Forge Intelligence

Handbook

Reasoning economy (system)

Maximize deterministic control flow; minimize ambiguous LLM spend; allow bounded LLM acceleration where judgment is needed. Tracks reasoning tiers separately from Platform execution worker-ladder steps.

Updated

Intent

Maximize deterministic control flow; minimize ambiguous LLM spend; allow bounded LLM acceleration where judgment is needed. Tracks reasoning tiers separately from Platform execution worker-ladder steps.

Behavior

Default reasoning ladder

Rung 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
human_escalate Lenses approval / operator scarce

Policy binding

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

Ambiguity outcomes

Emit ambiguity_assessment.v1 when deterministic stages abstain. Actions: continue_det, llm_break_in, worker_step, human_escalate.

Limits

Metrics (% zero-LLM sessions, mean tokens per class) are intended — verify pack maturity before production SLO claims.