SDLC

ISO/IEC/IEEE 12207 fixes the SDLC as six stages: requirements, design, development, testing and QA, deployment, maintenance, measured by DORA metrics, with one stated purpose, building software correctly and reliably. SFEIR points out what that purpose leaves out. Read against Marty Cagan's Four Big Risks, the SDLC natively covers technical feasibility alone, one risk out of four; value, usability and business viability sit in the PDLC that contains it.

Between May and August 2026 the argument is about execution mode, not sequence. Olivier Rafal (WeNvision) holds that define, build, verify, deploy, maintain stay identical and that AI abolishes no phase, it intensifies each one. Louis Claxton, of Anthropic's Applied AI team, keeps the six stages and turns them into a loop threaded by committed artifacts (`intent.md`, `spec.md`, `plan.md`), with skills as advisory controls and hooks as the deterministic layer that blocks. Jason Clinton, Anthropic's Deputy CISO, describes securing a lifecycle where Claude writes roughly 80% of merged code.

Chris Williams breaks with all of that. The SDLC was shaped over decades to counter human failure modes, ego, fatigue, forgetting, so running it on models is a category error, and his ADLC is built against it. The rest fork instead: SFEIR's cycle à 11 phases, Wescale's Usine Logicielle Augmentée, Thoughtworks' AI/works™, Klaassen's Compound Engineering. Whether these are variants of the same cycle or its replacement is the question none of them settles.