diff --git a/README.md b/README.md
index a2641f8..a66a225 100644
--- a/README.md
+++ b/README.md
@@ -633,6 +633,71 @@ The Accumulation phase block reduces to a single soft
 "Years until possible retirement: none" line; everything else
 behaves like the legacy projection display.
 
+### Realized actuals overlay (`--overlay-actuals`)
+
+Run `zfin projections --as-of <DATE> --overlay-actuals` to plot the
+**realized portfolio trajectory** from `<DATE>` up to today on top of
+the projected percentile bands. Answers the question "how accurate
+were my past projections compared to reality?"
+
+The TUI is the high-fidelity surface — open the projections tab,
+press `d` to set the as-of date, then press `o` to toggle the
+overlay. The CLI prints a tip pointing at the TUI; the braille chart
+itself doesn't render the overlay (the resolution doesn't do justice
+to a 12+ year actuals trajectory laid against percentile bands).
+
+**The overlay reads from two sources, snapshot-precedence:**
+
+1. **Native snapshots** in `<portfolio-dir>/history/*-portfolio.srf`
+   — produced by `zfin snapshot`. Highest fidelity (full lot-level
+   state, exact totals).
+2. **Imported values** in `<portfolio-dir>/history/imported_values.srf`
+   — a hand-maintained back-history file with one `liquid::` total
+   per date. Useful for backfilling a historical record from
+   spreadsheet data, statements, etc. Snapshots win on overlapping
+   dates.
+
+**As-of resolution against either source.** `--as-of <DATE>`
+resolves first against the snapshot directory (nearest-earlier
+`*-portfolio.srf`); if no snapshot exists at or before the date,
+it falls back to `imported_values.srf`. This means you can run
+`zfin projections --as-of 2018-06-01` even with zero snapshot
+files, as long as the imported back-history covers that date.
+
+When the resolution lands on an imported value, the projection
+bands are computed using **today's allocations scaled to the
+imported liquid total** — we can't reconstruct the historical lot
+composition from just a `liquid::` row, so today's stock/bond
+split is substituted as the best-available approximation. The
+header line reads `Projections (as of YYYY-MM-DD, imported value)`
+and a muted note flags the scaling.
+
+**How to read the chart:**
+
+- **Actuals line stays inside the bands** → the model was
+  directionally honest.
+- **Actuals line punches through the top band** → the model was
+  conservative (good problem to have).
+- **Actuals line punches through the bottom band** → the model was
+  optimistic (bands need wider envelope, or a bear scenario was
+  underweighted).
+- **A thin vertical "today" line** marks where the actuals end and
+  the projected future begins.
+
+**Critical caveat (must be loud, by design):**
+
+> This overlay shows whether the model was **directionally honest**,
+> not whether the SWR claim was **accurate**. The SWR claim is a
+> 30-year claim. We have at most ~12 years of weekly history
+> (post-import) and 1+ years of native snapshots. The overlay tells
+> you "did my actual trajectory fall within the bands the model would
+> have drawn." It does **not** tell you "did the safe withdrawal rate
+> hold up over a full retirement horizon." We will not have data to
+> answer that within either of our lifetimes.
+
+The TUI surfaces this caveat on a status line whenever the overlay is
+active.
+
 ### Life events
 
 Life events modify the simulation's annual cash flow. Positive