The distribution of bypasses shows what the system is missing

What it is

A bypass is a call site that finishes the same kind of UI outside system tags: a private shell, a native clone, a control copied from elsewhere. One bypass can be a rush job. The distribution of bypasses, laid out by page, task, and control kind, stops being “someone would not follow the rules” where it is dense, and becomes a missing piece of the system — no date picker of that kind, no multi-select of that kind, no action button that can enter toolbar density. Distribution is diagnosis, not a blacklist.

Read shape: scatter (occasional, everywhere) versus cluster (the same task, again and again). Clusters point at a gap. Scatter is more like a local expedient, and should not drive the system to open a component for every expedient.

Why it happens

Consumers rarely bypass because they prefer to build it themselves. They bypass because the task at the call site has no legal path through the system: no such state, no such slot, in-context copy never said how to combine, contributing a new component costs more than building privately. The same gap repeats on every page that hits the task, so bypasses cluster on the task dimension. Slice by team, and the cluster shatters into “these groups are not adopting”; slice by task, and it regroups as “everyone is privately building date ranges”.

Diagnosis reads the cluster’s label, not its headcount. Ten products each shipping a chip row on a filter bar means the system is missing filter chips, not that ten reprimands are due. Treat the distribution as a map of disobedience, and maintainers will go persuade people to change calls; the gap stays, and next quarter the cluster is still there.

Where it stops holding

Safety bypasses (payment and auth that must be platform-native, not swapped for a system control) form legal clusters and should be marked not-applicable, not a gap. Stale bypasses — the system has since filled in, the product has not migrated — are a migration queue, not a new gap; split them in the distribution by version. Non-system drawing in exploration files is not a product bypass. A single site whose task is unique enough that nothing else will hit it stays a local implementation; do not let “it counts as a point” in the distribution push an extraction. Scanner blind spots (unparseable native modules) invent false scatter or false holes; fix recognition before reading shape.

Applying it

• Tag every bypass with two labels: the task (filter, date, upload…) and the surface (settings, list, marketing). Cluster by task, not by team.
• When a cluster crosses a threshold (e.g. same task ≥ 5 sites, or ≥ 3 products), open a system backlog item: which capability is missing, not “notify team X”.
• For clusters that already have a matching component, inspect in-context copy and legal configuration — often the miss is usage that cannot be read, or a missing state; write the backlog as copy or state, not a new component.
• How to check: list the three largest clusters and write, for each, “if the system had X, these sites could disappear”. X must be a capability (control, state, slot), not “they were more disciplined”. After shipping X, rescan: the cluster should fall. If the only action was a remediation letter and the cluster does not move, diagnosis was used as a blacklist. Re-slice the same points by team: if the cluster vanishes, a team-only view would have lost the gap — do not ship team views alone after that.

Related

• Same group: R1.18.1 Adoption is counted at call sites, not by teams · R1.18.2 High coverage is not the same as correct use · R1.18.3 Metrics exist to locate gaps, not to score consumers
• Adjacent: R1.06 Contribution and governance · R1.02 Component library and variants
• Search terms: bypass distribution as diagnosis · bypass map · missing-capability signal