From reflection to persistent meta-cognition
Stage 8 introduced self-reflection: a mechanism for reviewing reasoning traces and emitting bounded improvement proposals. Stage 20 extends this into a persistent runtime capability with four components that work together: confidence estimation per operation type, calibration monitoring across operations, failure attribution by cognitive step, and introspection budgeting with watchdog agents.
The difference between Stage 8 and Stage 20 is not just more capability; it is continuous monitoring versus episodic review. Stage 8 reflection is triggered by events. Stage 20 meta-cognition runs continuously, tracking cognitive operations and their outcomes, and can intervene before a failure compounds into a larger problem.
Confidence estimation per operation type
The meta-cognitive engine attaches confidence estimates to specific operation types: retrieval, planning, prediction, critique, arbitration, and reflection itself. Rather than a single overall confidence measure, each operation type has its own confidence history.
This granularity matters for attribution. A system that fails on a planning step after confident retrieval has a different problem from one that fails after low-confidence retrieval. The first may indicate that the planner is over-relying on accurate memory; the second may indicate that retrieval quality is the upstream cause of the planning failure.
Experiment 27 illustrated this with three operations:
- A retrieval stated at confidence 0.9 that produced a failed outcome: calibration error 0.848.
- A planning step at confidence 0.659 that produced a successful outcome: calibration error 0.341.
- A tool call at confidence 0.9, discounted to 0.654 due to high risk (riskScore = 0.75), that failed: calibration error 0.654.
Mean calibration error across the three: 0.614. The watchdog threshold was 0.35. A mean of 0.614 is well above threshold and triggers an alert: calibration_drift_detected.
The discount applied to the tool call confidence (0.9 reduced to 0.654 because riskScore = 0.75) is the meta-cognitive system applying prior knowledge about uncertainty: a high-confidence prediction that involves a high-risk action should be treated with less confidence because risk introduces irreducible uncertainty. This is calibration adjustment before the fact, not just measurement after.
Failure attribution
When a high-confidence operation fails, meta-cognition runs attribution to locate the failure. The attribution in Experiment 27 for the failed tool call (confidence = 0.9, riskScore = 0.8, zero retrieved memories): failureAttribution = risk_underestimated. The system had stated high confidence on a risky action with no supporting memories, and it failed.
The attribution categories align with those introduced in the reflection article:
context_missing: relevant memories were not retrieved, or the memory base lacks coverage of this situation.risk_underestimated: the world model or the operation's own confidence estimate did not adequately account for risk.plan_flawed: the planner's strategy was structurally inadequate for the task.calibration_error: the system was consistently wrong while confident; the confidence measure needs adjustment.external_failure: the action was correctly planned and executed but failed due to environmental factors outside the cognitive system's control.
Attributing failures correctly is what prevents the reinforcement engine from penalizing correct strategies. If a well-reasoned action fails due to external factors, the attribution external_failure prevents that failure from propagating as a negative reinforcement signal to the memory and policy systems.
The self-modification proposal chain
When attribution finds a structural issue, the meta-cognitive engine emits a self-modification proposal. Experiment 27 produced a proposal with type strategy_adjustment, a stabilityRisk of 0.8, and supporting evidence from the failed high-risk tool call.
The stabilityRisk field is important. A self-modification proposal that would make a large change to reasoning strategy carries high stabilityRisk because the consequences are uncertain. A proposal that would make a minor adjustment (increase retrieval depth for high-risk actions) carries lower stabilityRisk.
The constitutional engine (described in the next article) uses stabilityRisk as a gating factor. High-stabilityRisk proposals are quarantined for review; low-stabilityRisk proposals may be applied automatically. This is the safety mechanism that prevents the meta-cognitive system from rewriting the architecture based on a few anomalous failures.
Recursive budget and the recursion trap
Meta-cognition consumes resources. Running calibration checks, attribution analyses, and proposal generation after every operation would consume more resources than the operations being evaluated. A system that spends all available compute evaluating its own evaluation has no capacity for actual work.
The engine enforces introspection budgets. Reflection can run when budget is available and the expected improvement from reflection exceeds its cost. This mirrors the cognitive economics logic applied to all operations, applied recursively to meta-cognition itself.
The recursion trap is real: meta-cognition can be applied to the meta-cognitive operations themselves, producing an infinite regress of self-evaluation. The budget enforcement cuts this regress by treating meta-cognitive operations the same as any other operation: they are approved when expected utility exceeds cost, and rejected otherwise. Meta-cognition does not receive a special exemption from the economics it is supposed to be monitoring.
Watchdog agents
Watchdog agents monitor for unsafe drift, repeated calibration failure, or stalled cognition without requiring explicit triggers. They run in the background, watching for specific signal patterns that indicate problems:
Calibration drift: mean calibration error rising above threshold across multiple operations. Triggers a review request and may reduce the weight given to confidence estimates in arbitration.
Repeated attribution failure: the same failure mode attributed consistently across many operations without improvement. Indicates the system is not learning from its mistakes, which may require a deeper structural change.
Stalled cognition: the system is generating outputs but all recent operations are failing or producing low-quality outcomes. May indicate that the current context is inadequate and a reset or re-retrieval is needed.
Watchdog agents can trigger reflection, reduce autonomy (requiring human confirmation before high-risk actions), or request stricter arbitration (raising the confidence threshold before approving actions) depending on policy.
This creates a runtime safety surface before constitutional stability is applied. The constitutional engine handles structural violations; the watchdogs handle performance degradation before it reaches that level.