diff --git a/docs/plans/2026-02-17-pixel-alignment-design.md b/docs/plans/2026-02-17-pixel-alignment-design.md
new file mode 100644
index 0000000..8ec0fa5
--- /dev/null
+++ b/docs/plans/2026-02-17-pixel-alignment-design.md
@@ -0,0 +1,31 @@
+# Pixel-Aligned Text Positioning
+
+## Problem
+
+Font detection now correctly identifies the font, but the overlay text position is wrong:
+1. Characters near the redaction edge are cut off ("Sent" becomes "Se") because the char filter uses `c.x + c.w <= rx` which misses chars whose right edge extends past the redaction
+2. `offset_x` is calculated from OCR positions + PIL rendering width, but browser canvas rendering differs — resulting in misaligned overlay
+3. `offset_y = 0` (hardcoded) — no vertical alignment at all
+
+## Solution
+
+### Fix 1: Character Filtering
+
+Change char filtering to use center-point: `c.x + c.w / 2 < rx` instead of `c.x + c.w <= rx`. This captures characters whose center is left of the redaction, recovering "nt" from "Sent". Similarly for right: `c.x + c.w / 2 > rx + rw`.
+
+### Fix 2: Pixel-Based Alignment Function
+
+New `_align_text_to_page()` function:
+1. Render the left text with the detected font onto a canvas
+2. Binarize both rendered text and page image crop
+3. Slide rendered text across the page crop (±20px X, ±10px Y) to find position with maximum Dice overlap
+4. Return `(offset_x, offset_y)` — the pixel-perfect position
+
+### Fix 3: Integration
+
+Replace the OCR-based offset calculation in `_run_analysis()` with the pixel alignment result. The frontend already applies `line.x + offsetX` and `line.y + offsetY` — no frontend changes needed.
+
+## Files Changed
+
+- `unredact/pipeline/font_detect.py` — Add `_align_text_to_page()` function
+- `unredact/app.py` — Fix char filtering, replace offset calculation with pixel alignment