At a glance
- What it is: the design for the kernel’s capacity layer (how much, and when to stop)
- Status: Parked. Designed, not built; activation is gated on S8
- Depends on: real tracking data from Metrics & status tracking
- Serves: the kernel
The mental model is Kubernetes: capacity is admission control plus eviction. The weekly budget is node capacity; every active entity declares a resource request; the scheduler refuses to overcommit; overload triggers priority-ordered eviction. Three mechanisms over existing primitives:
1. Declaration (structure layer carries it). Active entities declare a capacity block in frontmatter: cost (coarse weekly units — allocated slots, not tracked hours), optional peak: true, and priority (eviction order). Instance config declares the budget: units per normal week, reduced budgets for travel/crunch weeks. Costs are estimates by design — the point is making the sum visible, not accounting.
2. Admission (flow layer enforces it). Capacity is checked when something tries to become active, never audited after. At the weekly review, promoting/activating is a scheduling request: a capacity check sums active costs against the week’s budget and answers “fits” or “over by N — pause something or don’t admit.” Sequencing is one more admission rule: at most one peak: true per period. “Everything defaults to NO” stops being a vow and becomes a computed answer.
3. Eviction (capacity layer’s runtime). The signals→downshift evaluator, upgraded from warning to action list: when the rule fires, emit the lowest-priority active entities in eviction order with reclaimed units (“pause side-hustle (+2)”). v1 of this prints the list; actual state-flipping and calendar teeth are a later decision (pending O-A).
Deliberate omission: no actual-hours tracking, ever — allocation-based capacity (requests, not usage) gives ~90% of the protection for zero daily effort, and the burnout signals catch allocation/reality divergence; that’s what they’re for.
Free byproduct: snapshots already record state, so allocated-vs-budget per week becomes a historical chart — pressure curves visible in hindsight, feeding calibration (O-D).
Activation gate (S8): do not implement until tracking has produced enough real weeks to set the budget, the unit, and per-entity costs from evidence.