Four engines.
One causal graph.
Manifold is a causal-intelligence terminal built on four specialized engines running in parallel on a shared scored graph.
Recovers the causal topology of a critical system from observational data — which nodes act on which, and through what edges.
- Constraint-based, score-based, and gradient-based search across observational data
- Five discovery algorithms compared side-by-side — PC, FCI (CMI-knn), NOTEARS, PCMCI-linear, lag-correlation
- Auto → confirmed edge lifecycle with promote / demote affordance
Verifies geopolitical and structural claims about the graph — sanctions exposure, jurisdictional dependencies, control relationships.
- Formal logic over node and edge attributes
- Audit trail for every claim with traceable evidence
- Surfaces silent failure modes before they become incidents
Asks what-if questions on the causal graph — substitute a node, sever an edge, model an embargo — and computes the counterfactual world.
- Unified intervention surface — scenario input, ablation, and interdiction co-located on one pane
- APPLY ALL + SIMULATE — counterfactual cascade auto-runs after intervention
- Explainable cuts — every interdicted edge surfaces a WHY trace back to the graph
Simulates cascade dynamics across the graph and characterizes tail risk beyond traditional VaR.
- Monte Carlo cascade simulation with configurable shock distributions
- Tail-depth statistics — fat-tail severity beyond VaR
- System-level outputs — ΩSF, cascade reach, time-to-failure
Engines as tools. Conversation as control.
Every engine exposes itself to a conversational layer on top of the terminal. Operators run ablations, build counterfactuals, and chase cascade chains in natural language — by typing or speaking — and the engines execute against the same scored graph.
- Hands-free voice mode — barge-in interrupts the agent mid-response
- Every action emits a replay-able trace for audit
- Bulk-ablate TARSKI-restricted nodes with one command
- Tool registry — engines exposed as callable functions
- Model picker per session — Gemini default; Claude or local Ollama on demand
Criticality in fertilizer supply chains looks nothing like criticality in pancreatic biology — and the same Manifold substrate renders both. Engines plug per domain; the criticality core swaps; the rest of the system stays put. The same codebase already spans domains this far apart, from global supply chains to human biology.
See the engines on real data.
Trial accounts run 48 hours on a curated graph with live macro feeds — FRED, World Bank, NOAA and more.
