Align Space docs with CPU demo

README.md

@@ -6,9 +6,9 @@ colorTo: indigo
 sdk: gradio
 sdk_version: 6.14.0
 app_file: app.py
-pinned: false
+pinned: true
 license: apache-2.0
-short_description: Multi-modal RS stack for dark vessel detection (S1+S2+AIS).
+short_description: CPU-safe dark-vessel reasoning demo for S1/S2/AIS fusion.
 tags:
   - remote-sensing
   - sentinel-1
@@ -31,8 +31,8 @@ tags:
 
 # DarkVesselNet
 
-Multi-modal remote sensing stack for dark vessel detection. Sentinel-1 SAR plus Sentinel-2 optical plus AIS, fused through a geospatial foundation model backbone (Prithvi-2 / Clay v1 / SatMAE++ / DOFA / SatlasNet / RemoteCLIP), with physics-informed anomaly reasoning via TGARD and Pi-DPM.
+Multi-modal remote sensing stack for dark vessel detection. Sentinel-1 SAR plus Sentinel-2 optical plus AIS are fused through geospatial foundation-model backbones in the full repository, with physics-informed anomaly reasoning via TGARD and Pi-DPM.
 
-Click an area of interest on the Mapbox globe (Strait of Hormuz, South China Sea, Galapagos EEZ, Sea of Japan, Gulf of Oman) and the Space pulls the latest cloud-free Sentinel-2 chip with the nearest Sentinel-1 GRD scene from Microsoft Planetary Computer, runs the seven canonical EO heads, joins to AIS, and renders a Folium overlay of candidate dark vessels with a per-detection reasoning trace.
+This public Hugging Face Space is the CPU-safe demo path: pick an area of interest and the app generates shape-consistent synthetic Sentinel-1, Sentinel-2, and AIS tensors, then runs the same scoring trace used by the scaffold. It does not download live Planetary Computer scenes or MarineCadastre AIS on the free Space tier.
 
 See the [GitHub repository](https://github.com/arunshar/darkvessel-stack) for the full pipeline, training recipes, and xView3-SAR leaderboard reproduction.

app.py

@@ -1,17 +1,14 @@
 """Gradio HF Space entry point for DarkVesselNet.
 
-The Space ships a Mapbox dark-themed globe with five pre-built AOIs (Gulf of
-Oman, Strait of Hormuz, South China Sea, Galapagos EEZ, Sea of Japan). On
-click the Space pulls the latest cloud-free Sentinel-2 chip and the nearest
-Sentinel-1 GRD scene from Microsoft Planetary Computer, runs the seven
-heads, joins to MarineCadastre AIS, and renders a Folium overlay of
-candidate dark vessels with a reasoning trace per detection.
-
-For HF CPU smoke we ship a stubbed backbone that returns shape-consistent
-features without downloading 600M-parameter weights.
+The public Space is a CPU-safe scaffold: it exposes the same AOIs and
+reasoning trace as the full project, but uses deterministic synthetic
+Sentinel-1/Sentinel-2/AIS tensors instead of live external downloads or
+600M-parameter foundation-model weights.
 """
 from __future__ import annotations
 
+import zlib
+
 import gradio as gr
 import torch
 
@@ -28,7 +25,7 @@ AOIS = {
 def run_pipeline(aoi: str) -> str:
     if aoi not in AOIS:
         return "Unknown AOI."
-    torch.manual_seed(hash(aoi) & 0xFFFF)
+    torch.manual_seed(zlib.crc32(aoi.encode("utf-8")) & 0xFFFF)
     chip = torch.randn(1, 6, 224, 224)
     ais = torch.randn(1, 12, 5)
     optical_sar_energy = chip.square().mean()
@@ -37,16 +34,20 @@ def run_pipeline(aoi: str) -> str:
     lat, lon = AOIS[aoi]
     return (
         f"AOI: {aoi} ({lat:.3f}, {lon:.3f})\n"
-        f"backbone: prithvi-2 (stub mode on CPU)\n"
+        f"backbone: prithvi-2 interface (CPU demo mode)\n"
         f"dark vessel probability: {score:.3f}\n"
         f"reasoning: TGARD gap score 0.42, Pi-DPM kinematic residual 0.18 m/s^2.\n"
-        f"sensor stack used: Sentinel-1 VV/VH GRD, Sentinel-2 L2A, AIS DMA feed.\n"
+        f"sensor stack simulated: Sentinel-1 VV/VH GRD, Sentinel-2 L2A, AIS DMA feed.\n"
     )
 
 
 def build_ui() -> gr.Blocks:
     with gr.Blocks(title="DarkVesselNet") as demo:
-        gr.Markdown("# DarkVesselNet\nMulti-modal remote sensing for dark vessel detection.")
+        gr.Markdown(
+            "# DarkVesselNet\n"
+            "CPU-safe dark-vessel reasoning demo for S1/S2/AIS fusion. "
+            "The full repository contains the live data connectors and model-backed pipeline."
+        )
         aoi = gr.Dropdown(choices=list(AOIS), value="Gulf of Oman", label="Area of interest")
         out = gr.Textbox(label="Pipeline output", lines=8)
         btn = gr.Button("Run DarkVesselNet")

requirements.txt

@@ -1,3 +1,3 @@
 torch>=2.3
-gradio>=4.30
+gradio>=6.14.0
 audioop-lts>=0.2.0