Forge Intelligence

Handbook

Reasoning economy

Anyone designing or operating problem-class workflows who needs to minimize ambiguous LLM spend while still allowing bounded judgment where it is genuinely needed.

Updated

Who is this for?

Anyone designing or operating problem-class workflows who needs to minimize ambiguous LLM spend while still allowing bounded judgment where it is genuinely needed.

What problem does it solve?

Unbounded prompting is expensive, hard to audit, and easy to over-trust. Forge Intelligence enforces a reasoning ladder: try rules and cheap mechanisms first, escalate only when ambiguity is recorded and policy allows the next rung.

What can you do?

  • Route a session through ordered rungs: deterministic → cheap model → break-in → human escalate.
  • Record ambiguity as a first-class outcome instead of silently expanding prompts.
  • Apply token budgets per stage so spend is visible in the session ledger.
  • Align with Platform resource honesty — abundant resources (rules, local compute) before scarce ones (cloud LLM, human attention).

What can you not do?

  • Skip ladder rungs without policy allowing it.
  • Use open-ended LLM actions — break-in steps use a closed set of allowed actions.
  • Conflate reasoning ladder steps with execution worker-ladder steps — they are tracked separately.

Limitations

Policy and pack maturity vary by domain. A pack may be at contract-level readiness (I1) without domain-validated behavior. Check pack documentation before assuming production-grade judgment.

Next steps