Darkstar platform
Turn questions into models—and models into inspectable decisions.
Combine domain expertise and supported data workflows with probabilistic inference, causal analysis, and integration paths that keep assumptions and results visible.
A connected path, not a forced sequence
Data and expertise → models → inference → applications and agents
Begin with visual authoring, supported learning, or a model you already have. Some causal methods work directly from data; some learned structures need parameters and assumptions before inference. Darkstar keeps those handoffs explicit.
Pillar one
Build and learn models
Turn expertise and supported data workflows into explicit model structure and parameters. Visual authoring, statistical learning, prompt-generated candidate discovery, and synthetic sampling are different operations with different evidence.
- Input
- Domain knowledge, a supported dataset, or an existing model.
- Operation or question
- Author or refine a family-native model; use a supported learning workflow where applicable.
- Output
- An inspectable model or learned artifact with identity, assumptions, parameters, and limits.
- Useful next step
- Review the model, complete required assumptions or parameters, and choose a supported query.
See proof-backed learning workflows, their inputs and artifacts, and the current hosted-access boundary.
Pillar two
Understand what is likely—and how uncertain it is
Use evidence updates, posterior distributions, prediction, state estimation, and family-appropriate queries. Probabilistic inference answers questions under the declared model; it does not establish that an association is causal.
- Input
- A compatible model, evidence, target, and supported inference configuration.
- Operation or question
- Condition on evidence and run the model-appropriate probabilistic query.
- Output
- A typed result with probabilities or distributions, uncertainty, method identity, diagnostics, and provenance.
- Useful next step
- Inspect the result, revise evidence or assumptions, compare scenarios, or export the validated model.
The catalog distinguishes 31 fine-grained semantics from the 22 broad Online family guides.
Pillar three
Compare actions—and learn from what happened
Apply supported interventions, causal-effect methods, diagnostics, sensitivity analysis, and case-specific counterfactuals while keeping factual evidence, alternative actions, targets, and assumptions explicit.
- Input
- A compatible causal model or dataset, a supported question, and the assumptions required by that method.
- Operation or question
- Compare an intervention, estimate a supported effect, or evaluate a counterfactual alternative.
- Output
- A result with the estimand or target, uncertainty where supported, diagnostics, provenance, and explicit limitations.
- Useful next step
- Challenge assumptions, compare an alternative, or carry the result into a reviewed decision workflow.
Navigate 101 authority-derived identities by estimation, design, diagnostics, mediation, policy, and other tasks.
Pillar four
Bring model-based reasoning into applications and agent workflows
SDKs embed supported capabilities in customer applications and services. Agent interfaces can expose supported model operations and structured results without becoming a second inference engine or guaranteeing an autonomous decision.
- Input
- A validated model, a supported operation, an authorized integration, and an applicable runtime surface.
- Operation or question
- Invoke the operation through a maintained SDK today, or through an explicitly available agent interface.
- Output
- A typed result, refusal, or durable job identity with relevant diagnostics and provenance.
- Useful next step
- Inspect the result in the visual workflow, handle typed limits, and keep product-specific access boundaries intact.
Compare SDK embedding with the MCP interface and its authorization boundary.
For recurring organizational questions
Make the reasoning reusable—not invisible.
Use the visual workspace for review, supported SDKs for application integration, and explicit model operations for services and agent workflows. The result is an inspectable reasoning contract, not a turnkey answer for every organization.
Illustration: a project delivery question
- Assess modeled delivery risk from current evidence.
- Compare a declared capacity or dependency intervention.
- After an actual overrun, examine what an earlier decision might have changed.
This is an illustrative question sequence, not a delivered project-management module or measured ROI.