teleo-codex/domains/internet-finance/reactive-control-from-observed-state-outperforms-predictive-optimization-in-variable-load-systems.md

type	domain	description	confidence	source	created	secondary_domains
claim	internet-finance	Systems with expensive compute and unpredictable demand benefit more from feedback-based scaling than from load forecasting	likely	Corless et al. (SIAM 2016) AIMD principles applied to pipeline architecture	2026-03-11	mechanisms critical-systems

Reactive control from observed state outperforms predictive optimization in variable load systems

For systems characterized by expensive computational resources and variable demand patterns, reactive control mechanisms that respond to observed system state provide better performance and stability than predictive optimization approaches. This is particularly true when the cost of coordination and prediction exceeds the efficiency gains from optimal resource allocation.

The AIMD (Additive Increase Multiplicative Decrease) algorithm demonstrates this principle: rather than attempting to predict future load, model arrival patterns, or solve optimization problems, the system simply observes whether congestion is present and adjusts resource allocation accordingly. When queues are shrinking (no congestion), incrementally add capacity. When queues are growing (congestion detected), aggressively reduce capacity. This reactive approach is mathematically guaranteed to converge to stable throughput without requiring any predictive modeling.

The advantages of reactive control over predictive optimization in these contexts:

1. No modeling overhead: Eliminates the computational cost and complexity of building and maintaining predictive models
2. Robust to distribution changes: Works regardless of how demand patterns shift over time
3. Simpler implementation: Two-rule algorithm versus complex forecasting and optimization
4. Guaranteed convergence: Mathematical proofs of stability versus heuristic tuning
5. Local information only: No need to gather or process global system state

This principle applies broadly to distributed systems where coordination costs are high and demand is unpredictable—from network congestion control to computational pipeline management to energy grid balancing.

Evidence

TCP's dominance as an internet protocol provides decades of empirical validation: AIMD-based congestion control handles the majority of internet traffic without predictive load modeling. The SIAM monograph extends this to smart grid energy allocation and distributed computing, demonstrating the generality of the principle across domains.

For pipeline architectures with expensive evaluation compute (like LLM-based analysis) and variable extraction load, this suggests that reactive scaling based on queue depth observation will outperform attempts to predict optimal worker allocation through demand forecasting.

---

Relevant Notes:

• optimal governance requires mixing mechanisms because different decisions have different manipulation risk profiles
• coordination mechanisms
• trial and error is the only coordination strategy humanity has ever used

Topics:

• domains/internet-finance/_map
• core/mechanisms/_map
• foundations/critical-systems/_map