Keeping score honestly: zero improved, four made worse

The development period ran from August 5, 2026, to August 7, 2026, during which time the author focused on auditing their own work on the tooling that edits a tie-in novel and plays the game. This period produced the most useful measurement on the project to date. The first two rounds of auditing were standard, checking if the documentation matched the scripts and if a cold start worked, with both rounds uncovering real issues.

In the third round, the author asked a different question: had this pipeline ever improved a single chapter? They counted and found that there were zero chapters improved, but four chapters were damaged. The findings confirmed that everything was working, as each stage ran, checks passed, the output was well-formed, and the net effect on the artifact was negative.

The author emphasized the importance of defining outcomes and being willing to accept the answers, as it is much harder to do than consistency checks. The three recurring shapes in the findings were: guards that are always false, safety conditions that could not fire, and nothing errors; a finding that was found and then un-found; and change A, forget B, where a rule is adjusted but the dependency is forgotten.

The author adopted several habits from the auditing process, such as having someone argue against a fix before shipping it, stating each rule once in its owning place, and ensuring that the run's cleanup is tested, as cleanup code is the least exercised code. Finally, it was crucial to remember that nothing a run starts may outlive it, as long unattended runs can spawn processes, causing issues for subsequent runs.