Автоматическое доказательство
M7.7 furniture and fit PR #35
Состояние: merged
Что подтверждает: M7.7 was exact-head verified, product-owner accepted and squash-merged. It proves the searchable furniture catalogue, M2-derived fit explanations, selected-object inspector, atomic Apply and one-step semantic Undo described by that milestone. It does not prove final plan-recognition accuracy.
Открыть доказательство ↗
Автоматическое доказательство
M7.8A recognition benchmark PR #40
Состояние: merged
Что подтверждает: M7.8A established an eight-fixture public-safe benchmark, deterministic geometry/topology/opening/room/area/confidence metrics and real Chromium/OpenCV source execution. Product-owner calibration acceptance covered image-rectangle coordinates and undirected horizontal/vertical calibration. The accepted baseline intentionally exposed zero Source wall and opening F1 before quality tuning.
Открыть доказательство ↗
Автоматическое доказательство
M7.8B wall topology PR #41
Состояние: failed
Что подтверждает: Product-owner review failed on a representative clear apartment plan: the local Draft contained 417 wall candidates, zero openings and a symbol/furniture network instead of an architectural wall graph; AI reconciliation retained the polluted network and added unsupported long lines. Corrective structural masking reduced an isolated reproduction from about 948 to 115 unique segments and from 12,346 to 249 admissible parallel pairs, but current aggregate Source geometry/topology F1 is only approximately 0.492537/0.462687. PR #41 remains Draft and must not be described as accepted recognition quality.
Открыть доказательство ↗