b/docs/plans/2026-02-16-spot-redaction-design.md
new file mode 100644
index 0000000..e312a29
--- /dev/null
+++ b/docs/plans/2026-02-16-spot-redaction-design.md
@@ -0,0 +1,95 @@
+# Spot Redaction Detection Design
+
+## Problem
+
+The current auto-detection pipeline (`detect_redactions.py`) using OpenCV morphological operations has three failure modes:
+1. **False positives** — dark areas (shadows, header backgrounds) detected as redactions
+2. **Multi-line merging** — adjacent redactions on different lines merged into one box
+3. **Missed redactions** — small or low-contrast redactions not detected
+
+## Solution
+
+Replace auto-detection with user-guided spot detection. The user double-clicks directly on a redaction box. The backend uses connected-component analysis to find the exact extent of that dark region, then automatically runs OCR + font detection on the surrounding text.
+
+## Backend
+
+### New endpoint: `POST /api/redaction/spot`
+
+**Request:**
+```json
+{
+  "doc_id": "abc123",
+  "page": 1,
+  "click_x": 342,
+  "click_y": 187
+}
+```
+
+Coordinates in page-image pixels (same coordinate space as existing redaction boxes).
+
+**Algorithm:**
+1. Load the rasterized page image (already cached from upload)
+2. Convert to grayscale, binarize with threshold 40 (dark pixels -> white mask, rest -> black)
+3. Run `cv2.connectedComponents()` on the binary mask
+4. Look up which label is at `(click_y, click_x)` in the label matrix
+5. If label is 0 (background) -> return error
+6. Extract all pixels with that label, compute bounding box
+7. If bounding box area < 100px -> return error (noise/speck)
+8. Pass the bounding box into the existing `analyze_redaction()` pipeline
+9. Return the combined response
+
+**Success response:**
+```json
+{
+  "box": {"x": 310, "y": 180, "w": 150, "h": 16},
+  "segments": [...],
+  "gap": {...},
+  "font": {...},
+  "chars": [...],
+  "line": {...},
+  "offset_x": 5.2,
+  "offset_y": 0.0
+}
+```
+
+**Error response:**
+```json
+{"error": "no_redaction_found"}
+```
+
+### Upload changes
+
+Remove `detect_redactions()` call from the upload handler. Upload only rasterizes pages. `GET /api/doc/{id}/page/{p}/data` returns `{redactions: []}` initially. The `detect_redactions.py` module stays in the codebase but is no longer called.
+
+## Frontend
+
+### Double-click handler (replaces zoom)
+
+1. Double-click on canvas triggers spot detection
+2. Convert click from screen space to page-image space (inverse transform for zoom/pan)
+3. If click lands on an existing redaction -> activate it (current behavior), skip spot detection
+4. Send `POST /api/redaction/spot`
+5. While waiting: show loading indicator
+6. On success:
+   - Create redaction entry from returned `box`
+   - Populate `analysis` from response
+   - Set status to `analyzed`, initialize `overrides`
+   - Add to `state.redactions` and sidebar list
+   - Activate the redaction (center canvas, show green overlay, open toolbar)
+7. On error: show toast "No redaction found at click point"
+
+Redaction ID: `"m" + Date.now().toString(36)` (same as former manual marking).
+
+### Input remapping
+
+| Modifier | Old action | New action |
+|----------|-----------|------------|
+| Double-click | Zoom 2x | Spot redaction |
+| Shift+drag | Manual mark redaction | Adjust gap width |
+| Ctrl+drag | Adjust overlay offset | Adjust overlay offset (unchanged) |
+| Ctrl+Shift+drag | Adjust gap width | (removed) |
+| Plain drag | Pan | Pan (unchanged) |
+
+### Sidebar
+
+Starts empty. Redactions appear as user double-clicks them. Sorted top-to-bottom, left-to-right after each addition.