ARDA discovers governing equations, causal structures, and conservation laws from raw observational data — with zero prior knowledge and full governance at every step.
ARDA — The Discovery Contract
Most scientific software optimizes for speed. ARDA optimizes for a stronger contract: every claim is typed, every run produces an evidence ledger, every discovery can be reproduced with a single truth dial setting. If a discovery cannot be reproduced, it is not a discovery. It is noise.
Four Discovery Modes
ARDA routes your data through the right discovery mode automatically. Each mode produces typed scientific claims, feeds the same evidence ledger, and respects the same truth dial governance.
Governing equations from raw data
Genetic programming and symbolic regression across ODE, PDE, SDE, GR, and tabular domains. Routes through PySR and GP backends with universe-aware strategy registries. Discovers closed-form laws — not black-box fits.
Physics-informed latent dynamics
Multimodal episodes encoded through physics-aware architectures — HNN, LNN, Neural ODE, SSM, SE(3)-equivariant, EGNN — with learned decomposition, ensemble training, and regime classification.
Neural encode, symbolic distill
The best of both worlds. Neural encoders learn latent representations, then symbolic distillation extracts closed-form equations from the latent space. Produces both latent dynamics and observable dynamics claims.
Learn causal graphs, run interventions
Dual-field dynamics: SIREN base field + GNN causal field with optional steering. Active Causal Investigator runs multi-cycle interventions with BALD-scored probes and edge ablation fusion.
Governance
One control that governs the rigor-speed tradeoff across the entire pipeline. Each level enforces specific controls, determinism requirements, and claim promotion gates.
Fast iteration. No controls. Claims tagged as hypotheses.
Controlled testing. Claims promoted to provisional on pass.
Full determinism. All controls. Publish bundle with replay recipe.
Causal Dynamics Engine
The CDE learns dual-field dynamics where the total force is decomposed into a SIREN base field, a GNN causal field, and an optional steering field. An Active Causal Investigator runs multi-cycle interventions with BALD-scored probes to iteratively refine the causal graph.
Ftotal = Fbase(SIREN) + Fcausal(GNN) + Fsteer
Scientific Claims
Every discovery produces typed scientific claims — not free-text outputs. Each claim carries metadata, confidence scores, and provenance. The evidence ledger records everything: dataset hashes, config hashes, control results, hardware, library versions, and CDE beliefs.
Every run writes a LedgerEntry: dataset hash, config hash, metrics, controls, claims, git commit, hardware fingerprint, library versions, and CDE-specific fields.
Discovery cards, model cards, indeterminacy reports, and decomposition reports. Promotion gates enforce minimum controls passed, R² thresholds, and determinism tier requirements.
PerturBench Norman19 for combination perturbation prediction. Tokamak CDE for plasma dynamics. Registry-based adapters for custom benchmarks.
Agent-Native Platform
ARDA is designed from the ground up for persistent agent access. AI agents discover the platform automatically, connect through any of five surfaces, maintain stateful sessions across interactions, and operate within governed autonomy boundaries.
Every surface — API, SDK, MCP, CLI, and agent sessions — talks to the same discovery engine, evidence ledger, and governance stack. Agents can start a discovery campaign, pause, come back hours later, and resume where they left off. Autonomy policies control what agents can do without human approval.
FastAPI with full OpenAPI spec. Every resource — projects, runs, campaigns, claims, evidence, agents — is a first-class REST endpoint. Agents auto-discover capabilities via /.well-known/ai-plugin.json.
ADPClient (sync) and AsyncADPClient with typed models for every resource. Your agents import arda-sdk and call methods directly — no HTTP wrangling.
100+ tools via Model Context Protocol. stdio for local agents (Claude Desktop, Cursor), SSE for remote. Auto-discoverable at /.well-known/mcp.json.
Full API and SDK surface from the terminal. arda sdk call <method>, arda api request, arda workflow — scriptable, pipeable, agent-friendly.
Persistent sessions with lifecycle management. BYO or native scientist workflows. Task queues, heartbeats, autonomy policies, lineage, and monitoring.
Agents discover ARDA automatically via standard endpoints. /.well-known/ai-plugin.json for agent manifests, /.well-known/mcp.json for MCP tool discovery, /v1/capabilities for platform introspection.
Agent sessions persist across interactions. Stages: planning → ready → running → completed. Task queues with heartbeats, requeue on failure, and full lineage tracking. Sessions survive agent restarts.
Governance for what agents can do without human approval. Per-project and per-campaign policies control which operations require human-in-the-loop confirmation versus autonomous execution.
ARDA ships a built-in native_arda agent that autonomously designs experiments, runs discoveries, evaluates claims, and iterates — all within governed session boundaries.
Beyond tools, ARDA exposes MCP resources: arda://capabilities for platform state, arda://product-docs for documentation, arda://product-qa for agent Q&A — all queryable in-context.
Every surface publishes a typed manifest at /v1/{surface}/manifest. Agents can introspect available methods, transport options, and runtime requirements for any interface.
Extensible Architecture
ARDA's neural pipeline is fully composable. The DataProfiler selects modules from a typed registry. Every slot — temporal, spatial, relational, dynamics, controls, decomposition — is pluggable.
SSM, GRU, Transformer, Neural ODE, HNN, LNN
EGNN, SE(3), Conv, Spectral, PointNet, Invariant MLP
GAT, MPNN, NRI, Dynamic Graph
Neural ODE, HNN, Neural SDE, FNO, SIREN, CDE, Lyapunov
Time shuffle, Phase random, Bootstrap, Feature shuffle, OOD
Conservative-dissipative, Multi-scale, Null
Cross-attention, Gated, Token mixer, Concat
ACI, Experiment Policy, Probe Execution
Orchestration
Campaigns orchestrate multi-phase discovery workflows with budget management and automatic validation passes. Agents connect via any surface, maintain persistent sessions that survive restarts, and operate within governed autonomy boundaries — from fully autonomous to human-in-the-loop.
Pricing
Pay for the platform and compute you use.
If you're working with dynamical systems, causal inference, or scientific data where reproducibility matters, ARDA is built for you.