Minimum-state intelligence: the question before the method

Most optimization programmes begin with a method: quantize the weights, prune a structure, distil a model, alter a cache or move state through a different memory hierarchy. The method defines the search space before the system has been fully asked what it can surrender.

Minimum-state intelligence reverses that order. The primary object is not the gain produced by one technique. It is the smallest physical memory state that the complete intelligence can occupy while the declared quality boundary remains intact. Techniques are candidates inside that search, not the authority that defines its limit.

The method is a path. The minimum evidence-valid state is the objective.

A model file is not the same thing as an operating intelligence. At runtime, weights coexist with activations, caches, routing state, temporary workspaces, framework allocations and system overhead. The relevant state depends on the target, workload, batch regime, sequence profile, execution path and physical memory denominator.

For that reason, a minimum-state claim cannot be separated from its conditions. A state that is valid for one workload may not be valid for another. A reduction measured in storage may not be a reduction in peak accelerator memory. A smaller component may produce a larger complete runtime after interactions are resolved.

• Name the target and exact artifact.
• Name the workload and runtime conditions.
• Name the physical memory denominator.
• Judge the composed system, not an isolated component.

No reduction factor is selected in advance. Two times, ten times or one hundred times may become measured checkpoints, but none is the scientific definition of success. The search continues while valid candidate states remain and stops when the available evidence can no longer defend a smaller one.

This makes the objective asymmetrical. Ambition is intentionally unbounded; authority is deliberately bounded. The system may search beyond familiar technique families, but it may only return what its measurements and untouched validation evidence permit.

The objective can be universal without assuming that architectures are interchangeable. Dense, sparse, mixture-of-experts, recurrent, state-space, multimodal and hybrid systems carry different structures and interact with memory in different ways.

A universal system therefore needs a stable mandate and an adaptive search grammar. It asks the same physical question of every architecture, then allows the observed system to determine which mechanisms, interactions and validation paths are relevant.

Architecture-neutral means the objective does not favour a family. Architecture-aware means the evidence never ignores one.

Progress is not the publication of a large ratio in isolation. It is a chain of increasingly defensible states: a measured baseline, a candidate with an explicit denominator, a preserved quality boundary, a complete-system result and a clear statement of what remains unproven.

The final unit of communication should therefore be a bounded claim rather than a slogan. It should identify the system, physical measurement, workload, quality definition, evidence authority and unresolved limitations together.