Capability Resource Theory: What Trading Card Games Teach About AI Systems
Every capability is characterized by persistence, firing count, and independence. The full resource management framework — from opportunity cost to adaptive cycling — derived from TCG mechanics and grounded in 14 of 15 primitives.
Capability Resource Theory
What do trading card games, AI systems, and pharmacovigilance pipelines have in common? They are all resource allocation systems operating under constraint. This framework, derived from TCG mechanics and grounded in 14 of 15 primitives, provides the universal vocabulary for capability management.
Opportunity Cost
Every token spent on X is unavailable for Y. Context windows are finite consumable resources. The only decision is: what is worth more?
Budget is quantity bounded by boundary, allocated via comparison, consumed irreversibly. This is true for mana in Magic: The Gathering, for tokens in an LLM context window, and for attention in a pharmacovigilance signal review queue.
The Capability Persistence Taxonomy
Every capability is characterized by three variables:
- Persistence — how long it lasts (ephemeral vs permanent)
- Firing count — how many times it triggers (single-use vs unlimited)
- Independence — whether it stands alone (independent vs dependent)
The space defined by these three variables contains ALL capability types. Named types — tool, modifier, one-shot, emergent, environment — are specific points in this space. No capability exists outside it.
The Sensitivity Matrix
Every capability has predictable failure inputs (weaknesses) and predictable strengths (resistances).
Effect equals the mapping of input type to capability type, multiplied by base effect. Document this as a matrix: rows are capabilities, columns are input types, cells are amplification factors.
This is FMEA (Failure Mode and Effects Analysis) wearing a different hat. It is also the type effectiveness chart from any RPG. The underlying structure is identical: every entity has a response profile that can be measured and exploited.
Four Scheduling Modes
Resource scheduling has exactly four fundamental modes:
| Mode | Allocation | Timing | Frequency |
|---|---|---|---|
| Breadth-fast | Shallow, spread | Early | High |
| Depth-slow | Deep, concentrated | Late | Low |
| Pipeline-burst | Stored then burst | Triggered | Burst |
| Lean-steady | Efficient minimum | Continuous | Steady |
No single mode is universally optimal. Select dynamically based on task complexity. A signal detection pipeline uses pipeline-burst. A regulatory monitoring system uses lean-steady. A research sprint uses breadth-fast then depth-slow.
Mutual Exclusion
Committing a capability to Task A makes it unavailable for Task B until the next cycle. State cycles: available, then use, then committed, then cycle resets to available.
This is Dijkstra's semaphore in capability space. It is also why multitasking degrades quality — the context switch cost is a real resource expenditure, not just a cognitive illusion.
The Prerequisite DAG
Advanced capabilities require foundational prerequisites in strict order: Knowledge flows through Boundary to Skill, flows through Boundary to Behavior. Each gate is irreversible — you cannot un-learn a skill back into raw knowledge.
Shortcuts that skip stages are possible but risky: they bypass validation gates. This is why the KSB (Knowledge, Skills, Behaviors) framework in pharmacovigilance education requires sequential demonstration — not because bureaucracy demands it, but because the prerequisite structure is real.
Multi-Domain Task Signatures
Every task has a signature in domain-space: the sum of domain requirements at various depths. Cross-domain tasks sum multiple domain requirements. Expertise transfer follows a transfer matrix.
This explains why pharmacovigilance requires both clinical knowledge and data science skill — the task signature spans multiple domains, and no single-domain expert covers the full signature.
Adaptive Cycling
No static strategy survives adaptation. The next strategy must be the counter to whatever is currently dominant. Dominant strategies create predictable counters in a recursive cycle.
This is the Red Queen hypothesis from evolutionary biology. It is Nash equilibrium cycling from game theory. It is the meta-game from competitive TCGs. The structure is identical: in any system with multiple strategies and selection pressure, stasis is extinction.
Build adaptive selectors, not fixed configurations.
Option Value
More relevant options at decision time produces better outcomes. This means the system that preserves optionality — that keeps the most doors open longest — has a structural advantage over the system that commits early.
In pharmacovigilance: maintaining multiple signal detection methods (PRR, ROR, IC, EBGM) rather than standardizing on one preserves the option value of seeing the same data through different boundary operators.
The Unifying Insight
All nine principles reduce to resource allocation under constraint with irreversible commitment. Whether the resource is mana, tokens, attention, or regulatory bandwidth — the mathematics is identical. The domain changes. The structure does not.