Arxi Decision Gate Asset Core
Docs
Asset Core Docs

Deterministic world-state engine documentation and API references.

Decision Gate docs

Agent Reasoning Stack

This page summarizes the four-layer reasoning stack we give agents during Asset Core runs so decisions stay deterministic, auditable, and aligned with the operation catalog.

Think of it as the operational view of the mental model introduced in the Docs overview and Concepts.

Each layer is intentionally narrow: the agent only needs the layer that matches the current question.

The full reasoning corpus is a living set of documents; scenario manifests pin the exact revision used for any run pack.

The four layers

1) Reasoning (Conceptual Layer)

Defines the world model and invariants: namespaces, containers, determinism, and fixed-point rules.

  • Focus: why the system behaves the way it does.
  • Outputs: invariants that must never be violated.

2) Ontology (Entity Types and Fungibility)

Defines what exists: classes, instances, fungible quantities, and which container kinds can hold them.

  • Focus: what entities and containers mean.
  • Outputs: valid representations for a domain’s entities.

3) Decision (Tool Selection)

Maps intent to concrete reads, writes, and error handling rules.

  • Focus: which tool or operation to call next.
  • Outputs: a safe commit plan and required preflight reads.

4) Pattern (Operational Playbook)

Turns decisions into step-by-step sequences with verification checks.

  • Focus: how to execute common workflows correctly.
  • Outputs: commit sequences and verification steps.

How the stack is used

Agents move from intent to commit by walking the stack in order.

  1. Reasoning: validate invariants and model assumptions.
  2. Ontology: choose entity representation and container kinds.
  3. Decision: pick tools and structure the commit.
  4. Pattern: execute and verify.

What the agent receives

For scenario runs, we provide the four documents below plus the Action Reference and Tooling Manifest.

These downloadable guides are snapshots from production runs. They may lag current documentation and are not normative; use the Action Reference and Tooling Manifest for the canonical contract.

How to read the stack

If you are new, read Reasoning, then Ontology, then Decision, and finish with the Playbook.

If you are diagnosing a run, jump to the layer that matches the failure mode and verify the surrounding constraints.

Legitimacy notes

Scenario manifests record the commit SHA for the reasoning corpus used during the run so readers can inspect the exact inputs the agent saw even as the documents evolve.