Clarify distinct Cua and Omabot qualification runs

This commit is contained in:
Francesco Bonacci committed 2026-09-12 16:42:57 -05:00
1 parent 76bd68b2aa
commit 85e91947f6
1 file changed
+10 -5
+10 -5
View File
@@ -50,11 +50,16 @@ Cua's retained evidence covers background application effects, two-lane overlap,
third-owner refusal, primary-input preservation, conflicts, stale targets and
geometry, cancellation, desktop faults, recovery, and cold package transitions.
Production-package app checks and independent primary observers are separate
from trace-enabled diagnostics. The complete native Linux runner passed all
128 required cells: 87 deliveries and 41 expected refusals, with no failures
or skips. That runner used source-built released Driver 0.24.0; the real-app
checks separately used the actual Omarchy `cua-driver-bin 0.24.0-1` executable.
See the linked qualification record for exact artifacts and observation limits.
from trace-enabled diagnostics. Cua's retained canonical run
`433ce968ee164d5e8e3226e800db93a6` recorded all 128 required cells: 87
deliveries and 41 expected refusals, with no failures or skips; its completion
report has SHA-256
`1eda4cc008ea6b27d96c21d58f8041834398a43409642376bafe84ad38f0e112`.
That run used source-built released Driver 0.24.0; the real-app checks
separately used the actual Omarchy `cua-driver-bin 0.24.0-1` executable. A
separate later Omabot replay reported 124 passes and four failures, plus seven
incomplete native cases, so Omarchy replay remains an explicit merge gate. See
the linked qualification record for exact artifacts and observation limits.
Duplicate motion notifications are retained and counted. They are acceptable
only when pointer identity, coordinates, focus, held input, and foreground