ForgeLLM

System Overview

Ref: ARC-101Last Updated: 03.18.26

ForgeLLM is an enterprise orchestration layer for specialized AI assistants. It provides the runtime, policy engine, and tooling necessary to deploy reliable, deterministic intelligence.

Beta Release Note: v0.9.1

Documentation is currently synchronized with the private beta release. Core orchestration syntax (forge.yml) is subject to minor stylistic changes before RC1.

DAG Architecture

Unlike generic chat interfaces, a Forge worker is a composed DAG (Directed Acyclic Graph) of distinct functional nodes:

  • The Controller: Evaluates incoming objective intent against rigid policy layers before execution begins.
  • The Planner: Computes the required sequence of deterministic tool calls, returning a structured JSON graph.
  • The Executor: Securely dispatches external API requests within isolated memory spaces.
forge_manifest.yml
worker: "financial_reconciliation"
version: "1.0.4"
runtime: "vpc-isolated-us-east-1"

policy_engine:
  strict_determinism: true
  allow_external_http: false
  allowed_tools:
    - ledger_db_read
    - math_verifier_v2

guardrails:
  - no_hallucination_override: enforce
  - human_approval_on_discrepancy: enforce

entrypoint:
  max_recursion_depth: 3
  timeout_ms: 15000

Boundary Rules

To build effective workers on the ForgeLLM architecture, understanding action boundaries is critical. Action boundaries define exactly what state changes a worker is permitted to induce.

Deterministic Guards ensure that given the exact same input state and contextual prompt, the worker will produce structurally identical tool-call graphs. This eliminates the "retry lottery" associated with standard LLM interfaces.