One chain runs through every engagement. A break anywhere surfaces as a symptom somewhere else — usually somewhere convincing. Each of these was traced from the symptom to the layer that actually broke.
Where it looked broken: Web. Where it actually broke: Data.
The dashboard is wrong — the pipeline must be broken.
Fresh API data was being compared against finalized stored data.
Nothing was broken. Two sources were answering slightly different questions, and the reconciliation rule had never been written down. The fix was a data contract, not an engineer.
Where it looked broken: Web. Where it actually broke: Reliability.
Every site in the portfolio went down at once.
The verification system had lost connectivity — the sites were fine.
A monitor that cannot detect its own failure will confidently report a catastrophe. Abstention states and a fleet-level canary came out of this one.
Where it looked broken: Automation. Where it actually broke: Operations.
We need to automate the workflow. All of it.
Most of the task portfolio was not worth automating at all.
3,089 recurring tasks measured by frequency, duration, variance and risk. Most could run unattended once the rules were written down; some needed a human gate; a small share was simply deleted.