BrotherSBE: a Claude Code plugin built around receipts instead of claims

[khalilmaaouni]

Posting this in Show and tell because I think the mechanism is more
interesting to this audience than the pitch.

BrotherSBE is a backend and data engineering colleague skill. Before any
code, it produces a dossier (purpose, process, ADR, data model, diagrams,
verification), sized by a five-question intake that decides which of those
files a given change actually needs (T0 through T3, first match wins).

The part I'd like feedback on from people who write skills: the four hard
gates (numbers, migration, approval, ran) never infer a PASS. Each reads one
JSON receipt file and checks it for internal consistency, not just presence.
A numbers claim needs a second derivation, textually different from the
first, rerun to zero drift against a pinned snapshot. A migration needs
forward and reverse both run against a restored copy with matching row
counts. If the receipt is missing, the verdict is NO-DATA. If it exists but
records nothing, still NO-DATA. If it exists but can't be parsed, that's a
FAIL, since a broken claim is treated as worse than an absent one.

There's a meta-test (evals/test_no_data_class.py) that discovers every check
registry in tools/ on its own, rather than reading a written list, then
hollows out each check's own worked example one field at a time and asserts
none of the hollowed versions produce a PASS. Current count: 32 checks, 6
registries, 3780 scenarios, 0 failures.

Free, MIT, solo, no users yet: https://github.com/khalilmaaouni/BrotherSBE. Curious
whether other skill authors have hit the same "my check silently passed on
garbage" failure mode, and how you closed it.