Factory changelog

v3 added the verification factory: a second, read-only machine that reads an existing system and produces a behavior catalog with typed semantics, a tiered coverage-debt report, and the regression tests that close the gaps. New machinery in this generation: typed oracle tiers (only cross-confirmed and backend-grounded facts may be asserted as invariants), provenance gates that grep cited excerpts rather than trusting line numbers, enforcement matrices crossing frontend rules against backend guards, and axis-independence analysis to keep generated permutations honest.

Its first estate engagement covered 15 production targets, with generated suites delivered into the client's own QA repo and idiom. Against live staging, 23 of 24 generated specs ran green on the first run. The failure was a timing flake, and it produced two permanent fixes: a flake lint and a learned status correction. The factory keeps what it learns.

v3 also carried our own platform across a generation. Parity replay, replaying recorded operations against the new generation and diffing every result against the old, is how the platform moved from v2 to v3 with zero behavioral difference asserted, not assumed.

v2 ran a legacy system through the line end to end with no architect at the stations. The subject was Bugzilla: 25 years old, recognizable, and public, so anyone can check the port against the original.

Discovery mapped 7 bounded contexts and 18 cross-service events. The build produced five services (user, product, bug, attachment, notification), a frontend, and end-to-end journey tests, passing through roughly 30 stations and 40 gates on the way.

v1 built OTR Select, a production SaaS platform, in six weeks. Built on Factory v1: every line AI-written, with an architect reviewing at stations that are now automated gates.

Those reviews are the honest measure of this generation; the stations that needed them are the ones v2 and v3 went on to automate.