A manufacturer retires a quality application

A fifteen-year-old Java application that did tracking and trending for a regulated quality function. Reverse-engineered from source, data and outputs, then rebuilt as reports.

Built on synthetic data for a fictional company — Northwind Consumer Products. The method, the analysis and the outputs are real; the client is not.

Behavioural claims extracted: 170
Verified against real outputs: 129
Input feeds mapped: 16
Of loads were manual: 86%

What the estate looked like

Northwind built a quality tracking application in 2010 in response to an audit finding. It has done its job every month since. It has also been flagged as a security risk, the team that built it has gone, and eighty-six percent of its data loads are performed by hand by one person.

The application was not really an application. It was a data warehouse and a set of reports wearing an application costume, which is why it could be retired rather than rewritten.

What the analysis found

We extracted 170 specific behavioural claims from the source code, the functional documentation and the data, then tested each one against the system’s own outputs. 129 verified against the sample we were given; the remainder needed the full dataset rather than the one percent slice we had.

Ten places where the functional documentation and the running system disagreed. In a regulated context that is not a curiosity. It is the gap between what you tell an auditor and what the system does.

What the rebuild looked like

Sixteen feeds landed in a Fabric bronze layer, conformed upward, with the manual loads automated. The reports came back at parity in Power BI so the quality team did not have to relearn anything, and the audit-freeze problem, which is snapshotting a dataset at a point in time while the system keeps running, became a native capability rather than a manual database copy.

A fragment of the claim ledger

Every behaviour the harness extracted, tested against the system's own outputs. This is a small part of one demonstration's ledger — enough to show the shape of the work, not enough to hand over the method.

C-001	duration accrues from first-report, not first-confirmed	proven
C-002	weekend hours excluded from the rolling mean	proven
C-003	asset class inherits criticality from parent feeder	strong
C-004	nulls in the source are treated as zero, not skipped	proven
C-005	retired assets remain in the twelve-month trailing window	untested
C-006	threshold is evaluated per part, not per lot	proven
C-007	the monthly job re-reads the prior period and overwrites	refuted
C-008	manual corrections are applied after aggregation	strong
C-009	duration accrues from first-report, not first-confirmed	proven
C-010	weekend hours excluded from the rolling mean	proven
C-011	asset class inherits criticality from parent feeder	strong
C-012	nulls in the source are treated as zero, not skipped	proven
C-013	retired assets remain in the twelve-month trailing window	untested
C-014	threshold is evaluated per part, not per lot	proven
C-015	the monthly job re-reads the prior period and overwrites	refuted
C-016	manual corrections are applied after aggregation	strong
C-017	duration accrues from first-report, not first-confirmed	proven
C-018	weekend hours excluded from the rolling mean	proven
C-019	asset class inherits criticality from parent feeder	strong
C-020	nulls in the source are treated as zero, not skipped	proven
C-021	retired assets remain in the twelve-month trailing window	untested
C-022	threshold is evaluated per part, not per lot	proven
C-023	the monthly job re-reads the prior period and overwrites	refuted
C-024	manual corrections are applied after aggregation	strong
C-025	duration accrues from first-report, not first-confirmed	proven
C-026	weekend hours excluded from the rolling mean	proven
C-027	asset class inherits criticality from parent feeder	strong
C-028	nulls in the source are treated as zero, not skipped	proven
C-029	retired assets remain in the twelve-month trailing window	untested
C-030	threshold is evaluated per part, not per lot	proven
C-031	the monthly job re-reads the prior period and overwrites	refuted
C-032	manual corrections are applied after aggregation	strong
C-033	duration accrues from first-report, not first-confirmed	proven
C-034	weekend hours excluded from the rolling mean	proven

Synthetic. Truncated deliberately.

See what this method reveals in your environment.

Share one representative source, report set or business question. We will work out what can be demonstrated and what a production path would need, before you commit to anything.

Talk to an architect