The need for selection
A cognitive system with memory, goals, policy, and multiple agents can generate more possible context than any reasoning step can usefully consume. The planner could attend to any of hundreds of retrieved memories. The arbitration system could weigh any of dozens of competing signals. The critic could flag any of many potential risks. Without selection, reasoning degrades as the context becomes too large to reason coherently about.
Attention is the selection mechanism. It decides which memories, goals, signals, interrupts, and agent outputs deserve working-memory budget. Everything else is placed in background monitoring or suppressed.
Salience routing
Attention is not the same as importance. A high-importance memory that is completely irrelevant to the current task should not receive attention. A low-importance memory that directly contradicts the current plan deserves immediate attention. The engine routes items based on salience, which combines multiple dimensions:
Importance: the inherent value of the memory or signal, as scored at ingestion and refined by consolidation.
Urgency: how time-sensitive the item is. A time-critical task approaching its deadline has high urgency regardless of importance.
Novelty: how unexpected or surprising the item is given current context. A genuinely novel signal may indicate a situation the system's current model cannot explain.
Goal relevance: how directly the item relates to an active goal. Items aligned with high-priority goals receive additional salience.
Affective weight: how the affect system modulates attention. In a high-alert state (high norepinephrine), risky and urgent items receive additional salience. In a settled state, the same items receive normal salience.
Risk: items that signal potential harm or constraint violation receive elevated salience even when importance is low.
The calibration failure that shaped the engine
Experiment 6 tested the attention engine before calibration and found a systematic problem. The engine assigned a novelty coefficient of 0.14, meaning novelty could contribute at most 14% of the total score times a factor that was at most 0.9, for a maximum novelty contribution of 0.126. The importance coefficient was 0.35.
The result was that importance dominated attention, and the engine and the test harness disagreed 40% of the time on which items should be selected. The harness expected novelty to have more influence; the engine's weights made that impossible.
The fix was straightforward: raise the novelty coefficient from 0.14 to 0.30 and reduce the importance coefficient from 0.35 to 0.29. After the adjustment, engine and harness agreed 100% of the time at all tested novelty settings.
This calibration episode illustrates a general principle: attention systems need explicit weight calibration against expected behavior. The coefficients cannot be set by intuition alone because their interactions are not intuitive. A novelty weight that seems adequate in isolation may be overwhelmed by importance in practice.
Working-memory budget
Working memory is a bounded resource. The attention engine enforces that bound by assigning budget to candidate context items. Items that clear the salience threshold receive budget allocations. Items that do not are either placed in background monitoring (their state is tracked but they do not receive reasoning attention) or suppressed entirely.
Budgeting makes cognition inspectable in a way that unlimited context does not. When an item is excluded from the reasoning context, the exclusion is explicit and attributable: "this memory did not clear the salience threshold because its importance score was 0.3 and it had no goal alignment, no urgency, and low novelty." That is a diagnostic statement. "This memory was not used because it fell outside the context window" is not.
Explicit budget enforcement also protects the reasoning step from degradation. Retrieval and agent outputs can expand without limit when salience is unlimited; budget enforcement prevents the reasoning prompt from becoming incoherent.
The affect-attention coupling: a quantitative finding
Experiment 21 quantified one of the most important couplings in the attention system: the relationship between the affect state and the attention boost for urgent, high-risk candidates.
In a settled state (normal baseline after processing routine signals), the coupleAttention boost for a high-risk, high-urgency candidate was 0.030.
In a stressed state (after ten consecutive outage signals, norepinephrine = 0.861, contradictionStress = 0.895), the same candidate received a boost of 0.334.
That is an 11-fold difference. In a stressed state, the attention system prioritizes risky and urgent items more than eleven times as strongly as in a settled state.
This is the intended behavior. When the affect system signals high alertness and stress, the cognitive priority should shift toward dangerous and time-critical items. Routine memories should fade to background while incident signals come to the foreground. The norepinephrine and contradictionStress channels in the boost formula encode this shift mechanically.
The design implication is that affect modulation is not cosmetic; it directly and measurably changes which items receive cognitive resources. An outage does not just change the system's "mood." It restructures the attention allocation in a quantifiable way.
Focus persistence and avoiding thrash
Attention should not thrash. A system that switches focus every turn, chasing the highest momentary salience signal, has no ability to sustain work on a complex task. The engine includes persistence and decay mechanisms so that an actively attended item retains some advantage in the next turn even if a moderately competing signal arrives.
Persistence means the attention score for an item in active focus is augmented by a continuity bonus. Decay means the bonus diminishes over time if the item is not re-attended. Together these create a regime where sustained focus is possible and can resist moderate competing signals, while being breakable by sufficiently urgent interrupts.
This balance supports both concentration (staying on a hard problem long enough to make progress) and responsiveness (shifting when something genuinely important demands it). Without persistence, the system would be reactive. Without decay, it would be rigid.
Interrupt lanes
Some signals must preempt ordinary attention competition. A safety constraint violation, a critical system failure signal, or a constitutional quarantine trigger should not wait for the next attention cycle; it should insert immediately.
Interrupt lanes are priority queues that bypass the normal salience competition. An item in an interrupt lane is attended regardless of what is currently in focus, and the current focus item is paused (its state preserved for resumption) rather than discarded.
Interrupt lanes are deliberately limited to signals that cannot safely wait. Overusing them would degrade the focus-persistence mechanism. The distinction between "urgent enough to interrupt" and "important enough to compete" is itself a design decision that shapes how the system responds under pressure.