product · sparc desktop

A neural pipeline that simulates adaptation.

SPARC is a Tauri/React desktop app wrapped around one thing: a 7-stage neural pipeline with built-in causal inference. Build a project, run the pipeline, simulate adaptation — every stage is reproducible, every prediction has a posterior.

01 · pipeline · the product

Seven stages. None of them optional.

The pipeline is the product. Each stage gates the next. χ², σ, and physics constraints are checked before you can advance. Stages are reproducible from project.yml — bring the same file to a different machine and get bit-identical posteriors.

stages
ingest → correlogram → causal → neural → posterior → scenario → report
checks
χ² < 1.2 · σ < 0.3 · monotone signs honored
runtime
CUDA · 11.8 / Apple MPS / CPU fallback
audit
every run emits a signed manifest
pipeline · 7 stages
running · stage 04
00ingestraster + vector + project.yml✓ done
01correlogramMoran's I · Geary's C✓ done
02causal graph24 nodes · 41 edges✓ done
03neural trainGW3C · χ² 0.93✓ done
04posterior8 chains · σ 0.154▸ live
05scenarioadaptation · uncertaintyqueued
06reportaudit trail · pdf · mdqueued
02 · neural + causal

A neural model that learns under causal structure.

At the heart of SPARC: a locally-weighted neural ensemble trained under directed causal constraints. Edges are propose-then-adjudicate; signs are physically grounded; coefficients vary smoothly across space. The result is a model that doesn't just fit — it admits to a theory of how the world works.

model
geographically-weighted neural ensemble · 8-chain HMC
structure
directed causal graph · monotone signs · sign locks
constraints
physics-class per edge · cycle rejection · mass-balance
review
accept · reject · flip · constrain · audit
causal · edge review
3 of 41
Pct_Canopy AAT_z✓ accepted
sign · physics · shading|β| · 0.78
Impervious AAT_z✓ accepted
sign · +physics · albedo|β| · 0.82
Bldg_Height Wind_z▲ review
sign · physics · drag|β| · 0.41
03 · adaptation

Simulate adaptation, with error bars.

An adaptation scenario is a perturbation applied to the trained ensemble. Increase canopy. Lower impervious. Raise sea level by 0.6m. SPARC re-runs prediction across the grid and returns a posterior mean, σ, and a pixel-wise uncertainty raster — never a single number.

perturbations
additive · multiplicative · spatial mask
outputs
mean raster · σ raster · diff raster
compare
side-by-side, paired ttest, drift map
export
GeoTIFF · CSV · PDF report
adaptation · "more shade"
philadelphia · 2025
Δ AAT_z mean
−0.258 (σ 0.154)
peak temp
+9.4 °C ↓ from +11.8 °C
tracts improved
218 of 287
physics check
✓ monotone · χ² 0.93
philosophy · 01

No black boxes.

Every coefficient is interrogable. Every edge has a sign. Every prediction has a posterior.

philosophy · 02

Physics, not vibes.

Monotone constraints, mass balance, conservation laws — the model honors them or it doesn't run.

philosophy · 03

Reproducible by default.

project.yml is the source of truth. Every result is regenerable from a single command, 5 years from now.

under the hood

What's actually running.

frontend
Tauri 2.0 · React 18 · TypeScript
backend
Python 3.12 · sidecar binary :8008
model
GW3C neural · 8-chain HMC
geo stack
GDAL · rasterio · GeoPandas
compute
CUDA 11.8 · MPS · CPU fallback
storage
DuckDB · parquet · GeoTIFF
ai assist
Claude Sonnet 4.6 (BYOK)
license
AGPL · open core

Ready to run the pipeline?