v8: model card with LOAO results + plots

README.md

@@ -14,23 +14,48 @@ language:
 library_name: pytorch
 ---
 
-# Flight-JEPA v7
+# Flight-JEPA v8
 
 A trajectory forecasting model for aircraft on terminal-area approach,
 specialized for the **blindspot continuation** task: given an observed
 past track, predict the trajectory through a coverage gap of variable
 length and the reappearance distribution.
 
-The headline contribution of v7 is a **JEPA-style past-track masked
+The headline contribution is a **JEPA-style past-track masked
 pretraining recipe** that produces representations more robust to test-
-time radar coverage gaps. Pretrained-then-fine-tuned models maintain
-significantly lower forecasting error and better calibrated uncertainty
-when up to half of the past observations are missing — the regime
-aviation deployment cares about.
+time radar coverage gaps, with the gain *generalizing across airports*.
+Pretrained-then-fine-tuned models maintain significantly lower
+forecasting error and better-calibrated uncertainty when up to 70% of
+past observations are missing — including on **completely held-out
+airports the fine-tuning never saw**.
 
-## Quick numbers
+## v8 — leave-one-airport-out (LOAO) headline
 
-On RKSIa (Incheon arrivals, 8092 test trajectories, n=3 seeds):
+Across 4 LOAO folds (held out: RKSIa / RKSId / ESSA / LSZH, n=3 seeds
+each), pretrained beats scratch with significance:
+
+| Past-track dropout | Mean Δ FDE | p |
+|---|---:|---:|
+| 0% (clean — no regression) | +2.8% | 0.41 |
+| 30% | −6.6% | 0.04 ✓ |
+| **50%** | **−23.4%** | **<0.001 ✓** |
+| **70%** | **−22.5%** | **<0.001 ✓** |
+
+11 of 12 comparisons at ≥30% dropout reach p<0.05. All 4 LOAO folds
+pass the locked criterion independently.
+
+![v8 summary](plots/summary_v8.png)
+![v8 FDE per airport](plots/fde_per_airport.png)
+![v8 coverage per airport](plots/coverage_per_airport.png)
+
+The result generalizes across very different airports (Korea, Sweden,
+Switzerland — different runway geometries and procedures). It uses an
+airport-ID token (UniTraj recipe, arxiv:2403.15098) for conditioning.
+
+## v7 — single-airport reference
+
+The v7 prerequisite (single-airport, RKSIa-only) showed the same effect
+on a within-airport split:
 
 | Past-track dropout | Scratch FDE | Pretrained FDE | Δ | p (Welch) |
 |---|---:|---:|---:|---:|