Forge Intelligence

Handbook

General intelligence

General intelligence answers: which problem class, ladder, and budget apply before any domain-specific pack runs?

Updated

Problem-class catalog

packs/catalog.yaml maps problem_class_id → pack directory. Loaded by forge_intelligence.catalog.list_packs().

Reasoning ladder (default)

Rung Typical mechanisms Token posture
deterministic rules, regex, margin gate, FTS, guards 0
cheap_model context packs, local worker, micro-pack bounded
break_in forgeDecide, closed allowed_actions capped calls + facts budget
human_escalate Lenses approval / operator scarce

Spend abundant resources (rules, local compute) before scarce ones (cloud LLM, human attention). Aligns with Platform resource honesty — but tracks reasoning tiers separately from execution worker-ladder steps.

Policy selection

An intelligence_policy_ref (see schemas/intelligence_policy.v1.schema.json) binds:

  • problem_class_id
  • reasoning_ladder ordering
  • token_budget caps
  • allowed_actions for break-in

Consumers pass problem_class_id at session start; general intelligence resolves the pack and records the ladder in route_trace.json.

Ambiguity as first-class outcome

When deterministic stages abstain or disagree, emit ambiguity_assessment.v1 rather than expanding the prompt. Recommended actions:

  • continue_det — stay on deterministic path
  • llm_break_in — invoke governed break-in
  • worker_step — step execution worker ladder (not human)
  • human_escalate — governance boundary

Metrics (intended)

  • % sessions with zero LLM calls
  • mean tokens per problem_class_id
  • ambiguity → escalate rate (separate from worker-ladder steps)