Can the Question and Judgment Show Why They Still Hold? Version 1.3 — The LimFlex Judgment Terrarium Ecosystem

Large language models can generate useful candidates, but organizations also need to preserve what currently holds, what actually changed, what remains unresolved, and who is authorized to commit or execute. This paper presents the LimFlex Judgment Terrarium Ecosystem, an externalized judgment environment in which qualified judgment is maintained outside any single AI model as a state that can be shared, inherited, reused, selectively updated, requalified, and audited.

The standard engineering perspective treats qualified current state, candidate semantics, claimed transition, the actual prior-to-candidate relation, and semantic Question Adequacy as distinct qualification objects. The main evidence is a prospective Corrected Fresh product qualification frozen before candidate generation.

Across 192 fresh slots produced by GPT-5.6 Sol, Claude Sonnet 5, and Gemini 3.8 Flash, structural transition admission was 192/192, the Qualified Action Envelope and Qualified Routing Envelope were each 192/192, semantic-coordinate coverage was 191/192, and the strict combined product outcome was 191/192. Across two fresh passes, exact action agreement was 87/96, while structural admission agreement was 96/96 and strict qualified-outcome agreement was 95/96. Thus, action paths varied while the qualified external boundary remained substantially more stable.

Median transition qualification time was 0.287 ms versus 4.762 s median provider latency, approximately 0.0060%. Supporting deterministic evidence includes 1000/1000 exact replays, 1000/1000 historical-version preservation, 25/25 lineage hops, zero protected-state drift, zero authority leakage, and 10/10 recovery.

The evidence is internal and bounded; it does not establish external interoperability, security certification, population-level model reliability, or real-world outcome correctness.