SPARC is a Tauri/React desktop app wrapped around one thing: a seven-stage neural pipeline with built-in causal inference. Bring your own data, watch each stage open as you scroll, and finish with a report shaped for who's reading it.
Drop in rasters, shapefiles, GeoJSON boundaries, parquet tables, or a project.yml — SPARC reprojects, resamples, and clips them to your study area without ever phoning home. Pull in canopy, impervious, parcels, permits, and equity layers, then pin the whole thing to a single manifest.
Before fitting any model, SPARC maps how tightly each variable clusters in space — measuring the distance at which two locations stop influencing each other. That range shapes every downstream step.
This stage sweeps a local statistical model across every location — stripping noise and identifying which variables carry a real signal in each neighborhood. Only the meaningful ones move forward, with their direction of effect preserved.
The core SPARC stage: a neural model trained at every location simultaneously, with effect directions locked to what the previous stage found and physics enforced at every checkpoint. If a result violates either, it doesn't advance.
Causal structure is explored, not assumed. SPARC searches the space of valid cause-and-effect maps that fit your data, then builds a probability distribution over which connections are most credible. Every link is auditable.
Push the model with real constraints. Set a dollar budget. Toggle equity layers — heat vulnerability, income, canopy access — as inputs the optimizer must honor. Get back a posterior raster, not a single number.
The same run produces three reports: Technical (methods, posterior diagnostics, code refs), Planner (briefs, tradeoffs, equity overlays), and Public (plain-language, before/after, FAQ). Each one ships with a signed manifest so anyone can re-run it.
Every number has a source. Every effect has a direction. Every prediction comes with a confidence range.
The model must respect how heat, water, and air actually behave — or it doesn't run.
Every project saves as a single file. Any result can be regenerated exactly, years from now, on any machine.
Everything runs on your machine. Nothing is uploaded or shared without your explicit say-so.
