Appendix: defects found while building

This is an appendix, not the argument. It exists because the checks described here are running in the repository and someone maintaining the code needs to know why they are there. The full narrative for each entry is in reports/decorative_guards.md.

Defect	What it cost	What catches it now
Card patterns named modules that no longer existed Regional modules were renamed `local` → `family_local`; the source cards still granted the old names. An unmatched glob is an empty permission set, not an error.	88.8% of run 2's parameters — the entire family-indexed regional model — received no gradient for 8,700 steps. The loss fell anyway, because the 285k reachable parameters were enough to fit something.	`tests/foundation/test_card_patterns_reach_the_model.py`: no module unreachable by every card, and no grant pattern that names nothing in any architecture.
Stage behaviour selected by matching stage names Six gates in the trainer keyed on run-1 stage names. Run 2 renamed every stage, so five returned the wrong answer.	Nine hours of training on simulation alone, with every metric healthy. No gradient was ever taken on a recording.	Stage admission is declared in config (`extra.curriculum`) and a stage that declares nothing is refused rather than defaulted.
A permission flag that nothing read `allow_derived` appeared once, in a signature. The config option that fed it defaulted to `False` and was discarded.	An evidence-rejected cortical partition (Yeo-7, 6 of 21 pairs separating under a spin null) ran silently whenever an anatomy declared no families.	The refusal exists and sits where the fallback is taken; opting in records `REJECTS:` in the partition's own notes.
Checks that could not fail Guards whose assertion was true by construction — a lookup with a permissive default, a count that matched itself, a regex that matched nothing.	Several defects passed review while the check written to catch them reported green.	Mutation testing is the standing requirement: a guard is not accepted until it has been made to fail on purpose.
Two directories for one object `configs/source_cards/` read by release, `configs/curriculum/source_cards/` read by training.	The published licence manifest was computed from a different card set than the one that trained the weights.	The release path derives the directory from the checkpoint's recorded config, and a disagreement between checkpoint and config is a blocker.
Artifact claims recomputed from current configuration The model card's gradient-reach figure was read from today's source cards rather than the run's.	A routine republish would have replaced 88.8% with 0.9% — a number describing a run that never happened.	`tests/release/test_card_is_computed_from_the_run.py`: anything said about weights is computed at the checkpoint's recorded commit, and a dirty tree is disclosed.
Instruments that failed silently A health check reading a log path that did not exist; a sweep whose recorded exit code came from `basename` rather than the command.	A finished run reported as dead; a killed run reported as `exit=0`.	`scripts/health.sh` fails on its own preconditions and reads the checkpoint, not only the log. `tests/release/test_shell_exit_capture.py` rejects `$?` read after a command substitution.
Six same-name class pairs across modules `derive_families`, `ClaimManifest`, `RegionFamily`, and the R12 rule itself, each existing twice in unrelated namespaces.	Repeated misdiagnosis; in one case a refusal's remedy named a field the object it validated could not carry.	Partition agreement is enforced (`tests/foundation/test_one_region_ontology.py`); full unification is `ARCHITECTURE.md` O-3/O-7.
Global state set at import time Eleven test modules called `torch.set_default_dtype(float64)` at module level, which runs during collection.	The whole suite ran in float64 whenever those files were collected, inflating a failure list from 10 files to 38.	An autouse fixture owns the dtype per test and restores it.

The one generalisation worth keeping

A published artifact and the code that generates it are two objects with separate histories. Green tests on the generator say nothing about the bytes already published, and a healthy artifact says nothing about the generator. Only a diff between them is evidence about both.

Three of the entries above were found that way and by no other means.

Appendix: defects and the checks that now catch them

The one generalisation worth keeping