Tier 3 · Lifecycle
Configured work types are a hypothesis about the work; the matters are the evidence. When one type absorbs two-fifths of the book, it is usually hiding several distinct kinds of work that deserve their own routing, reporting, and resourcing. And when a phase quietly holds matters for years, the phase plan is leaking. This audit reads the titles to find the taxonomy mismatch, then reads the phase history to put observed numbers (median ages, reopen rates, close velocity) on the lifecycle.
On-premise embedding analysis of matter titles segments a bloated configured type into the linguistically distinct kinds of work living inside it, each a candidate for its own routing, reporting, and resourcing.
Every phase with open work gets a median-age profile, and any phase past the 365-day threshold is flagged. In the demonstration corpus, two closing-and-reopening queues held matters at median ages of roughly four and five years.
The share of closed matters later reopened, and median open-to-close duration per work type. A wide velocity spread inside one type is itself evidence of dissimilar work sharing a single configuration.
Multiple start phases, unreachable phases, and a missing mandatory Draft phase: structural defects that confuse routing, reporting, and automation.
Work types holding no matters at all, and forms referenced by no live work type: configuration that costs maintenance attention while serving nothing.
Work-type inheritance beyond the sane limit (the five-level limit with the multi-dimension test), a depth that makes every change risky to test and to reason about.
Phase guards that block legitimate transitions, or are absent where a control is needed, checked against the configured phase plan.
Matter titles are embedded with a local model on an on-premise stack, so client data never leaves a controlled environment, and then segmented; segmentation is exploratory and reviewed with your team. The lifecycle lens works from phase history: every count, median age, and reopen rate is an observed measure, and the domain health score is composed from disclosed indicators with published thresholds and weights.
One configured type absorbed 41% of the book (12,231 of 30,000 matters). Title embeddings, computed on-premise and labelled by a local model, segmented it into 7 distinct kinds of work (two of the eight AI-proposed segments read as the same kind of work and were merged before review; the audit surfaces a candidate list, and a human confirms it). The lifecycle lens then put numbers on the drag: 2 of 3 phases with open work carried a median age at or above 365 days (Re-Open at a median of 1,834 days, roughly 5 years), and 10.0% of closed matters were later reopened.
A taxonomy that matches the work you actually do (sharper reporting, routing, and resourcing), a dated stuck-queue list your team can clear matter by matter, and a reopen-rate driver review, all anchored to a health score whose methodology is on the table.
