Constellation v3.0 · public beta

Three modules, one substrate.

Nova-Sim simulates. Nova-Obs ingests. Nova-Loop decides. They look like three products, but they are three doors into the same runtime, sharing one asset store, one scheduler, one provenance graph.

Request access How the substrate works →
The three modules

Each one is a face of the runtime.

No exports between them. No file hand-offs. Whatever one module produces, the next reads in place, by hash, with lineage intact.

// 01 · simulate

Nova-Sim

GPU-accelerated simulation primitives for radiative transfer, fluid dynamics, and orbital mechanics, composed into experiment graphs that the scheduler places and the provenance graph records.

kernels: 240+ precision: fp64 / mixed checkpointing: deterministic graphs: DAG, branch & merge backends: CUDA, ROCm, CPU
// 02 · ingest

Nova-Obs

The pipeline for raw observational data: ingest, align, calibrate, and version against an immutable, content-addressed asset store. A calibrated frame is the exact object Nova-Sim consumes next.

throughput: 12 GB/s/node formats: FITS, HDF5, Zarr provenance: W3C-PROV addressing: BLAKE3 tiers: hot / warm / cold
// 03 · decide

Nova-Loop

Closed-loop experimentation. State a hypothesis, give it a budget, and Nova-Loop plans runs against the scheduler, surfacing only the results worth your time and debiting the grant as it goes.

schedulers: 4 (incl. Bayesian) quotas: per-PI, per-grant notify: Slack, email, page stopping: budget / convergence replay: deterministic
How a question moves through it

One question, three modules, no copies.

A real run does not stay in one module. It flows, and because everything shares the substrate, the flow leaves a single, queryable trail.

OBS
Land the data

Nova-Obs ingests raw frames, calibrates them, and writes immutable assets addressed by hash.

SIM
Run the model

Nova-Sim builds an experiment graph that references those exact assets, no export, no duplication.

LOOP
Close the loop

Nova-Loop reads results, proposes the next runs against the budget, and schedules them automatically.

PROV
Account for it

The provenance graph ties every artifact to its inputs, its code, and the grant that paid for it.

Go deeper

Read what the modules made.

The modules are the how. The white papers and open datasets are the what, the results these three doors have produced for our research partners.