Development covered 5 Aug 2026 to 7 Aug 2026 (commit dates).
This stretch was mostly me auditing my own work, six rounds of it, on the tooling that edits the tie-in novel and the tooling that plays the game. It produced the single most useful measurement I have taken on this project, and I want to write it down while the sting is fresh.
Round three stopped asking whether it was correct
The first two rounds were the ordinary kind. Does the documentation match the scripts. Does a cold start work. Do the paths resolve. Both rounds found real problems and both left me feeling productive.
The third round asked a different question: has this pipeline ever made a single chapter better?
I went and counted. Chapters improved: zero. Chapters damaged: four.
Everything was working. Each stage ran, each check passed, the output was well formed, and the net effect on the actual artifact was negative. Two rounds of review had been auditing the machinery against its own description, which is a closed loop that can be perfectly consistent and still describe something useless.
The measurement changed what I built next far more than the previous forty findings combined. It is a lot harder to arrange than a consistency check, because you have to define the outcome and then be willing to read the answer. That, I think, is exactly why it does not get done.
The three shapes that kept recurring
Across the rounds, most findings were one of three things.
A guard that is always false is not a guard. Several safety conditions could not fire, because the value they tested was never in the state that would trip them. Nothing errors. The guard sits there being reassuring, and reads to a later maintainer as protection already handled.
A finding that is found and then un-found. A previous round had identified a real defect, and the fix for it had been reverted or bypassed in later work, and no record survived that this had ever been discovered. Without a written trail, the same defect gets found again as a new one every few weeks, and each of those finds costs full price.
Change A, forget B. The overwhelming majority of the genuine findings, and the whole content of one outside review. Somebody adjusted a rule and left the place that depended on it. Every one is trivial in isolation and there is no upper bound on how many of them a system can contain.
The refuter
The habit that came out of this and that I have kept: before shipping a fix, have something argue against it.
Concretely, four findings were written up as ready to ship, and passing them to a critic whose only job was to attack them killed three. Not because they were poorly implemented, but because the thing they described was not actually wrong, or was already handled somewhere else, or would have broken a rule the fix's author had not read.
A seventy-five percent kill rate on my own confident conclusions is a humbling number and a very cheap one to buy. It also inverts the usual worry about review: I had assumed the risk was missing problems, and the measured risk was inventing them.
Two more habits fell out of the same rounds. Say each rule once, in the place that owns it, because a rule stated four times is four things to keep in step and I had several of those. And when the review is of documentation, simulate a cold reader following it step by step rather than reading it for correctness, because those two exercises fail in completely different places.
Nothing a run starts may outlive it
One operational note, since it cost several nights.
Long unattended runs spawn things: browsers, servers, helpers. A run that finishes cleanly, writes its report and exits can still leave a browser driving away in the background, and the next run inherits a machine with several of those on it. Two of my nights were lost to memory pressure caused entirely by the previous night's leftovers.
The rule is now absolute. Nothing a run starts may outlive it, and the run's own cleanup has to be tested, because cleanup code is the least exercised code in any tool and it only matters when everything else already went right.
The live build is Cultivation Game.
