Architecture

Layered model

Consumers (Lenses, KA, Cockpit, Campaigns, Dark Factory)
        |  problem_class_id / intelligence_policy_ref
        v
forge-intelligence
  bundles/       human-life-v4.1-audited (immutable vendor drop)
  registries/    capability_index, adjacency_graph, allowlist
  orchestration/ router → composer (DAG) → node_executor (thin adapter) → session M/H
  artifacts/     typed envelope store (artifact-envelope schema)
  lcdl/          FI-local contracts + validators (NOT forge-lcdl)
  lmeta/         FI-local flows/units (NOT forge-lcdl)
  packs/         runtime policy entrypoints
  wiki/          concept dual-wiki (semantic + execution projections)
        |  composes (does not fork)
        v
forge-lcdl — tasks, operators, FlowExecutor, forgeDecide
        |
        v
Evidence attach — route_trace, token_ledger, ambiguity (schemas v1 + v2)

Three dual-wiki surfaces

See DUAL-WIKI.md, VOCABULARY.md, and INTELLIGENCE-NODE-RUNTIME.md.

Artifact envelope bus

Capabilities (intelligence nodes) exchange artifact envelopes only — see bundle schema schemas/artifact-envelope.schema.json. Session artifacts.json holds { "items": [ … ] }. Maturity registry (registries/capability_maturity.yaml) tracks technical (I0–I4) and autonomy (A0–A3) per capability; rollups in intelligence_maturity.yaml. for all 297 capabilities; runtime execution is gated by implemented_allowlist.yaml.

Non-linear orchestration (graph, not ladder)

Multi-pack cases use a task graph with depends_on, parallel_group, and fan_in_gate. The router returns a set of intelligences (not a single winner); general.reasoning acts as coordinator when multiple packs activate.

Router (activation / abstain / adjacent)
        |
        +-- general.reasoning --+
        |                      |
        +-- life.* packs ------+--> TaskGraphComposer (DAG)
                                      |
                                      v
                               CapabilityRunner + maturity gate
                                      |
                                      v
                               Session dual-wiki M/H

Modules: forge_intelligence.orchestration.{router,composer,maturity,runner}.

Orthogonal indexes (hierarchy)

Inward-dependency rule

forge-intelligence  --depends on-->  forge-lcdl (substrate tasks/operators)
        (owns: bundles, manifests, FI-local lcdl/lmeta, orchestration, session dual-wiki)

forge-lcdl does NOT receive intelligence_* contracts or pack manifests.

Sessions (`sessions/<id>/`) are **non-authoritative projections** — attach to ForgeRun/Lenses
as evidence; see [SESSION-DATA-GOVERNANCE.md](SESSION-DATA-GOVERNANCE.md) and ADR-0002.

General intelligence

Answers: which intelligences and task graph apply?

Reasoning ladder (default):

1. deterministic — validators, guards, substrate reuse
2. cheap_model — bounded context packs
3. break_in — closed-set forgeDecide
4. human_escalate — governance boundary

Specific intelligence

Each runtime pack under packs/<domain>/ contains policy.yaml, optional flows/, and references to FI-local lcdl/contracts/ and lmeta/units/.

Human-Life pack: 39 manifests in bundles/human-life-v4.1-audited/manifests/.

Schemas

v1 (session attach): schemas/*.v1.schema.json

v2 (manifest + graph): schemas/v2/

• intelligence_manifest.ref.schema.json
• task_graph.schema.json
• maturity_axes.schema.json
• artifact_ref.schema.json

Python package

Handbook pair

Product Markdown in fi; HTML deploy in fiw. See HANDBOOK-PAIR.md.