Appendix EWhat do the quality numbers say?

The article counts stories, lines and findings. This appendix puts numbers on the technical quality of what was built: how long a build takes, what the product depends on and whether those parts have known flaws, what a static analyser makes of the code, and how fast it runs. Two of the numbers were not what the quality system assumed.

Status, 2 November 2026

This appendix is a snapshot of work in progress, measured on the code as it stood on 19 October. Open: the dependency advisories (OPEN-116), the missing software bill of materials and scan (OPEN-115), and the improvements ranked at the end. What comes of them will be in a later version of this appendix. Numbers in the form OPEN-nnn are entries in the project's register of open items, part of its quality system (chapter 06).

What was measured?

Measure Result
Full build and test of the three products, from clean about 6 minutes, on one core
Automated tests run by that build about 1,650, every one passing, every quality gate held
Runtime components shipped in the three packages 11
Components with a known advisory 2: the PostgreSQL driver in the configuration tool, three high; a template library, one moderate
Shipped components missing from the register of third-party software 6 of the 11
Static analysis warnings, all priorities 178, of which 3 high priority
Software bill of materials not produced since 30 September
Package sizes 1.4 to 4.0 MB
Deployment on the application server 2 to 3.5 seconds
Pages: landing, page composer, search, list, edit under a second for most

What does the product depend on?

Very little. The three packages ship eleven components between them: the PostgreSQL driver, a connection pool, the Thymeleaf template engine and the small libraries it brings, a logging interface and an annotation library. Everything else comes from the application server. The register of software of unknown provenance that a Class C process requires lists the three the product asks for by name, and none of the six those three bring with them, among them the library with an advisory below.

Two of them have known advisories. The configuration tool still ships version 42.7.4 of the PostgreSQL driver, while the device and access control ship 42.7.13. That older version has three advisories rated high: two concern how the driver authenticates to the database, and one lets a crafted authentication exhaust the processor. All three were published before 19 October, and are fixed in later versions of the same driver. The expression library Thymeleaf uses has one moderate advisory, published two days after the snapshot. The exposure is probably low, since the database runs on the same machine and only templates written for the product reach the library, but a Class C process does not accept "probably": the assessment and the upgrade are recorded as OPEN-116.

Why did nothing flag them?

The step that should have flagged them was not running. The quality system asks for a software bill of materials on every build, and for no unresolved high or critical vulnerability before a story is done. The bill is produced by a build plugin, set up and verified on 16 September. Two weeks later, the build script began running offline, and in quiet mode. The plugin needs the network: offline, it prints that it is skipping, and the quiet mode hides the line. Since then, no bill has been produced, and nothing has scanned the dependencies. The item that set the plugin up is still marked resolved.

It is the same kind of slip as chapter 07's wiring: a check that looked in place and did nothing, with nothing to notice its absence. It is recorded as OPEN-115.

What does a static analyser make of the code?

SpotBugs, run at its most thorough on the three products, reports 178 warnings: 39 on the device, 124 on the configuration tool and 15 on access control. Most are routine for this kind of code: objects that hand out their internal lists instead of copies, and fields that would not survive being copied between servers. The three high-priority warnings are of that second kind, in what the configuration tool keeps in a user's session: harmless on one server, a blocker if the hospital ever runs two. The configuration tool also catches a null reference after the fact, 21 times in its page composer, where it could have checked first: a habit of code grown story by story.

One warning is security-related: a database statement assembled from a string that is not a constant. It is one of the two places that keep to the constant-SQL rule only in spirit, found by hand for chapter 07 and recorded as OPEN-107. The analyser found it on its own.

How fast does it run?

A full build and test of the three products takes about six minutes, which keeps a check before every change affordable. The packages weigh 1.4 to 4 MB each, and the application server deploys them in two to three and a half seconds.

The pages were timed in a browser's network panel, on a range of everyday tasks: the landing page of both modules; in the configuration tool, opening the patient page with its live preview; in the application, opening the search, searching and seeing the list update, and editing a record. Most come back in under a second. Everything ran on one workstation, bought in 2014, with a 2.1 GHz processor: the database, the application server, the identity provider and the browser. A hospital's server would do better.

What did I learn?

Two things about Claude and BMAD come out of these numbers. The code is lean: few dependencies, small packages, fast to build and to deploy, and the static analyser's warnings are mostly of the routine kind. And the quality gates the build runs, on coverage and traceability, all held.

About the method, each finding points at how the work was done:

Finding What it says about the method
The software bill of materials stopped being produced, unnoticed The check could skip itself in silence, and nothing verified that its output existed
A driver with three high advisories shipped for weeks Dependencies were registered when added, and never scanned afterwards
Six of the eleven shipped components are missing from the register The register records what the product asks for, and misses what comes with it
One module lags the others on the same component Nothing compares versions across modules
The analyser confirmed a finding made by hand A cheap tool that was not run would have found it first

What would I change?

These proposals are ranked by the benefit I expect; each comes with the figure that would show, at a later check, whether it worked.

Rank Improvement Why at this rank Measure
1 Every build produces the bill of materials and scans it, and fails when the bill is missing or a high advisory has no recorded assessment It closes the gap that let the advisories through, for every later dependency Advisories open longer than a week
2 Every check in the build proves it ran: its output is checked to exist It removes the class of slip seen here and in chapter 07 Checks that verify their own output
3 The static analyser runs in the build, and its count may not rise Cheap, and it found by itself what took a person to find Warnings over time
4 The register of third-party software is generated from the bill of materials, so it lists everything shipped, and shared components carry the same version in every module It removes the gap in the register and the lag that kept the old driver Shipped components missing from the register; components with more than one version

Where does this leave the project?

The project stands at a gate here too. The code is lean and fast, and the checks the build runs hold. Two checks the quality system counted on were not running, and that is the main lesson of these numbers. The two open items and these four proposals are among the decisions waiting at the gate. The next version of this appendix will report on them.