This note proposes a working lens — not a law — for comparing systems that produce outcomes through constrained agents.
Let:
- CORE (C) = intrinsic capability under ideal conditions (skill, compute, model quality, institutional power)
- GIFT (G) = resources transferred across boundaries (attention, capital, trust, context, training data)
- H = entropy / friction / loss (coordination tax, misalignment, decay, adversarial noise)
The expression
Effective output ≈ (C × G) − H
Multiplication matters: high capability with zero gift is inert. High gift with low capability dissipates. Subtraction is relentless: entropy does not negotiate.
Design readings
- Optimization trap: Teams often maximize C while H grows invisibly (process, politics, tech debt). Output flatlines.
- Gift without structure: Injecting resources without interfaces increases H faster than G.
- Entropy-aware leadership: The best operators spend as much energy lowering H as raising C.
AI translation
For LLM systems:
- C ≈ base model + tools + eval discipline
- G ≈ retrieval context + user intent + organizational knowledge shared with the runtime
- H ≈ prompt ambiguity + stale data + safety filters misfiring + latency-induced truncation
Improvement is not always “bigger model.” Sometimes it is gift routing (better context) or entropy surgery (clearer schemas, tighter feedback loops).
Status
Framework in revision. Counterexamples welcome. The goal is a portable vocabulary, not a manifesto.