Switch to the 2026-10-25 CAS-Canglong SST model: 12-month mean, spread and quintile probabilities; held-out skill and hindcast tab; Oct 2026 IRI plume; atmosphere 2026-10-22

.gitattributes

@@ -93,3 +93,29 @@ assets/maps/ssta_t15.png filter=lfs diff=lfs merge=lfs -text
 assets/maps/ssta_t16.png filter=lfs diff=lfs merge=lfs -text
 data/canglong_atmos_start20261008.nc filter=lfs diff=lfs merge=lfs -text
 data/canglong_sst_init202609.nc filter=lfs diff=lfs merge=lfs -text
+assets/maps/pupper_t01.png filter=lfs diff=lfs merge=lfs -text
+assets/maps/pupper_t02.png filter=lfs diff=lfs merge=lfs -text
+assets/maps/pupper_t03.png filter=lfs diff=lfs merge=lfs -text
+assets/maps/pupper_t04.png filter=lfs diff=lfs merge=lfs -text
+assets/maps/pupper_t05.png filter=lfs diff=lfs merge=lfs -text
+assets/maps/pupper_t06.png filter=lfs diff=lfs merge=lfs -text
+assets/maps/pupper_t07.png filter=lfs diff=lfs merge=lfs -text
+assets/maps/pupper_t08.png filter=lfs diff=lfs merge=lfs -text
+assets/maps/pupper_t09.png filter=lfs diff=lfs merge=lfs -text
+assets/maps/pupper_t10.png filter=lfs diff=lfs merge=lfs -text
+assets/maps/pupper_t11.png filter=lfs diff=lfs merge=lfs -text
+assets/maps/pupper_t12.png filter=lfs diff=lfs merge=lfs -text
+assets/maps/sigma_t01.png filter=lfs diff=lfs merge=lfs -text
+assets/maps/sigma_t02.png filter=lfs diff=lfs merge=lfs -text
+assets/maps/sigma_t03.png filter=lfs diff=lfs merge=lfs -text
+assets/maps/sigma_t04.png filter=lfs diff=lfs merge=lfs -text
+assets/maps/sigma_t05.png filter=lfs diff=lfs merge=lfs -text
+assets/maps/sigma_t06.png filter=lfs diff=lfs merge=lfs -text
+assets/maps/sigma_t07.png filter=lfs diff=lfs merge=lfs -text
+assets/maps/sigma_t08.png filter=lfs diff=lfs merge=lfs -text
+assets/maps/sigma_t09.png filter=lfs diff=lfs merge=lfs -text
+assets/maps/sigma_t10.png filter=lfs diff=lfs merge=lfs -text
+assets/maps/sigma_t11.png filter=lfs diff=lfs merge=lfs -text
+assets/maps/sigma_t12.png filter=lfs diff=lfs merge=lfs -text
+data/canglong_atmos_start20261022.nc filter=lfs diff=lfs merge=lfs -text
+data/canglong_sst_round5_init202609.nc filter=lfs diff=lfs merge=lfs -text

README.md

@@ -9,7 +9,11 @@ python_version: "3.12"
 app_file: app.py
 pinned: true
 license: mit
-short_description: Live 16-month global SST and ENSO forecast from CAS-Canglong
+short_description: 12-month global SST & ENSO outlook with probabilities
+models:
+  - LonghaoWang/CAS-Canglong-SST
+datasets:
+  - LonghaoWang/SST-Hindcast-2016-2022-1deg-lead1-12
 tags:
   - weather
   - climate
@@ -22,19 +26,21 @@ tags:
 
 # 🐉 CAS-Canglong · global SST & ENSO forecast
 
-The current operational run of **CAS-Canglong**, a 3-D Earth-specific transformer for sub-seasonal to seasonal prediction
-of global sea-surface temperature ([*Advanced Science* 2026](https://doi.org/10.1002/advs.78271) ·
-[code](https://github.com/GISWLH/CAS-Canglong) · live portal [canglong.earth](https://canglong.earth)).
+The latest run of the **2026-10-25 release of the CAS-Canglong 1° SST model**
+([weights and code](https://huggingface.co/LonghaoWang/CAS-Canglong-SST) ·
+[2016–2022 hindcast](https://huggingface.co/datasets/LonghaoWang/SST-Hindcast-2016-2022-1deg-lead1-12) ·
+[paper, *Advanced Science* 2026](https://doi.org/10.1002/advs.78271) · [code](https://github.com/GISWLH/CAS-Canglong) ·
+live portal [canglong.earth](https://canglong.earth)).
 
-* **Ocean & ENSO** — 16 monthly SST and SST-anomaly maps (1° model grid and 0.25° refined field) and the Niño 3.4
-  outlook against the IRI ENSO plume.
+* **Ocean & ENSO** — 12 monthly maps of SST, SST anomaly, the chance of a top-quintile month and the forecast spread, and
+  the Niño 3.4 outlook against the IRI ENSO plume.
 * **Regional outlook** — Niño regions, the Indian Ocean Dipole, the tropical North Atlantic, the Kuroshio, the
-  Mediterranean and more, month by month.
+  Mediterranean and more, with the last 12 observed months.
+* **Skill & hindcasts** — a held-out test on 2023–2026 targets and the 2016–2022 Niño 3.4 hindcast, lead by lead.
 * **Atmosphere · 6 weeks (experimental)** — weekly 2 m temperature, precipitation, sea-level pressure and 500 hPa height
   from the CAS-Canglong atmospheric run, read with [WeatherPre](https://github.com/GISWLH/WeatherPre).
-* **Data** — the NetCDF and CSV files behind every map.
+* **Model & data** — the NetCDF and CSV files behind every map, and code to load the model and the hindcast.
 
-The fields are identical to those on canglong.earth. They come from re-running the frozen checkpoints on a GPU
-workstation; the Space only displays them. Assets are rebuilt with
-[`hf_space/scripts/build_assets.py`](https://github.com/GISWLH/CAS-Canglong/tree/main/hf_space).
-This is a research forecast, not an official warning product; see the *About* tab for limitations.
+The Space only displays precomputed fields; they come from the frozen model run on a GPU workstation and are rebuilt with
+[`hf_space/scripts/build_assets.py`](https://github.com/GISWLH/CAS-Canglong/tree/main/hf_space). This is a research
+forecast, not an official warning product; see the *About* tab for its limits, in particular for ENSO.

app.py

@@ -1,6 +1,6 @@
 """CAS-Canglong forecast showcase (Hugging Face Space).
 
-Everything shown is read from `data/` and `assets/`, written by `scripts/build_assets.py` from one forecast run.
+Everything shown is read from `data/` and `assets/`, written by `scripts/build_assets.py` from one forecast release.
 """
 from __future__ import annotations
 
@@ -18,16 +18,22 @@ DATA, MAPS = ROOT / "data", ROOT / "assets" / "maps"
 RUN = json.loads((DATA / "run.json").read_text())
 SST = xr.open_dataset(DATA / RUN["sst_file"]).load()
 IRI = json.loads((DATA / "iri_plume.json").read_text())
+HINDCAST = pd.read_csv(DATA / "hindcast_nino34.csv", parse_dates=["target_month"])
+HINDCAST_SKILL = pd.read_csv(DATA / "hindcast_nino34_skill.csv")
+HELDOUT = pd.read_csv(DATA / "round6_heldout_metrics.csv")
+MODEL_REPO, HINDCAST_REPO = RUN["model_repo"], RUN["hindcast_dataset"]
 
 SURFACE, GRIDLINE, BASELINE, INK, INK2, MUTED = "#fcfcfb", "#e1e0d9", "#c3c2b7", "#0b0b0b", "#52514e", "#898781"
 CAS_BLUE, DYN_ORANGE, STAT_AQUA, CONTEXT = "#2a78d6", "#eb6834", "#1baf7a", "#c3c2b7"
 FONT = "system-ui, -apple-system, Segoe UI, sans-serif"
 
-FRAMES = RUN["frames"]                       # index 0 = observed initial month, 1..16 = forecast leads
+FRAMES = RUN["frames"]                       # index 0 = observed last input month, 1..12 = forecast leads
 WEEKS = RUN["weeks"]
 MONTHS = pd.to_datetime(SST.time.values)
 OBS_MONTH = pd.Timestamp(SST.obs_time.values[-1])
-NINO = SST.nino34.values.astype(float)
+HIST = pd.to_datetime(SST.hist_time.values)  # last 12 observed months, ending with OBS_MONTH
+NINO = SST.nino34_9120.values.astype(float)
+NINO_HIST = SST.nino34_history_9120.values.astype(float)
 
 # lat/lon boxes on the 0..360 grid; DMI is west minus east
 REGIONS = {
@@ -44,6 +50,18 @@ REGIONS = {
     "Mediterranean (30°N–46°N, 0°–36°E)": [(30, 46, 0, 36)],
     "Global ocean (60°S–60°N)": [(-60, 60, 0, 360)],
 }
+FIELDS = {"SST anomaly": "ssta", "SST": "sst", "Chance of a top-quintile month": "pupper", "Forecast uncertainty (σ)": "sigma"}
+HELDOUT_TRACKS = {
+    "12 common start months, Jul 2023–Jun 2024 (8 models)": "all_available_models",
+    "25 common start months, Jul 2023–Jul 2025 (7 models)": "longer_common_period",
+    "34 fixed target months, Dec 2023–Sep 2026 (own models only)": "own_fixed_target_months",
+}
+HELDOUT_REGIONS = {"Global": "Global", "Niño 3.4": "Nino3.4", "Tropical Indo-Pacific (20°S–20°N)": "WP",
+                   "North Pacific (20°N–70°N)": "PDO", "Arctic (70°N–90°N)": "AO", "South Indian (60°S–40°S)": "SIO",
+                   "South Pacific (60°S–40°S)": "SPO", "South Atlantic (60°S–40°S)": "SAO"}
+HELDOUT_LABELS = {"Cutoff-2022 model": "CAS-Canglong method, refit with data to 2022",
+                  "Seasonal baseline": "frozen seasonal model (before calibration)",
+                  "Original baseline": "original head", "Surface baseline": "surface head", "Joint baseline": "joint head"}
 
 
 def box_mean(field: xr.DataArray, boxes) -> np.ndarray:
@@ -56,7 +74,8 @@ def box_mean(field: xr.DataArray, boxes) -> np.ndarray:
 
 
 def region_series(name: str) -> np.ndarray:
-    fields = xr.concat([SST.observed_ssta.expand_dims(time=[np.datetime64(OBS_MONTH, "ns")]), SST.ssta], "time")
+    """12 observed months (ending with the last input month) followed by the 12 forecast months, 1991-2020 base."""
+    fields = xr.concat([SST.history_ssta_9120.rename(hist_time="t"), SST.ssta_9120.rename(time="t")], "t")
     spec = REGIONS[name]
     if spec == "dmi":
         return box_mean(fields, [(-10, 10, 50, 70)]) - box_mean(fields, [(-10, 0, 90, 110)])
@@ -89,18 +108,8 @@ def _center(season: str) -> pd.Timestamp | None:
 
 def plume_figure() -> go.Figure:
     fig = go.Figure()
-    obs = [(_center(o["season"]), o["data"], o["label"]) for o in IRI["observed"]]
-    iri_anchor = max(obs, key=lambda o: o[0])
-
-    def connect_observation(anchor, x, y, color, width, group):
-        # A visual bridge to the first forecast; no extra forecast values are introduced.
-        fig.add_trace(go.Scatter(x=[anchor[0], x[0]], y=[anchor[1], y[0]], mode="lines",
-                                 line=dict(color=color, width=width, dash="dot"),
-                                 legendgroup=group, showlegend=False, hoverinfo="skip"))
-
     for k, m in enumerate(IRI["models"]):
         x = [_center(s) for s in m["seasons"]]
-        connect_observation(iri_anchor, x, m["data"], CONTEXT, 1, "iri")
         fig.add_trace(go.Scatter(x=x, y=m["data"], mode="lines", line=dict(color=CONTEXT, width=1),
                                  name="IRI plume models", legendgroup="iri", showlegend=k == 0, opacity=0.9,
                                  hovertemplate=f"{m['model']} ({m['type'].lower()}): %{{y:+.2f}} °C<extra></extra>"))
@@ -109,26 +118,20 @@ def plume_figure() -> go.Figure:
         if avg["model"] not in colors: continue
         x = [_center(s) for s in avg["seasons"]]
         name = f"IRI {avg['model']} mean"
-        connect_observation(iri_anchor, x, avg["data"], colors[avg["model"]], 2, name)
-        fig.add_trace(go.Scatter(x=x, y=avg["data"], mode="lines+markers", name=name, legendgroup=name,
+        fig.add_trace(go.Scatter(x=x, y=avg["data"], mode="lines+markers", name=name,
                                  line=dict(color=colors[avg["model"]], width=2), marker=dict(size=8, line=dict(color=SURFACE, width=2)),
                                  hovertemplate=f"{name}: %{{y:+.2f}} °C<extra></extra>"))
-    run = pd.Series(NINO).rolling(3, center=True).mean().values[1:-1]
+    fig.add_trace(go.Scatter(x=HIST, y=NINO_HIST, mode="lines", name="ERA5 observed (monthly)",
+                             line=dict(color=INK, width=1.5), hovertemplate="ERA5 %{x|%b %Y}: %{y:+.2f} °C<extra></extra>"))
+    run = pd.Series(NINO).rolling(3, center=True).mean().values[1:-1]   # forecast-only seasons, OND .. JAS
     xs = MONTHS[1:-1]
-    cas_anchor = (OBS_MONTH, float(SST.observed_nino34.values[-1]))
-    connect_observation(cas_anchor, xs, run, CAS_BLUE, 3, "cas")
     fig.add_trace(go.Scatter(x=xs, y=run, mode="lines+markers", name="CAS-Canglong (3-month mean)",
-                             legendgroup="cas",
                              line=dict(color=CAS_BLUE, width=3), marker=dict(size=8, line=dict(color=SURFACE, width=2)),
                              hovertemplate="CAS-Canglong: %{y:+.2f} °C<extra></extra>"))
     fig.add_annotation(x=xs[-1], y=run[-1], text="CAS-Canglong", xanchor="left", xshift=8, showarrow=False,
                        font=dict(color=INK, size=12))
-    fig.add_trace(go.Scatter(x=[cas_anchor[0]], y=[cas_anchor[1]], mode="markers",
-                             name=f"Observed (ERA5, {OBS_MONTH:%b %Y})",
-                             marker=dict(color=CAS_BLUE, size=10, symbol="diamond", line=dict(color=SURFACE, width=2)),
-                             hovertemplate="ERA5 observed %{x|%b %Y}: %{y:+.2f} °C<extra></extra>"))
-    fig.add_trace(go.Scatter(x=[o[0] for o in obs], y=[o[1] for o in obs], mode="lines+markers", name="Observed (IRI)",
-                             line=dict(color=INK, width=2),
+    obs = [(_center(o["season"]), o["data"], o["label"]) for o in IRI["observed"]]
+    fig.add_trace(go.Scatter(x=[o[0] for o in obs], y=[o[1] for o in obs], mode="markers", name="Observed (IRI)",
                              marker=dict(color=INK, size=9, symbol="diamond", line=dict(color=SURFACE, width=2)),
                              text=[o[2] for o in obs], hovertemplate="%{text}: %{y:+.2f} °C<extra></extra>"))
     for y, label in ((0.5, "El Niño +0.5 °C"), (-0.5, "La Niña −0.5 °C")):
@@ -136,38 +139,48 @@ def plume_figure() -> go.Figure:
         fig.add_annotation(x=0, xref="paper", y=y, text=label, xanchor="left", yanchor="bottom", xshift=4,
                            showarrow=False, font=dict(color=MUTED, size=11))
     fig.update_xaxes(dtick="M2", tickformat="%b<br>%Y", tickangle=0)
-    return _layout(fig, "Niño 3.4 SST anomaly (°C)", 460)
+    return _layout(fig, "Niño 3.4 SST anomaly (°C, 1991–2020 base)", 460)
 
 
 def region_figure(name: str) -> go.Figure:
     y = region_series(name)
-    x = [OBS_MONTH, *MONTHS]
+    x = [*HIST, *MONTHS]
+    n = len(HIST)
     fig = go.Figure()
-    fig.add_trace(go.Scatter(x=x, y=y, mode="lines+markers", name="CAS-Canglong forecast",
-                             line=dict(color=CAS_BLUE, width=2),
-                             marker=dict(size=[0] + [8] * len(MONTHS), line=dict(color=SURFACE, width=2)),
-                             text=["ERA5 observed"] + ["CAS-Canglong forecast"] * len(MONTHS),
-                             hovertemplate="%{text} %{x|%b %Y}: %{y:+.2f} °C<extra></extra>"))
-    fig.add_trace(go.Scatter(x=x[:1], y=y[:1], mode="markers", name=f"ERA5 observed {OBS_MONTH:%b %Y}",
-                             marker=dict(color=INK, size=10, symbol="diamond", line=dict(color=SURFACE, width=2)),
+    fig.add_trace(go.Scatter(x=x[:n], y=y[:n], mode="lines+markers", name="ERA5 observed",
+                             line=dict(color=INK, width=1.5), marker=dict(size=6),
                              hovertemplate="observed %{x|%b %Y}: %{y:+.2f} °C<extra></extra>"))
+    fig.add_trace(go.Scatter(x=x[n - 1:n + 1], y=y[n - 1:n + 1], mode="lines", line=dict(color=CAS_BLUE, width=2, dash="dot"),
+                             showlegend=False, hoverinfo="skip"))
+    fig.add_trace(go.Scatter(x=x[n:], y=y[n:], mode="lines+markers", name="CAS-Canglong forecast",
+                             line=dict(color=CAS_BLUE, width=2), marker=dict(size=8, line=dict(color=SURFACE, width=2)),
+                             hovertemplate="%{x|%b %Y}: %{y:+.2f} °C<extra></extra>"))
     if name.startswith("Niño"):
         for v in (0.5, -0.5):
             fig.add_hline(y=v, line=dict(color=MUTED, width=1, dash="dot"))
     fig.update_xaxes(dtick="M2", tickformat="%b<br>%Y", tickangle=0)
-    return _layout(fig, "SST anomaly (°C)", 400)
+    return _layout(fig, "SST anomaly (°C, 1991–2020 base)", 400)
 
 
 def region_table(name: str) -> pd.DataFrame:
-    y = region_series(name)
+    y = region_series(name)[len(HIST) - 1:]
     labels = [f"{OBS_MONTH:%b %Y} (observed)"] + [f"{t:%b %Y}" for t in MONTHS]
     return pd.DataFrame({"Month": labels, "Lead (months)": ["0"] + [str(i) for i in range(1, len(MONTHS) + 1)],
                          "SST anomaly (°C)": [f"{v:+.2f}" for v in y]})
 
 
 def map_path(field: str, index: int) -> str:
-    key = FRAMES[int(index)]["key"]
-    return str(MAPS / f"{'ssta' if field.startswith('SST anomaly') else 'sst'}_{key}.png")
+    code, i = FIELDS[field], int(index)
+    if code in ("pupper", "sigma"):
+        i = max(i, 1)                           # probability and spread exist for forecast months only
+    return str(MAPS / f"{code}_{FRAMES[i]['key']}.png")
+
+
+def frame_label(field: str, index: int) -> str:
+    i = int(index)
+    if FIELDS[field] in ("pupper", "sigma") and i == 0:
+        return f"**{FRAMES[1]['label']}** — probability and spread maps start at lead 1"
+    return f"**{FRAMES[i]['label']}**"
 
 
 def atmos_path(var: str, week: int) -> str:
@@ -176,80 +189,152 @@ def atmos_path(var: str, week: int) -> str:
     return str(MAPS / f"atm_{code}_w{int(week)}.png")
 
 
-def frame_label(index: int) -> str:
-    return f"**{FRAMES[int(index)]['label']}**"
-
-
 def week_label(week: int) -> str:
     return f"**{WEEKS[int(week) - 1]['label']}**"
 
 
-above = [t for t, v in zip(MONTHS, NINO) if v > 0.5]
-peak_i = int(np.argmax(NINO))
+def heldout_figure(track: str, region: str, metric: str) -> go.Figure:
+    col = "acc_spatial" if metric.startswith("Spatial ACC") else "rmse_K"
+    d = HELDOUT[(HELDOUT.track == HELDOUT_TRACKS[track]) & (HELDOUT.region == HELDOUT_REGIONS[region])]
+    fig = go.Figure()
+    greys = iter(["#898781", "#a9a79f", "#c3c2b7", "#d9d8d2"])
+    nwp = iter([DYN_ORANGE, "#b8421b", "#f3a5a1"])
+    for model, g in sorted(d.groupby("model"), key=lambda kv: (kv[0] != "Cutoff-2022 model", "baseline" not in kv[0], kv[0])):
+        g = g.sort_values("lead")
+        main = model == "Cutoff-2022 model"
+        color = CAS_BLUE if main else next(greys) if "baseline" in model else next(nwp)
+        fig.add_trace(go.Scatter(x=g.lead, y=g[col], mode="lines+markers", name=HELDOUT_LABELS.get(model, f"{model} (raw archive)"),
+                                 line=dict(color=color, width=3 if main else 1.6, dash=None if main or "baseline" in model else "dash"),
+                                 marker=dict(size=7 if main else 5),
+                                 hovertemplate=f"{model} · lead %{{x}}: %{{y:.3f}}<extra></extra>"))
+    fig.update_xaxes(title="Lead (months)", dtick=1)
+    fig = _layout(fig, "Spatial ACC" if col == "acc_spatial" else "RMSE (°C)", 480)
+    fig.update_layout(hovermode="closest", margin=dict(r=40, b=110), legend=dict(y=-0.2))
+    return fig
+
+
+def heldout_table(track: str, region: str) -> pd.DataFrame:
+    d = HELDOUT[(HELDOUT.track == HELDOUT_TRACKS[track]) & (HELDOUT.region == HELDOUT_REGIONS[region])]
+    t = d.groupby("model").agg(acc=("acc_spatial", "mean"), rmse=("rmse_K", "mean"), months=("n_months", "first"),
+                               first=("first_valid", "min"), last=("last_valid", "max")).sort_values("acc", ascending=False)
+    return pd.DataFrame({"Model": [HELDOUT_LABELS.get(m, f"{m} (raw archive)") for m in t.index],
+                         "Mean spatial ACC (leads 1–12)": t.acc.round(3).values,
+                         "Mean RMSE °C (leads 1–12)": t.rmse.round(3).values, "Samples per lead": t.months.values,
+                         "Target months": [f"{a} – {b}" for a, b in zip(t["first"], t["last"])]})
+
+
+def hindcast_figure(lead: int) -> go.Figure:
+    fig = go.Figure()
+    fig.add_trace(go.Scatter(x=HINDCAST.target_month, y=HINDCAST.obs, mode="lines", name="ERA5 observed",
+                             line=dict(color=INK, width=2), hovertemplate="observed %{x|%b %Y}: %{y:+.2f} °C<extra></extra>"))
+    fig.add_trace(go.Scatter(x=HINDCAST.target_month, y=HINDCAST[f"lead{int(lead)}"], mode="lines",
+                             name=f"CAS-Canglong hindcast, lead {int(lead)}", line=dict(color=CAS_BLUE, width=2),
+                             hovertemplate="hindcast %{x|%b %Y}: %{y:+.2f} °C<extra></extra>"))
+    for v in (0.5, -0.5):
+        fig.add_hline(y=v, line=dict(color=MUTED, width=1, dash="dot"))
+    fig.update_xaxes(dtick="M6", tickformat="%b<br>%Y", tickangle=0)
+    return _layout(fig, "Niño 3.4 anomaly (°C, 1959–2009 base)", 400)
+
+
+def hindcast_caption(lead: int) -> str:
+    r = HINDCAST_SKILL[HINDCAST_SKILL.lead == int(lead)].iloc[0]
+    return (f"Lead {int(lead)}: correlation **{r['corr']:.2f}**, RMSE **{r['rmse_K']:.2f} °C** over {int(r['n'])} target months "
+            "(2016–2022). These targets were used to calibrate the released model, so the scores are in-sample.")
+
+
+above = [t for t, v in zip(MONTHS, NINO) if v >= 0.5]
 glob = region_series("Global ocean (60°S–60°N)")
 KPIS = [
-    ("Niño 3.4 · " + f"{MONTHS[0]:%b %Y}", f"{NINO[0]:+.2f} °C", "first forecast month"),
-    ("Niño 3.4 · " + f"{MONTHS[3]:%b %Y}", f"{NINO[3]:+.2f} °C", "boreal winter, 4-month lead"),
-    ("El Niño (> +0.5 °C)", f"{above[0]:%b %Y} – {above[-1]:%b %Y}" if above else "–", f"{len(above)} of {len(NINO)} forecast months"),
-    ("Global ocean 60°S–60°N", f"{glob[1]:+.2f} °C", f"area-mean anomaly, {MONTHS[0]:%b %Y}"),
+    (f"Niño 3.4 · {OBS_MONTH:%b %Y}", f"{NINO_HIST[-1]:+.2f} °C", "observed (ERA5), last input month"),
+    (f"Niño 3.4 · {MONTHS[0]:%b %Y}", f"{NINO[0]:+.2f} °C", "forecast, 1-month lead"),
+    (f"Niño 3.4 · {MONTHS[3]:%b %Y}", f"{NINO[3]:+.2f} °C", "forecast, 4-month lead"),
+    ("El Niño (≥ +0.5 °C)", f"{above[0]:%b %Y} – {above[-1]:%b %Y}" if above else "–", f"{len(above)} of {len(NINO)} forecast months"),
+    ("Global ocean 60°S–60°N", f"{glob[len(HIST)]:+.2f} °C", f"area-mean anomaly, {MONTHS[0]:%b %Y}"),
 ]
 kpi_html = "".join(f'<div class="kpi"><div class="kpi-label">{a}</div><div class="kpi-value">{b}</div>'
                    f'<div class="kpi-note">{c}</div></div>' for a, b, c in KPIS)
 
 HEADER = f"""
 <div class="hero">
-  <div class="hero-kicker">CAS-Canglong · live research forecast</div>
+  <div class="hero-kicker">CAS-Canglong · research forecast · model release 2026-10-25</div>
   <h1>🐉 Global sea-surface temperature &amp; ENSO outlook</h1>
-  <p>Input: ERA5 monthly means through <b>{OBS_MONTH:%B %Y}</b> → forecasts for <b>{MONTHS[0]:%b %Y} – {MONTHS[-1]:%b %Y}</b>
-  (16 months) · six-week atmosphere from <b>{RUN['atmos_start']}</b></p>
+  <p>Input: ERA5 monthly means through <b>{OBS_MONTH:%B %Y}</b> → mean, spread and calibrated quintile probabilities for
+  <b>{MONTHS[0]:%b %Y} – {MONTHS[-1]:%b %Y}</b> (12 months) · six-week atmosphere from <b>{RUN['atmos_start']}</b></p>
   <div class="chips">
+    <a href="https://huggingface.co/{MODEL_REPO}" target="_blank">🧠 Model weights</a>
+    <a href="https://huggingface.co/datasets/{HINDCAST_REPO}" target="_blank">🗂️ Hindcast 2016–2022</a>
     <a href="https://doi.org/10.1002/advs.78271" target="_blank">📄 Paper · Advanced Science</a>
     <a href="https://github.com/GISWLH/CAS-Canglong" target="_blank">⭐ GitHub</a>
     <a href="https://canglong.earth" target="_blank">🌍 canglong.earth</a>
-    <a href="https://github.com/GISWLH/WeatherAI" target="_blank">🧩 WeatherAI (PyTorch port)</a>
     <a href="https://github.com/GISWLH/WeatherPre" target="_blank">📊 WeatherPre</a>
   </div>
 </div>
 <div class="kpis">{kpi_html}</div>
 """
 
+HELDOUT_NOTE = """
+**What was held out.** These curves come from a copy of the released method that was re-fitted with targets up to
+December 2022 only (calibration settings chosen on 2021–2022) and then scored on 2023-01 to 2026-09 targets it never saw.
+The four base heads were trained on targets up to August 2009 and validated up to April 2016. The published weights themselves were calibrated with
+targets through September 2026, so they have no held-out score of their own. The method design was shaped by earlier
+looks at the post-2022 period, so this is a retrospective out-of-time check, not a pristine blind test.
+Spatial ACC is the area-weighted anomaly correlation of each map (not the grid-point temporal correlation). The three
+dynamical systems are raw ensemble-mean archives with no drift or bias correction, which inflates their RMSE. 12–34
+monthly samples per lead are serially correlated; no significance test was made. **Niño 3.4 is the weak spot:** there the
+refit's ACC is below that of its own uncalibrated seasonal model at every lead.
+"""
+
 ABOUT = f"""
 ### What this is
-CAS-Canglong is a 3-D Earth-specific transformer for sub-seasonal to seasonal prediction of global sea-surface
-temperature (SST) ([Wang et al., *Advanced Science* 2026](https://doi.org/10.1002/advs.78271);
-preprint [arXiv:2409.05369](https://arxiv.org/abs/2409.05369)). This Space shows the **current operational run**, the same
-forecast that [canglong.earth](https://canglong.earth) serves:
-
-* **Ocean (16 months)** — a one-pass Swin-transformer SST model on a 1° grid reads 16 months of eight ERA5 monthly
-  fields (SST, MSLP, latent heat flux, 10 m winds, 500/850 hPa geopotential, net solar radiation) and predicts the next
-  16 monthly SST fields in one forward pass; a learned refiner adds 0.25° detail without changing the 1° block means.
+CAS-Canglong learns the coupled ocean–atmosphere system from ERA5 to predict global sea-surface temperature (SST)
+([Wang et al., *Advanced Science* 2026](https://doi.org/10.1002/advs.78271); preprint
+[arXiv:2409.05369](https://arxiv.org/abs/2409.05369)). This Space shows the **2026-10-25 release of the 1° monthly SST
+model** ([weights and code on Hugging Face](https://huggingface.co/{MODEL_REPO})):
+
+* **Inputs** — the last 120 months of eight ERA5 monthly fields on a 1° grid: SST, mean sea-level pressure, surface latent
+  heat flux, 10 m u and v wind, 500 and 850 hPa geopotential, net surface solar radiation.
+* **Core** — four frozen heads built on a 1° CAS-Canglong network: the original network/EOF blend, a slow-background
+  branch, a joint ocean–atmosphere branch and a seasonal-state branch whose response depends on the calendar month.
+* **Calibration** — a per-grid-cell ridge correction (α = 0.05, 19 coefficients per lead and cell, using the four head
+  forecasts, the observed SST anomaly of the last 1, 3 and 12 months, annual/semi-annual harmonics, a time trend and a
+  few seasonal interactions),
+  mixed 10 % with 90 % of the seasonal model; a CRPS-fitted Gaussian spread; and quintile probabilities calibrated in the
+  spirit of the probability-based correction of [Guan et al., *Nature* 2026](https://www.nature.com/articles/s41586-026-11126-x).
+* **Output** — 12 monthly leads on 40,701 ocean cells (nominal 1°): mean SST, Gaussian σ and quantiles, and the
+  probability of each fifth of the 1959–2009 climate. Leads 13–16 are not produced.
 * **Atmosphere (6 weeks, experimental)** — CAS-Canglong V3.5 weekly means published on canglong.earth, read through
   [WeatherPre](https://github.com/GISWLH/WeatherPre) (`wp.forecast("canglong", "t2m", "6w")`).
 
 ### How to read the maps
-* **SST anomaly** maps are the native 1° model output minus the ERA5 1959–2016 monthly climatology (the model's training
-  period, the same reference as its Niño 3.4 index). On the 1991–2020 base that IRI uses, Niño 3.4 values would be
-  about 0.1 °C lower (0.04–0.18 °C, depending on the month).
-* **SST** maps are the 0.25° field. Within about 2° of land it is blended with the latest ERA5 SST for display (8 grid
-  rings replaced, 6 rings faded), exactly as on canglong.earth. That coastal belt is a visual repair, not model skill.
-* The **ENSO plume** compares CAS-Canglong's 3-month running mean with the
-  [IRI ENSO plume]({IRI['source']}) published {IRI['published']}. The plume is redrawn from IRI's own figure.
+* **SST anomaly** maps and all regional and Niño 3.4 series here use the WMO 1991–2020 normal, computed from the same
+  ERA5 record as the inputs, so they can be compared with the [IRI ENSO plume]({IRI['source']}) (published
+  {IRI['published']}, redrawn). The model's own reference is 1959-01–2009-08; on that base Niño 3.4 is 0.04–0.23 °C higher,
+  depending on the month. Both are in the NetCDF.
+* **Chance of a top-quintile month** is the calibrated probability that the monthly SST falls in the warmest fifth of the
+  1959–2009 climate for that calendar month (20 % = climatological odds). Because the ocean has warmed since then, high
+  values are widespread.
+* **Forecast uncertainty (σ)** is a per-cell Gaussian spread. The probabilities are marginal per grid cell, not a coherent
+  ensemble, so they cannot be summed into regional or Niño 3.4 uncertainties.
 
 ### Limits — please read
-* A research forecast, not an official or operational warning product. No probability is implied: this is one
-  deterministic run.
-* Checkpoint acceptance for the operational ocean model used agreement with the historical NWP multi-model mean and
-  with the current IRI plume, so this run is **not an independent test of skill**. Published hindcast skill is in the paper.
-* The six-week atmospheric run is **not drift-corrected**. Against the ERA5 1990–2017 weekly climatology, its
-  area-mean 2 m temperature poleward of 60°N is about +5 °C in week 1 and +14 °C in week 6, while 60°S–60°N stays within
-  ±1 °C. Treat high-latitude fields as model climate, not as anomalies; anomaly maps are therefore not shown.
+* A research forecast, not an official or operational warning product.
+* The released weights were calibrated with targets through 2026-09, and the calibration settings were selected on
+  year-blocked diagnostics of the same years. Every historical score of these exact weights, including the published
+  2016–2022 hindcast, is therefore in-sample. The *Skill & hindcasts* tab shows the held-out check of a refit copy.
+* **ENSO:** the forecast damps the 2026 El Niño quickly. Niño 3.4 drops from {NINO_HIST[-1]:+.2f} °C observed in
+  {OBS_MONTH:%B} to {NINO[0]:+.2f} °C in {MONTHS[0]:%B}, well below the IRI dynamical models. The damping is already in
+  the four frozen base heads (lead 1: +1.5 to +1.8 °C on the model base), and Niño 3.4 is the region where the held-out
+  check found this method weakest. Read the Niño 3.4 outlook with that in mind. canglong.earth still serves the previous
+  16-month Public-Swin run, whose Niño 3.4 stays at +3.2 °C or more through December 2026 (on its 1959–2016 base).
+* The six-week atmospheric run is **not drift-corrected**; high-latitude 2 m temperature drifts warm with lead time
+  (about +5 °C in week 1 to +14 °C in week 6 poleward of 60°N in earlier runs). Read those fields as model output; anomaly
+  maps are not shown.
 
 ### Provenance of this run
-* Ocean run: {RUN['run_time_utc'][:16].replace('T', ' ')} UTC, re-run from the frozen checkpoints in an isolated
-  directory; identical to the operational product (max |ΔSST| = {RUN['check_vs_operational']['max_abs_sst_diff_K']} K,
-  max |ΔNiño 3.4| = {RUN['check_vs_operational']['max_abs_nino34_diff_K']} K).
-* Checkpoint SHA-256: 1° `{RUN['checkpoint_sha256']['lowres_checkpoint'][:16]}…`, 0.25° refiner
-  `{RUN['checkpoint_sha256']['highres_checkpoint'][:16]}…`; input cache `{RUN['input_cache_sha256'][:16]}…`.
+* Forecast file: `{RUN['sst_file']}`, derived from the release NetCDF (SHA-256 `{(RUN['source_forecast_sha256'] or '')[:16]}…`)
+  without changing any value. Checks: SST − anomaly reproduces the stored climatology to
+  {RUN['checks']['forecast_sst_minus_ssta_vs_climatology_max_abs_K']:.1e} K, and the observed {OBS_MONTH:%B} anomaly
+  equals the model's own input context (max |Δ| = {RUN['checks']['observed_ssta_vs_model_context_max_abs_K']} K).
 * Atmosphere: {RUN['atmos_source']}.
 
 ### Cite
@@ -275,7 +360,7 @@ CSS = """
 .hero .chips a {text-decoration: none !important; padding: 5px 12px; border-radius: 999px; font-size: .9rem;
   border: 1px solid var(--border-color-primary); color: var(--body-text-color) !important; background: var(--block-background-fill);}
 .hero .chips a:hover {border-color: var(--color-accent);}
-.kpis {display: grid; grid-template-columns: repeat(auto-fit, minmax(190px, 1fr)); gap: 10px; margin: 14px 0 4px;}
+.kpis {display: grid; grid-template-columns: repeat(auto-fit, minmax(170px, 1fr)); gap: 10px; margin: 14px 0 4px;}
 .kpi {padding: 12px 14px; border-radius: 12px; border: 1px solid var(--border-color-primary); background: var(--block-background-fill);}
 .kpi-label {font-size: .8rem; color: var(--body-text-color-subdued);}
 .kpi-value {font-size: 1.55rem; font-weight: 650; margin: 2px 0;}
@@ -287,26 +372,43 @@ with gr.Blocks(title="CAS-Canglong · SST & ENSO forecast") as demo:
     with gr.Tabs():
         with gr.Tab("🌊 Ocean & ENSO"):
             with gr.Row(equal_height=True):
-                field = gr.Radio(["SST anomaly (1°)", "SST (0.25°)"], value="SST anomaly (1°)", label="Field", scale=1)
+                field = gr.Radio(list(FIELDS), value="SST anomaly", label="Field", scale=2)
                 lead = gr.Slider(0, len(FRAMES) - 1, value=1, step=1, label="Month (0 = last observed input month)", scale=3)
-            caption = gr.Markdown(frame_label(1))
+            caption = gr.Markdown(frame_label("SST anomaly", 1))
             image = gr.Image(map_path("SST anomaly", 1), show_label=False, interactive=False, buttons=["download", "fullscreen"])
             gr.Markdown("#### Niño 3.4 outlook vs the IRI ENSO plume")
             gr.Plot(plume_figure(), show_label=False)
-            gr.Markdown("Dotted lines connect each forecast curve to its latest available monthly observation "
-                        "(IRI for the IRI models; ERA5 for CAS-Canglong). These are visual guides from monthly "
-                        "observations to 3-month forecast means; diamonds mark observations.")
-            gr.Markdown("Monthly CAS-Canglong Niño 3.4 (°C): " + ", ".join(
-                f"{t:%b %Y} {v:+.2f}" for t, v in zip(MONTHS, NINO)) + ".")
-            lead.change(lambda f, i: (map_path(f, i), frame_label(i)), [field, lead], [image, caption])
-            field.change(lambda f, i: map_path(f, i), [field, lead], image)
+            gr.Markdown("Monthly CAS-Canglong Niño 3.4 (°C, 1991–2020 base): " + ", ".join(
+                f"{t:%b %Y} {v:+.2f}" for t, v in zip(MONTHS, NINO)) + ". The forecast damps the observed El Niño "
+                "much faster than the IRI dynamical models; see *About* and *Skill & hindcasts*.")
+            lead.change(lambda f, i: (map_path(f, i), frame_label(f, i)), [field, lead], [image, caption])
+            field.change(lambda f, i: (map_path(f, i), frame_label(f, i)), [field, lead], [image, caption])
         with gr.Tab("📍 Regional outlook"):
             region = gr.Dropdown(list(REGIONS), value=list(REGIONS)[0], label="Region")
             rplot = gr.Plot(region_figure(list(REGIONS)[0]), show_label=False)
-            gr.Markdown(f"Box means of the native 1° anomaly field (cos-latitude weighted, ocean cells only). The headline "
-                        f"Niño 3.4 ({NINO[0]:+.2f} °C for {MONTHS[0]:%b %Y}) is the operational index from the 0.25° field.")
+            gr.Markdown("Box means of the native 1° anomaly field (cos-latitude weighted, model ocean cells only), "
+                        "relative to ERA5 1991–2020, with the last 12 observed months for context.")
             rtable = gr.Dataframe(region_table(list(REGIONS)[0]), interactive=False)
             region.change(lambda r: (region_figure(r), region_table(r)), region, [rplot, rtable])
+        with gr.Tab("📈 Skill & hindcasts"):
+            gr.Markdown("#### Held-out test on 2023–2026")
+            with gr.Row():
+                htrack = gr.Dropdown(list(HELDOUT_TRACKS), value=list(HELDOUT_TRACKS)[0], label="Comparison", scale=3)
+                hregion = gr.Dropdown(list(HELDOUT_REGIONS), value="Global", label="Region", scale=2)
+                hmetric = gr.Radio(["Spatial ACC (higher is better)", "RMSE (lower is better)"],
+                                   value="Spatial ACC (higher is better)", label="Metric", scale=2)
+            hplot = gr.Plot(heldout_figure(list(HELDOUT_TRACKS)[0], "Global", "Spatial ACC"), show_label=False)
+            htable = gr.Dataframe(heldout_table(list(HELDOUT_TRACKS)[0], "Global"), interactive=False)
+            gr.Markdown(HELDOUT_NOTE)
+            for c in (htrack, hregion, hmetric):
+                c.change(lambda t, r, m: (heldout_figure(t, r, m), heldout_table(t, r)), [htrack, hregion, hmetric], [hplot, htable])
+            gr.Markdown(f"#### Hindcast 2016–2022 · [dataset](https://huggingface.co/datasets/{HINDCAST_REPO})")
+            hlead = gr.Slider(1, 12, value=3, step=1, label="Lead (months)")
+            hcplot = gr.Plot(hindcast_figure(3), show_label=False)
+            hccap = gr.Markdown(hindcast_caption(3))
+            hlead.change(lambda k: (hindcast_figure(k), hindcast_caption(k)), hlead, [hcplot, hccap])
+            gr.Dataframe(HINDCAST_SKILL.rename(columns={"lead": "Lead", "n": "Months", "corr": "Correlation", "rmse_K": "RMSE (°C)"}),
+                         interactive=False, label="Niño 3.4 hindcast scores by lead (in-sample)")
         with gr.Tab("🌦️ Atmosphere · 6 weeks (experimental)"):
             gr.Markdown("Weekly-mean fields of the CAS-Canglong V3.5 atmospheric run published on canglong.earth. "
                         "**Not drift-corrected** — high-latitude temperatures drift warm with lead time (see *About*).")
@@ -318,15 +420,31 @@ with gr.Blocks(title="CAS-Canglong · SST & ENSO forecast") as demo:
             aimg = gr.Image(atmos_path("2 m temperature", 1), show_label=False, interactive=False, buttons=["download", "fullscreen"])
             week.change(lambda v, w: (atmos_path(v, w), week_label(w)), [avar, week], [aimg, wcap])
             avar.change(lambda v, w: atmos_path(v, w), [avar, week], aimg)
-        with gr.Tab("📦 Data"):
-            gr.Markdown("NetCDF and CSV files behind every map and chart on this page.")
-            gr.File([str(DATA / RUN["sst_file"]), str(DATA / RUN["atmos_file"]), str(DATA / "nino34.csv"), str(DATA / "run.json")],
+        with gr.Tab("📦 Model & data"):
+            gr.Markdown(f"* **Model** — frozen weights, calibration states, inference code and a CPU example: "
+                        f"[{MODEL_REPO}](https://huggingface.co/{MODEL_REPO})\n"
+                        f"* **Hindcast** — observed and lead 1–12 SST for every target month of 2016–2022: "
+                        f"[{HINDCAST_REPO}](https://huggingface.co/datasets/{HINDCAST_REPO})\n"
+                        "* **This run** — the files behind every map and chart on this page:")
+            gr.File([str(DATA / RUN["sst_file"]), str(DATA / "nino34.csv"), str(DATA / RUN["atmos_file"]),
+                     str(DATA / "hindcast_nino34.csv"), str(DATA / "round6_heldout_metrics.csv"), str(DATA / "run.json")],
                     file_count="multiple", label="Downloads", interactive=False)
             gr.Code(f"""import xarray as xr
-from huggingface_hub import hf_hub_download
-path = hf_hub_download("LonghaoWang/CAS-Canglong", "data/{RUN['sst_file']}", repo_type="space")
-ds = xr.open_dataset(path)          # ssta / sst (1°), sst_025 (0.25°), nino34, observed fields
-print(ds.nino34.to_series())""", language="python", label="Load in Python")
+from huggingface_hub import hf_hub_download, snapshot_download
+
+# this run: mean, anomaly (both bases), sigma, 10/90 % quantiles, quintile probabilities, Nino 3.4
+ds = xr.open_dataset(hf_hub_download("LonghaoWang/CAS-Canglong", "data/{RUN['sst_file']}", repo_type="space"))
+print(ds.nino34_9120.to_series().round(2))
+
+# the model: verify the frozen package and replay the October 2026 forecast on a CPU
+pkg = snapshot_download("{MODEL_REPO}", allow_patterns=["frozen_model_20261025_v1/*"])
+# cd <pkg>/frozen_model_20261025_v1 && python -B run_frozen.py --verify-only
+# python -B run_frozen.py --prepared-input examples/input_only_202610.npz --output /tmp/forecast_202610.nc
+
+# the 2016-2022 hindcast (obs + lead1..lead12 per target month)
+hc = xr.open_dataset(hf_hub_download("{HINDCAST_REPO}",
+                     "hindcast_2016_2022_v1/sst_hindcast_201601_202212_lead01_12.nc", repo_type="dataset"))""",
+                    language="python", label="Load in Python")
         with gr.Tab("ℹ️ About"):
             gr.Markdown(ABOUT)