ψ Vikshep
Sign inDownloadTalk to founders

Documentation

From Geant4 output to physics answer.

Vikshep is a deterministic feature-extraction plane for physics data. Its feature plane is the wavelet scattering transform: nothing is learned during feature extraction — the wavelet filters are fixed analytic Morlets chosen by geometry, not trained. Translation invariance and deformation stability are theorems about the transform, not benchmarks. The transform is delivered by the open deterministic core, in development; see Engine status below for what the released CLI does today.

Engine status

  • Available nowThe vikshep-ingest / vikshep-recipe CLI: Geant4 CSV ingest, per-event 2-D rasterization into shared memory, per-event aggregates, calibration regression, DisCo-penalised tagging on the aggregates (training gradient: Pearson proxy; reported dCorr²: exact weighted), and the benchmark harness. Tested today: byte-identical reports under a fixed seed on the same machine.
  • In developmentThe open deterministic scattering core (samvardhan03/vikshep-compute), which computes S₀/S₁/S₂ and r₂ = S₂/S₁ on the ingested grids. It is built against a written determinism specification (VDS-1) with a cross-platform conformance suite.
  • PlannedDesktop app. GPU acceleration.

Free vs. paid boundary

The control plane is AGPL-3.0 and ships in full. The private engine is intended to ship as binaries; none has been released yet. The table below is the single source of truth; it is also referenced from Pricing.

Free forever (AGPL, local)

Trial (hosted plane)

Lab / Enterprise (paid)

  • +G4 Direct Interface — Geant4 CSV → features in one command
  • +All loaders: g4, root-uproot, hdf5, well (local CPU)
  • +Per-event 2-D rasterization and aggregate features
  • +DisCo mass-decorrelation training and calibration CLIs
  • +Benchmark harness (Asimov proxy, JSD)
  • +Agent orchestrator and all recipes
  • +Web preview dashboard (client-side)
  • +AGPL-3.0 — full source, modifications must be open
  • +Community support via GitHub Issues
  • ~Hosted plane access — dual-capped by TRIAL_DAYS and TRIAL_GPU_SECONDS
  • ~No card required
  • ~One trial per verified email
  • ~The download itself (AGPL source, local CLI) never expires
  • +Hosted GPU scattering runs (Lab / Enterprise)
  • +Well-scale ingest over 1 TB (hosted)
  • +SW1 catalogue search (hosted)
  • +Harness-as-a-service
  • +Premium 3-D SO(3) engine paths
  • +Commercial license (no AGPL copyleft)
  • +Priority support and SLA

The download itself never expires — the AGPL source and local CLI run indefinitely on your own hardware at no cost. The trial covers the hosted plane only and is dual-capped (time + GPU-seconds); both caps are enforced server-side and require no card.

How it fits together

The control plane (agent, loaders, recipes, site — all AGPL-3.0) speaks to the open MCP data plane over a frozen seam: 28-char SHA3-256 object IDs in POSIX shared memory, line-delimited JSON-RPC 2.0. Raw tensors never reach TypeScript or the browser. Build the MCP binary (vikshep-mcp) from the open core and set VIKSHEP_MCP_BIN to its path. The loaders and CLI recipes are pure Python; the scattering path requires the [scattering] extra and a Rust toolchain.