What went wrong, and what it was worth

Our paper's first required comparison is structured regional state versus a pooled vector. We built the second one, named it after the first, and measured it losing. The controlling defect was not the narrowing — it was that nobody had written the narrowing down.

A parameter with a closed-form optimum was left to gradient descent for two minutes, got 20% of the way there, and lost the entire run. The model had the best conditional mean of all seven arms and the second-worst score.

A catalogue of instruments whose reading was constant with respect to the question being asked. All were green. All looked like evidence. Plus the inverse category — guards asserting a defect still exists, which fail the moment someone fixes it.

We built the fine/coarse resolution pair the paper asks for, measured whether it validates at its boundary, and it does not. The failing measurement is the result — and it says something specific about what regional state can carry.

Every divergence between what we wrote and what we built, in a register any contributor may add to and none may delete from. A stated narrowing is a decision you can attack. An unstated one is invisible to a process built entirely out of attacking stated things.

A trailing slash in .gitignore that could not match a symlink and gave seven checkouts ELOOP. An anatomy loader whose exception was swallowed by a bare except. A provenance field that was structurally incapable of ever reading clean.

A normalizing flow reported −log q of 4.489e9, bit-identical across runs, from the first step. Three separate causes, each invisible to the others' tests. What found them was a counter, not a bound — and the fourth fix was committed as "the root cause" and then withdrawn.

Six of seven defects blocking a training run were the same defect wearing six different files. Two of them would not have crashed — they would have produced a right-looking number from a model that quietly lost the mechanism the experiment exists to test.

We built a refusal to stop a model shipping under a name it had not earned. Then an evaluation stamped the previous model's name on its own results, in the file the public card reads, and nothing objected — the guard watches emission, not writing.

A health check read a log path that did not exist. It printed nothing and exited zero — output shape-identical to a healthy run — and it was the only thing watching a multi-hour unattended job.

Relating two supports meant hand-writing a map. An algebra now derives what the declarations support and refuses the rest — because a fabricated correspondence is indistinguishable at the type level from a real one. Nine sources, one view, and the lead field is a buffer.

We renamed the training stages. Six gates matched on the old names and five gave the wrong answer, so the model trained on simulation alone. A complete fix had been sitting unapplied since that morning, over six red tests that named the defect.

A 513-line registration module, the volumetric atlas it needs, the template it registers to, and the solver — all present, all unused, and every source card describing the capability as missing. The distance to a second modality was a caller, not a component.