Skip to content

symfonic.capabilities.governance.rulebook

rulebook

The canonical order and the failure-mode declarations (T3.4.4).

This module is the reviewable artefact. Seven concerns, one order, and for each one a written answer to "what happens when this stage itself breaks?". The order is not alphabetical and it is not the order the concerns were built in — each position is argued below.

Why this order:

  1. credential_hygiene first, because every later stage observes the payload. A scrubber that runs after the classifier, the meter, or the detector is a scrubber that runs after three components have already copied the secret into a trace, a prompt, or a provider request.
  2. intent_filter second, because its verdict is an input to fabrication (v7.0.6: an action intent with no tool call is itself a finding). Classifying after the draft exists would make that finding unreachable.
  3. tool_preconditions before policy_steering, because a precondition is a fact about the call ("the cart is not hydrated") and a policy is a fact about the deployment ("nobody deletes without approval"). Steering a call that was never admissible wastes the model's turn on a policy debate about an impossible action.
  4. budget last among the effect stages, because it must measure what actually survives. Metering before the preceding two have narrowed the payload measures a payload nobody is going to send.
  5. fabrication before metacognition, because reflection is the expensive stage: it calls a model. A draft that fails a free deterministic scan should never reach a paid critic, and refusing at fabrication provably stops before the reflection call.