CAPRI

The ocean has no walls, yet it holds rooms.

CAPRI stacks twenty-one channels of Sentinel-3 ocean colour — raw spectral bands and the biogeochemistry drawn from them — into a navigable terrain you can orbit, lift apart, and enter. Where the colour of water settles into recurrence, a state appears.

The framework does not name these states. It shows you where they begin.

Grab the latest build from Releases:

• …aarch64.dmg — macOS (Apple Silicon) desktop app. Open it, drag to Applications. It's an unsigned beta, so on first launch right-click the app → Open → Open.
• capri-source-….zip — the full source for that release, to read or build yourself.

Double-click it. One self-contained file: no terminal, no server, no internet, no install. Six instruments, the real datasets baked in, and drag-in support for your own CSVs (parsed locally — nothing is uploaded). See bench/README.md.

capri/eef_viewer.html is the legacy viewer. It requires python capri/server.py running and internet access for its CDN scripts. Use bench/ instead.

A satellite scene carries 21 (or more) Sentinel-3 OLCI channels per pixel — raw spectral bands and the products derived from them: chlorophyll, phytoplankton absorption, backscatter, suspended sediment, light attenuation, fluorescence. Read flat, in a table, that structure stays hidden. CAPRI renders it as a navigable 3D stack of spatial layers — each variable given height and colour — and then reads the stack for state: the regimes, fronts, and anomalies a single flat map keeps to itself.

It runs entirely on your machine. No account, no server, no telemetry.

Make sure to include the following bands: Oa1-Oa11,Oa13,ADG443_NN, CHL_NN, KD490_M07, PAR, TSM_NN, aphy_443, bbp_443, latitude, longitude

• A layered terrain — every channel becomes a displaced, colour-mapped surface. Orbit it, drag the layers into any order, toggle each one, drop into a true orthographic plan view where every layer reads at the same scale.
• UMAP — projects each cell's full spectral signature into a 2-D map so similar water lands together and regimes separate as clusters.
• PLS regression + VIP — asks what drives a chosen variable. Partial Least Squares is built for collinear spectral bands, so its importance scores stay honest where correlated bands would mislead a random forest. Validated by 5-fold cross-validation.
• Correlation & latent ecology — the relationships and covariances between all channels, plus a PCA/k-means read of recurring regimes.
• Embedders (in progress) — create and persist contrastive embedders, feeding them scenes over time; the learned encoder is the next step.
• Real data in — synthetic streams for exploration, or your own GeoTIFF / CSV / sparse-sensor data. Derived products (e.g. ESA-standard Fluorescence Line Height) are computed from the raw bands, not invented.

Everything is computed from the observed (or uploaded) values, seeded for reproducibility — no fabricated scores.

cd eef-dashboard
npm ci
npm run dev          # http://localhost:5173  — the dashboard in your browser
npm run tauri:build  # build the native macOS app (.dmg)

The shipping application lives in eef-dashboard/ — a TypeScript / React + Three.js front end in a Tauri (Rust) shell.

An honest beta. The macOS builds are unsigned; contrastive-embedder training is not yet wired (the tab accumulates and persists training data today). Read the in-app scores as relative skill across a single scene, not absolute accuracy.

Apache-2.0 — permissive, with an explicit patent grant. Use it, fork it, build on it.