Theory Advanced 1 minute read Updated 2026-06-26 UTC

Benefit-centered viability

A revised viability frame that gives positive human and institutional gains first-class weight in model-breeding decisions.

Research statusConceptual synthesis Publication statePublished Reviewed byMichael Kappel Source reports3

Why viability needs a benefit term

Earlier viability pages emphasize predictive gain, robustness, diversity, memory, latency, energy, and risk. Version 1.5.0 adds a positive emphasis: human and institutional benefit should be a first-class term, not an afterthought.

A model can be accurate and cheap while still not worth deploying if it does not improve the user's actual situation. Conversely, a model with modest benchmark gain may be highly valuable if it transfers skill, reduces drudgery, or preserves critical knowledge.

Benefit-centered score

TermMeaning
ΔTaskUtilityChange in task success.
ΔHumanCapabilityChange in user competence after assistance.
ΔReuseHow often outputs become durable artifacts.
ΔAutonomyExport, override, and migration health.
ΔEfficiencyResource savings or cost reduction.
pseudocode
FUNCTION benefit_centered_viability(candidate, incumbent)
    positive = delta_task_utility(candidate, incumbent)
    positive += delta_human_capability(candidate, incumbent)
    positive += delta_reuse(candidate, incumbent)
    positive += delta_autonomy(candidate, incumbent)
    positive += delta_efficiency(candidate, incumbent)

    costs = delta_latency(candidate, incumbent)
    costs += delta_memory(candidate, incumbent)
    costs += delta_maintenance(candidate, incumbent)

    RETURN positive - costs
END FUNCTION

Theory implication

A descendant should be promoted when it improves the lived ecology around it, not only when it improves a benchmark number.

Source reports used for this guide

These reports are preserved verbatim in the site archive. The guide above is an editorial synthesis and may narrow, qualify, or reorganize claims from the source material.