Concepts/Concept
ADPS Agent Systems Blue Book · Engineering Concept
Reasoning Assetization: Move Repeated Decisions into Runtime Assets
Move repeated, reviewable decisions into versioned capability assets before runtime.
Start with GIS publishing
Publishing map data repeats the same format checks, processing, styling, and service configuration. Asking a model to recreate mechanical names, coordinates, and bounds on every run exposes a stable task to fresh probabilistic variation. Xuanxu Technology moved the open-ended work into pipeline design and let runtime consume certified pipelines.
Definition
Reasoning assetization places low-frequency, open-ended, reviewable decisions at design time and stores the conclusions as versioned capability assets. A usable asset has an input contract, version, validation record, lifecycle state, and invalidation signal. Runtime reads certified pipelines, rules, templates, or skill packages instead of repeating the same open reasoning.
Design time: knowledge + examples + model + review -> versioned capability asset
Runtime: input signature -> certified asset -> deterministic execution -> external acceptance
Caching one answer is insufficient. A cache does not define an operating boundary, certification evidence, or version drift.
Boundary
This design fits enumerable task families, repeatable work, high error cost, and results that can be validated offline. Open intent may still require runtime interpretation; mechanical values should continue to come from tools and fact contracts.
Provenance
- Initial case: Xuanxu Technology GIS publishing agent, contributed by Yuke Xiong.
- Lineage: partial evaluation, program specialization, build-time staging, and artifact promotion.
- ADPS contribution: model-assisted generation, human review, certified versions, and operating rights.
- Definition status: case-coined concept.
Initial source: Xuanxu Technology GIS Publishing Agent Chinese case report; contributed by Yuke Xiong.