What it was for
The phased lifecycle was a misreading from the moment it was born. Winston Royce's 1970 paper is the one usually credited with waterfall, and it does contain the diagram — but Royce called the pure sequential version risky and said it invites failure. His argument was for iteration between the steps: testing should influence program design, program design should influence requirements. The industry adopted the picture and discarded the warning.
What kept the picture alive for fifty years was economics, not the paper. Specialists were scarce and expensive, and work moved between them. Each pass was slow, and rework that crossed a phase boundary was ruinous, so you put a gate at every boundary to catch errors before they compounded into the next expensive stage. Phases are a scheduling artifact for scarce human labour, and the gates are the interest payment on how costly it was to be wrong late.
The verdict
TRANSFORMS, and the entry turns on separating two things the diagram bundles together.
The dependency order survives. You cannot verify what you never specified, and you cannot specify what nobody has framed. That ordering is logic, not scheduling, and cheap execution does nothing to it. If anything it matters more: when a machine will confidently build whatever you describe, the cost of an unframed intent goes up.
The phase does not. A phase is a bounded period, owned by a group, ending in a handoff and a gate. Every one of those properties exists because human throughput was the constraint. Remove the scarcity and a phase is just a dependency someone drew a box around.
What changes
Both AI-native playbooks kept the boxes. AWS's AI-DLC draws them, and so does Anthropic's playbook (Louis Claxton, August 2026), which lists six stages while its own text says requirements and design "collapse into one session." Six stages where two of them happen simultaneously is not six stages. It is a loop that someone is still drawing as a line.
This is not pedantry about diagrams, because the diagram becomes the org chart. Phases named on a wall turn into teams, teams turn into handoffs, and handoffs turn into a queue. You can make every box take an afternoon instead of a quarter and still rebuild the pipeline, because the shape is what people organise around.
The test for whether a phase has actually dissolved is not how long it takes. Ask instead: does anyone schedule it as a phase, does anyone own it as a phase, and does an artifact cross its boundary? Answer no three times and what you have is a dependency you satisfy, not a stage you enter.
The caution
Do not over-read this into "sequence does not matter." It does. Intent before specification before build before verification is a real dependency, and skipping it produces exactly the fluent, plausible, wrong output the rest of this ledger exists to catch. The verdict is about phases as units of organisation. A loop still has an order. What it does not have is a gate and an owner per segment.
And the compliance guard applies here more than almost anywhere else. Some phase gates exist because a regulator, an auditor or a safety case requires them, not because anyone thinks they are efficient. Those are CONSTRAINED. Freeze them, document why, and audit the rest.