Meta-cognition engine
The subsystem that monitors reasoning quality, measures calibration error, attributes failures by operation type, and emits self-modification proposals. Operates under a budget constraint: introspection has a cost and must be economically justified like any other cognitive operation.
The meta-cognition engine is the mature control layer that the reflection loop grows into: where early reflection reviews specific traces in response to failures, the meta-cognition engine runs continuous monitoring across all cognitive operations with dedicated watchdog agents for each key signal. It tracks calibration error per operation type (retrieval, planning, prediction, execution) and computes rolling means that reveal systematic overconfidence or underconfidence patterns. It attributes failures to the specific cognitive step where they originated - missing context, weak plan, overconfident prediction, poor arbitration weight, external execution failure - producing actionable findings rather than generic negative reward. It manages introspection budgets so that monitoring cost stays proportional to operational value: a well-performing system with low calibration error and stable policy drift gets minimal introspection overhead; a system showing degrading patterns gets increased monitoring. Proposals from the meta-cognition engine route through the constitutional engine before taking effect, just like any other self-modification proposal. The watchdog agents it deploys - calibration watchdog, identity drift watchdog, retrieval quality watchdog - each monitor one signal with one threshold and emit structured alerts when that threshold is crossed.