diff --git a/docs/runbook_propagation_pipeline.md b/docs/runbook_propagation_pipeline.md
index d4e6acbc..e6490d8e 100644
--- a/docs/runbook_propagation_pipeline.md
+++ b/docs/runbook_propagation_pipeline.md
@@ -18,8 +18,10 @@ prop-grid-rs (worker nodes)      Postgres (10.0.15.30)            prop pods (nod
 1. Rust completes `{band, valid_time}` compute and writes the `.prop`
    file via atomic rename. (Readers also accept legacy `.ntms` files
    with `NTMS` magic during a rolling deploy.)
-2. Rust emits `NOTIFY propagation_ready '<iso valid_time>'`.
-3. `NotifyListener` on each Elixir pod receives it, prunes
+2. Rust emits `NOTIFY propagation_ready '<run_time_iso>|<valid_time_iso>'`
+   (pipe-delimited: run_time first, valid_time second — needed for
+   `retain_scores_window` which uses run_time as the retention anchor).
+3. `NotifyListener` on each Elixir pod splits the payload, prunes
    ScoreCache entries outside the active forecast window, and
    broadcasts `propagation_updated` over PubSub.
 4. LiveView clients see the broadcast and re-request scores. The
@@ -91,6 +93,19 @@ to drain. All fetch/decode/score helpers moved to
 `rust/prop_grid_rs/src/pipeline.rs`; the Elixir module is retained
 only as the cron entry point, not as a fallback compute path.
 
+## On-disk file formats
+
+Score files live under `/data/scores/<band_mhz>/<valid_time_iso>.<ext>`:
+
+| Extension | Source | Writer | Description |
+|-----------|--------|--------|-------------|
+| `.prop` | HRRR (f00–f48) | `prop-grid-rs` (Rust) | Primary CONUS scores, 0.125° grid, 92k cells per band |
+| `.hrdps.prop` | HRDPS (Canadian) | `hrdps_grid_worker` (Elixir) | Canadian coverage, merged at read with HRRR > HRDPS priority |
+| `.gefs.prop` | GEFS (f024–f168) | `gefs_fetch_worker` (Elixir) | Extended-range ensemble mean, merged at read with HRRR > HRDPS > GEFS priority |
+| `.ntms` | Legacy | (deprecated) | Old format, still accepted by readers during rolling deploys |
+
+All files use atomic rename on write (write to `.tmp.<uuid>`, then `rename(2)`).
+
 ## Monitoring signals
 
 - `avg_over_time(beam_memory_total_bytes[1h])` on prop pods — steady
diff --git a/prediction.md b/prediction.md
index 6bc131fa..b806324a 100644
--- a/prediction.md
+++ b/prediction.md
@@ -1,5 +1,7 @@
 # Propagation Prediction System Design
 
+> **Status (2026-08-01): All three phases are implemented.** The Rust pipeline (`prop-grid-rs`) fetches HRRR f00–f48 hourly, scores all 22 bands across the full CONUS grid, and writes per-valid_time `.prop` files. The map timeline shows the next 48 hours of real NWP-forecast scores (not persistence). `Propagation.point_forecast/3` provides the per-point sparkline. See `docs/runbook_propagation_pipeline.md` for operations; `scripts/validate_forecast.py` measures forecast skill by lead time.
+
 ## The Opportunity
 
 No tool exists for microwave operators to answer: **"When should I get on the air in the next 24 hours?"**