Five stages, four handovers

The five stages run in a fixed order: reconnaissance, definition, delivery, feedback loop, extraction. The order is not a preference. It exists so that each stage has something to check before the next one begins, and so that a hollow project can be stopped at a boundary rather than discovered much later.

Most post-mortems look for the failure inside a stage: a discovery interview that missed the real workflow, a build that shipped without reaching the business endpoint, a metric nobody wrote down. The more reliable place to look is the seam. A stage ends when its exit condition is met, but a handover ends when a named artefact has changed hands. Those are different events, and teams routinely mistake the first for the second.

That is the claim this article defends: stages are structurally safe, seams are not. Inside a stage the work is visible, a failure is local, and the people doing it can see the edge. At the seam, one stage's output becomes another's input, and the only thing carrying it across is a document that either exists or does not.

What each stage must hand over

Stage	Exit condition	Artefact handed to the next stage
Reconnaissance	Seven questions answered or explicitly left blank; three maps complete with at least one named role under "who loses"; work order filed the same day; the opponent from question seven actually contacted, otherwise marked as inference	The six-part reconnaissance work order
Definition	Five columns of the metric table complete; one row locked as the top-priority metric; quote carrying all four blocks and its defensive clause; signed by the customer	Signed metric table, quote and contract annex
Delivery	Prototype meets all five conditions; three environment checks passed in order; evaluation set of at least three cases in the repository; evidence reaches the business endpoint; four-week window not closed early	Evidence chain from system acceptance to business endpoint, plus the observation window data
Feedback loop	Six sections written with a concrete death cause in the sixth; at least one method-library entry with all five fields; three-layer anonymisation passed	Six-section anatomy and the method-library entry
Extraction	Successor manual exists and was written by the successor; weaning complete; four-week rule passed; sunset review archived	Handover manual and archived sunset review

The work order, the signed annex and the handover manual are documents a customer can hold. The evidence chain and the anatomy are records. None of them is a meeting.

Where the seams fail

The reconnaissance-to-definition seam fails when the work order exists but its answers cannot be traced to anything the customer actually does. Seven questions were answered from memory, the three maps were redrawn from the kickoff slide, and the artefact is technically complete. Definition then locks a metric against a workflow nobody observed.

The definition-to-delivery seam fails when the annex is thinner than the conversation. Delivery starts from what the team remembers agreeing to, which is always slightly more than what was signed.

The delivery-to-feedback seam fails when evidence stops at the system boundary. A 200 response proves the system accepted the request and nothing more. A feedback loop fed on acceptance logs will report that everything ran, which is how a project reaches extraction without ever learning which parts of it were used.

The feedback-to-extraction seam fails when the anatomy is written and the handover is improvised. The knowledge exists, in the FDE's head, on the last day.

One diagnostic covers all four: at every boundary, ask for the artefact by name. If the answer is an explanation rather than a document or a data set, the previous stage has not ended.

The health structure

The method carries a rough allocation of project effort: investigate 20, define 15, deliver 40, feed back 15, extract 10. This is a book-internal estimate rather than a measured industry figure, and it reads best as a diagnostic rather than a schedule. If definition is reduced to a short conversation at kickoff, it did not happen. If delivery swallows most of the calendar, the team is redesigning while building, which usually means definition handed over a scope it had not resolved. The two stages most often compressed to nothing are the last two, and they are the two that decide whether the work outlives the FDE.

The one legal shortcut

When the customer's business faces a risk to personal safety or a major public interest, the order can be compressed. The stages cannot be skipped, and the compression has to be recorded in the fourth section of the six-section anatomy, where the judgement was wrong and why speed was worth it. The exception is narrow on purpose: it covers emergencies, not impatience.

What to do at the next boundary

Before starting a stage, name the artefact the previous one owed you. If you cannot name it, write down what is missing and finish the previous stage; time spent there is cheaper than rework built on an assumption. Then put that artefact on the next stage's input list, so the seam has an owner instead of a hope. The stages themselves rarely need defending; the metric table, the quote and the four-week observation are what travel between them.