Week 5 Session: What the System Used to Be

Can a fault be explained by something that is no longer in the code at all? This session treats version history and prior configuration as first-class evidence, not background.

22 March 2027

Before the session#

Bring your Week 3 structural model of fulfil and your Week 4 instinct for runtime evidence — this week’s fault is invisible to both until you add a third source, but neither one stops being useful.

The incident#

A small fraction of fulfil’s orders are being handed to Continental Parcel — a carrier Northfield Mercantile stopped using at the start of the year. Nothing in the present-day routing code names Continental Parcel; the lookup table your Week 3 model already covers lists only current carriers. Runtime traces (Week 4’s evidence class) confirm the symptom is real: for the affected orders, fulfil genuinely issues an outbound call to Continental Parcel’s API. Reading the current source harder will not explain a call to a carrier the source doesn’t mention.

Introduced near the failure, or caused it?#

Version history shows a carrier-table consolidation migration that ran fourteen weeks ago, and — in the same week — an unrelated logging refactor that happened to touch the same file. Bisecting by proximity in time alone would leave you unable to tell which of the two is relevant, or whether either is. “This change is the closest one in time” is not the same claim as “this change is why the failure happens,” and the practical is built so that conflating the two produces a wrong answer that still feels confident.

In the session#

You bisect through version history toward a candidate change, then use structural and/or runtime evidence — not the commit timestamp — to test whether that change is actually responsible. A candidate that survives this check is more than temporally adjacent to the failure; a candidate that doesn’t survive it is a coincidence you were right to suspect and right to rule out.

Afterwards#

This is the last of three evidence classes introduced deliberately, one per week — structural, runtime, historical. From Week 6 onward you’re expected to combine them without being told which to reach for.