Documentation

Core Object Model

TENSOR investigations are represented as directed graphs of nodes and edges. Nodes represent investigative questions. Edges represent branching decisions: yes, no, or unknown.

A graph package includes metadata (namespace, product, version, schemaVersion) plus nodes and edges.

Node Contract

Each node contains a unique id, full question text, and a domain category. Optional fields include short label, localized translations, and custom extensions metadata.

Node IDs follow the pattern Q<number> and must remain stable across versions unless explicitly deprecated.

Edge Contract

Each edge contains source, target, and a decision. Edge IDs encode the same relationship (example: Q12-yes-Q1).

This ID encoding is deliberate: it makes graph diffs and integrity checks deterministic in CI/CD pipelines.

Execution Semantics

The model supports loops so investigations can return to shared questions when new evidence appears. This avoids duplicating logic across branches.

Conventional playbooks are often mostly linear with occasional decision points and loop-backs. TENSOR captures that baseline behavior while also supporting cross-branch reuse and reconvergence.

• Branching: each decision is explicit, never implicit.
• Re-entry: cycles are valid and expected for iterative investigations.
• Unknown handling: uncertainty is first-class via unknown edges.
• Traceability: every traversal can be reconstructed from node/edge IDs.

Versioning And Compatibility

Minor versions can add categories, nodes, and edges without breaking existing implementations. Breaking changes are reserved for major versions and require migration notes from the consortium.

Implementers should pin schemaVersion and validate graph packages in CI before release.

Extensibility Model

Organizations can add custom metadata in extensions while preserving compatibility with the core schema. Use namespace-scoped keys such as acme.confidence or acme.controlMapping to avoid collisions.

LLM Integration Guidance

TENSOR is designed to let AI assist analysis without allowing AI to redefine workflow semantics. Keep graph logic deterministic and treat model output as evidence, not policy.

• Run prompts against stable node IDs, not free-form step text.
• Capture model output with confidence metadata in extensions.
• Require explicit branch decisions before state transitions.
• Use repeated sampling plus adjudication for high-impact decisions.

Related Pages

For implementation details on probabilistic AI behavior and mitigation patterns, see AI Reliability.