Gorde:
| Egile nagusia: | |
|---|---|
| Formatua: | Recurso digital |
| Hizkuntza: | |
| Argitaratua: |
Zenodo
2026
|
| Gaiak: | |
| Sarrera elektronikoa: | https://doi.org/10.5281/zenodo.20380892 |
| Etiketak: |
Etiketa erantsi
Etiketarik gabe, Izan zaitez lehena erregistro honi etiketa jartzen!
|
Aurkibidea:
- <p>This paper proposes Field Engineering as an emerging scientific and practical discipline for designing, testing, promoting, governing, correcting, and retiring operational fields before intelligence becomes action. The discipline is grounded in our recent work on the Two-Field Principle, the LLF Existence Problems, the field above the model, the Math of Fields, and the Field Machine. Its central claim is bounded: field engineering does not assume that every proposed field is real, valid, or ready to govern. Instead, it treats fields as operational candidates that must be specified, varied, tested, falsified, promoted, governed, corrected, or retired. The paper defines the Field Contract as the basic unit of the discipline and introduces a lifecycle: Field Design → Field Candidate → Field Trial → Field Promotion → Field Governance → Return → Redesign or Retirement. It distinguishes Field Engineering from prompt engineering, context engineering, RAG engineering, agent engineering, software architecture, and AI governance, while showing how it integrates them into a broader regime of passage. Practical examples include institutional RAG, tool-using agents, legal interpretation, clinical escalation, hiring support, financial boundaries, production release gates, and customer resolution. The paper argues that the next phase of AI reliability will require not only better models, prompts, retrieval, tools, and policies, but engineered fields that determine when intelligent candidates may become trusted artifacts, decisions, actions, or HOLD. This requires a shift from improving isolated components to governing the passage of candidates across fields. A model may generate a stronger answer candidate, a prompt may shape it more clearly, retrieval may supply more relevant evidence, tools may expand what the system can execute, and policies may define acceptable use. Yet none of these alone determines whether the candidate deserves promotion. Field Engineering addresses this missing layer by designing the regimes through which candidates must travel before they acquire higher status. An answer candidate must survive evidence and authority fields before becoming trust. A tool-call candidate must survive permission, reversibility, and audit fields before becoming action. A classification candidate must survive fairness, contestability, and consequence fields before becoming institutional judgment. A tool-using agent candidate must survive an Agent Gateway field before becoming operational action, because tool availability is not permission, API execution is not authority, and logs are not yet receipts. The gateway governs whether an agentic request may be promoted, held, returned, rejected, or used to trigger redesign of the surrounding workflow. In this sense, AI reliability depends not only on better generation, retrieval, reasoning, or tool use, but on engineered passage: the disciplined movement from candidate formation to promotion, HOLD, correction, rejection, return, or redesign.</p>