Forge Intelligence

Intelligence node runtime

Forge Intelligence executes intelligence nodes (capabilities) through typed artifact envelopes — one lmeta unit and one FI-local LCDL contract per capability.

Updated

Architecture

Layer Role
src/forge_intelligence/artifacts/ Envelope schema validation, artifact store, I1 payloads
src/forge_intelligence/orchestration/node_executor.py Thin adapter: lmeta unit + FI contract; substrate via forge-lcdl only
lmeta/units/ Declarative steps: validate_inputs, emit_envelope, fan_in, budget_guard
lcdl/contracts/ FI-local I/O contracts (intelligence_*_v1)
registries/capability_maturity.yaml Per-capability technical + autonomy axes (schema 1.1)
registries/intelligence_maturity.yaml Intelligence + arena rollups (generated)
registries/implemented_allowlist.yaml Runtime gate — only allowlisted capabilities execute

Sessions persist envelopes in artifacts.json ({ "items": [ … ] }). Route trace rungs include technical and autonomy from the registry. See DUAL-WIKI.md and VOCABULARY.md.

Maturity model (dual axis)

Axis Levels Meaning
technical I0–I4 Implementation readiness (scaffold → assured)
autonomy A0–A3 FI decision-autonomy (assisted → monitored)

Autonomy is not Forge L0–L8 execution autonomy (blueprints-owned). See MATURITY-HIERARCHY.md for capability → intelligence → arena rollups.

  • I0_scaffold — unit + contract stub; deterministic envelope emitter
  • I1_runtime — on allowlist; envelope/adapter runtime (schema-valid payloads; contract evidence; tests green). Not domain assurance.
  • A0_assisted — default; human approves each emit
  • A1_supervised — seeded for general.reasoning (13 caps with validators)

Shallow-I1 honesty

Most I1 packs (all life intelligences and general packs except general.reasoning) use a shallow adapter: lmeta units run validate_inputs and emit_envelope with contract-level golden fixtures. Eval suites pass at this level. Do not claim domain-complete reasoning, wired validators, or I2 integration for these packs until validators and semantic fixtures are added per capability.

Nothing left behind: every capability_index.json key has a maturity row (tools/audit_capability_maturity.py).

Reference pack: general.reasoning

All 13 reasoning capabilities are I1_runtime / A1_supervised with golden fixtures under packs/general/reasoning/fixtures/ and coordinator flow lmeta/library/general/reasoning/coordinator.lmeta.

Cross-pack handoff

Case graphs (e.g. reskilling T1–T10) chain nodes via the artifact store. Capability aliases (e.g. build_cashflowbuild_financial_picture) resolve through registries/capability_aliases.yaml.

Bundle source (curated)

Topic Bundle path
Pack contract bundles/human-life-v4.1-audited/core/pack_contract.yaml
Cross-intelligence protocol bundles/human-life-v4.1-audited/core/cross_intelligence_protocol.yaml
Artifact envelope schema bundles/human-life-v4.1-audited/schemas/artifact-envelope.schema.json
Reskilling example DAG bundles/human-life-v4.1-audited/examples/cross_intelligence_reskilling_case.yaml

Full manifests and evaluation suites remain in the bundle; this handbook links product-facing summaries only.

Verification

./scripts/ci-validate-bundle.sh
pytest -q
python3 tools/audit_capability_maturity.py --fail-on-missing --fail-on-orphan-files
python3 tools/generate_intelligence_maturity.py
python3 tools/generate_node_catalog.py