Verification-Bounded Stage Gates: A Phase-Weighted Readiness–Risk Index for LLM Adoption Across the Software Development Life Cycle

Large language models (LLMs) now touch every phase of the software development life cycle (SDLC), yet most adoption guidance remains organization-wide and qualitative. A recent IEEE review by Radenković et al. synthesized LLM use across the SDLC and proposed a framework for assessing organizational readiness for LLM adoption. This paper extends that work in three directions. First, we decompose readiness to the level of individual SDLC phases aligned with ISO/IEC/IEEE 12207, because the evidence base shows that both benefit and risk vary sharply between, for example, requirements elicitation and production deployment. Second, we pair readiness with an explicit, phase-weighted risk-exposure term and combine the two into the LLM–SDLC Readiness–Risk Index (LRRI), which ranks phases for investment. Third, we convert the assessment into operational governance through verification-bounded stage gates that map each phase’s readiness, risk exposure, and verification capacity to one of four graded autonomy levels, under the principle that the autonomy granted to an LLM in a phase must not exceed the organization’s capacity to verify that phase’s artifacts. We illustrate the framework on three synthetic organizational archetypes and run a 90,000-draw Monte Carlo sensitivity analysis. Within the model, verification capacity blocks advancement from assistive to suggestive use in 11 of the 12 phase–archetype cells held at the assistive level, and gate decisions are markedly more sensitive to score-elicitation error than to disagreement over weights. We close with an empirical validation protocol based on a Delphi panel and DORA delivery metrics.