What your spend produced
The triangulation layer — comparing what was promised, paid, planned and reported against what actually exists.
The July 2025 readiness target was missed; the Sep–Oct 2025 window was missed (scope hardened); December was partial and insufficient; a filing-window-driven MVP finally launched Jan–Feb 2026 and now runs under live use.
Raw output is high (380K lines, 17 feature areas, 99 modules), but 27–29% is documented AI-generated (likely 40–50%), and 20+ duplicated DataGrids (~10–15K lines) inflate the totals — unique human-authored logic is well below the headline.
654 errors and 167 warnings unresolved, 152 TODO/FIXME across 109 files, and 20+ duplicated grids — debt accrues faster than it's paid down, and without tests each change risks regressing shipped features.
The root CLAUDE.md instructs commit messages to omit AI co-authorship, and 'known fragile areas' (permissions/role leakage, advanced search, notification reliability) ship without automated verification — what's reported diverges from the record.
Each comparison is stated as dates and counts only, confirmed against two independent sources before it ships. This is the section a vendor contests hardest — so it carries only facts, never a characterization of intent.