Inspecting cognition
The system has begun to act, debate, and adapt. Self-reflection adds the ability to inspect those processes and evaluate whether they are working.
Reflection reads reasoning traces, outcome records, confidence estimates, and debate artifacts to evaluate how cognition performed. The goal is not introspective mysticism; it is operational meta-cognition: measuring reasoning quality, identifying failure modes, and proposing improvements under budget constraints. Self-reflection is what turns the activity trace index (created by multi-agent decomposition) from a log into a learning resource.
Why reflection appears early in the self-regulation arc
This stage appears before the full meta-cognitive control layer because the architecture needs a trace-review mechanism early. Later stages (attention, temporal cognition, budget control, affect, and identity) add more surfaces for the system to monitor, but all of them produce traces that reflection will eventually review. Building reflection capability before those systems exist allows traces to be reviewed from the start of their existence.
The Stage 8 reflection loop is intentionally narrow: it reviews concrete reasoning and debate artifacts, then emits bounded improvement proposals. The Stage 20 meta-cognition layer extends this into continuous monitoring, calibration tracking, watchdog agents, and introspection budgeting. Stage 8 is the first reflective instrument; Stage 20 is the mature control layer.
Failure attribution
When an outcome is poor, the question is not just "what went wrong?" but "where in the reasoning process did it go wrong?" Reflection attempts to locate the failure in the cognitive chain.
The attribution categories are distinct and require different remediation:
Missing context means the hydration step retrieved the wrong memories, or no relevant memories existed. The fix is improved retrieval strategy or additional experience accumulation.
Weak plan means the planner's proposal was structurally flawed. The fix is stronger critic participation, additional world-model input, or a prompt adjustment for the planner role.
Overconfident prediction means the world model assigned high confidence to an incorrect prediction. The fix is calibration adjustment and potentially requiring more context before accepting high-confidence predictions.
Poor arbitration weight means the scoring function selected the wrong candidate. The fix is adjusting the weights between reward, coherence, memory alignment, and risk.
External execution failure means the plan was sound but execution failed for reasons outside the cognitive system's control. No internal fix is needed; the system should log the failure and not over-penalize the strategy.
Without attribution, all failures become generic negative reward. The reinforcement engine may punish a correct strategy because it happened to be followed by an external failure, or it may miss a systematic problem with the planner because failures are attributed to execution. Attribution makes reinforcement signals more accurate.
Confidence and calibration
The reflection loop evaluates whether confidence matched outcome. An agent that is wrong but uncertain is safer than one that is wrong and confident. Calibration (the degree to which confidence predicts accuracy) therefore becomes a first-class signal.
A retrieval that returned with high confidence but led to poor reasoning exposes a retrieval calibration problem. A world-model prediction that was highly confident but incorrect exposes a prediction calibration problem. These are separate; they require separate remediation.
Experiment 27 demonstrated the calibration measurement in practice. Three operations were evaluated: a retrieval with stated confidence 0.9 that succeeded: calibration error = 0.1. A planning operation with confidence 0.659 that succeeded: calibration error = 0.341. A tool call with confidence 0.9 discounted to 0.654 (due to risk) that failed: calibration error = 0.654. The mean calibration error across these three operations was 0.614, well above the 0.35 threshold that triggers a watchdog alert.
High mean calibration error is an architectural finding, not just an episode finding. It indicates that the system's stated confidence across multiple operation types consistently overestimates accuracy, which means arbitration (which relies on confidence as a 30% weight) is being misled by inflated confidence signals.
Strategy reflection
Beyond individual failures, reflection can identify reasoning strategies that work well for some task classes and poorly for others. A pattern of high-confidence failures on a specific query type suggests that the world model is poorly calibrated for that type. A pattern of the critic's risk flags being ignored and the action then failing suggests the arbitration weights under-value risk.
These strategy-level findings become recommendations: increase retrieval depth for this task class, require stronger world-model involvement for decisions of this type, weight the critic's risk score more heavily when uncertainty is above a threshold.
Recommendations are proposals. They do not directly modify the architecture. The system emits selfmod.proposed with a structured description of the proposed change, a stability risk assessment, and supporting evidence from the trace history. Constitutional and self-modification stages (described later) determine whether proposals are safe to apply.
The reflection budget
Reflection is not free. Reading traces, comparing predictions against outcomes, running calibration calculations, and synthesizing strategy findings all consume compute. A system that spends most of its time reflecting on its own reasoning has little time for actual work.
Stage 8 begins the discipline of bounded introspection. Reflection runs on a schedule (not after every action), is triggered by signals that suggest it is warranted (a calibration watchdog alert, a large coherence drop, a string of similar failures), and produces a bounded set of outputs rather than an unbounded analysis.
The full meta-cognitive budgeting system (Stage 20) formalizes this discipline. Stage 8 establishes the precedent: reflection is a resource to be allocated, not an unlimited entitlement.