diff --git a/docs/plans/2026-02-16-interactive-redaction-ui-design.md b/docs/plans/2026-02-16-interactive-redaction-ui-design.md
new file mode 100644
index 0000000..f8777d5
--- /dev/null
+++ b/docs/plans/2026-02-16-interactive-redaction-ui-design.md
@@ -0,0 +1,169 @@
+# Interactive Redaction UI Design
+
+**Date:** 2026-02-16
+**Status:** Approved
+
+## Problem
+
+The current UI does upfront OCR + font detection on every line of every page at upload time. This is slow, often inaccurate, and doesn't fit the app's workflow — users care about redactions, not every line of text. Redaction gaps must be manually marked with Ctrl+Space, which is tedious and error-prone.
+
+## Solution
+
+Shift to a redaction-centric workflow:
+
+1. **Upload:** Rasterize PDF + detect redaction boxes via OpenCV (no OCR)
+2. **Browse:** Clickable redaction indicators on the canvas
+3. **Analyze:** On-demand OCR + font detection when a user clicks a specific redaction
+4. **Solve:** Popover with solver controls anchored to the redaction
+
+## Architecture
+
+### Upload Flow (Simplified)
+
+```
+PDF → rasterize pages → OpenCV detect black rectangles → return redaction bboxes
+```
+
+No OCR. No font detection. Fast.
+
+### OpenCV Redaction Detection
+
+- Grayscale → binary threshold (near-black pixels)
+- Find contours → filter for rectangles (aspect ratio > 2:1, minimum area to reject noise)
+- Merge overlapping/adjacent boxes
+- Return bounding boxes in page-pixel coordinates
+
+### Manual Marking
+
+Users can click-drag on the canvas to draw custom redaction boxes for cases OpenCV misses. These have the same data shape as auto-detected ones.
+
+## Frontend: Canvas-Native Approach
+
+### Redaction Indicators
+
+- Each detected redaction rendered as a semi-transparent colored overlay on the `<canvas>`
+- Hover: cursor → pointer, box gets brighter highlight
+- Hit-testing via inverse-transformed coordinates (accounts for zoom/pan)
+
+### Click → Analyze → Popover
+
+1. User clicks a redaction overlay
+2. Frontend sends `POST /api/redaction/analyze` with `{doc_id, page, redaction: {x, y, w, h}}`
+3. Backend crops the line, runs OCR + font detection on just that line, returns segments + gap
+4. Popover opens anchored near the clicked redaction
+
+### Popover Contents
+
+- Detected line text with redaction gap visually indicated
+- Font name + size (editable)
+- Gap width in px (editable, +/- buttons)
+- Solve button → solver controls (charset, tolerance, mode, filters)
+- Results stream via SSE
+- Click result → preview renders on canvas in the gap
+- Accept button → saves solution
+
+## Frontend: Left Panel Redesign
+
+**Old:** Scrollable list of OCR'd lines with font info.
+
+**New:** Scrollable list of detected redactions.
+
+### Redaction Card Contents
+
+- Thumbnail crop of the redaction area
+- Status indicator: `unanalyzed` → `analyzed` → `solved`
+- Sequential number (e.g. "Redaction #3") + page number
+- Once analyzed: surrounding text snippet (e.g. "...from ████ to the...")
+- Once solved: accepted solution text
+
+### Interactions
+
+- Click card → canvas scrolls/pans to center on redaction, opens popover
+- Accepted solutions render persistently on canvas (green text in the gap)
+
+### State Model
+
+```javascript
+state = {
+  docId,
+  pageCount,
+  currentPage,
+  redactions: {
+    "r1": {x, y, w, h, page, status: "unanalyzed"},
+    "r2": {x, y, w, h, page, status: "analyzed", analysis: {segments, gap, font, chars}, solution: null},
+    "r3": {x, y, w, h, page, status: "solved", analysis: {...}, solution: {text: "Smith", font, size, width_px}},
+  },
+  fonts: [...],
+  zoom, panX, panY,
+  associates,
+}
+```
+
+## Backend: API Changes
+
+### Modified Endpoints
+
+| Endpoint | Change |
+|----------|--------|
+| `POST /api/upload` | Simplified: rasterize + OpenCV detect only |
+| `GET /api/doc/{id}/page/{p}/data` | Returns redaction bboxes instead of OCR lines |
+
+### New Endpoint
+
+**`POST /api/redaction/analyze`**
+
+- Input: `{doc_id, page, redaction: {x, y, w, h}}`
+- Process: crop line from page image → OCR the crop → detect font → build segments
+- Output: `{segments: [{text, x, w}, ...], gap: {x, w}, font: {name, id, size_px}, chars: [...]}`
+
+### Removed Endpoints
+
+| Endpoint | Reason |
+|----------|--------|
+| `GET /api/doc/{id}/page/{p}/overlay` | No more upfront green overlay |
+
+### Unchanged Endpoints
+
+- `POST /api/solve` — solver still takes font/width/charset, streams results
+- `GET /api/fonts`, `GET /api/font/{id}` — font serving
+- Dictionary/associates/email endpoints
+- `GET /api/doc/{id}/page/{p}/original` — page image serving
+
+### Data Storage
+
+```python
+_docs[doc_id] = {
+    "page_count": int,
+    "pages": {
+        page_num: {
+            "image": PIL.Image,
+            "redactions": [{"x": int, "y": int, "w": int, "h": int, "id": str}]
+        }
+    }
+}
+```
+
+## Removals
+
+### Frontend
+
+- Line list rendering + line selection logic
+- Segment editing text inputs (Ctrl+Space splitting)
+- Font controls / text edit bar / d-pad in the right panel
+- Green overlay image display
+- `state.pageData`, `state.selectedLine`, `state.lineOverrides`, `state.activeSegment`
+
+### Backend
+
+- Upfront `ocr_page()` + `detect_fonts()` calls in upload handler
+- `render_overlay()` and overlay endpoint
+- Line/font data from `/api/doc/{id}/page/{p}/data`
+
+### Kept Intact
+
+- All solver logic (Python + Rust)
+- Width table construction
+- Font serving endpoints
+- Dictionary/associates/email endpoints
+- PDF rasterization
+- Core OCR and font detection modules (called on-demand now)