Repository navigation
Local AI Visual Decision 1.1.0 (app 1.0.52 preview)
Pre-releaseOffline visual yes/no decisions as a separately imported .mhblock plugin for a
compatible Macro Handler 1.0.52 preview build. Earlier 1.0.51 builds are not
advertised as compatible. The host app provides the reviewed CPU runtime; model
weights are separately installed data, not included in this release or the APK.
Read the English setup/capability guide
before downloading local_visual_decision.mhblock. The guide covers selecting
SmolVLM or the optional LFM fast profile, explicit model-license acceptance,
regions/prompts/variables, True/False/Unknown/Error actions and verified imports.
Measured on Samsung SM-A055F and LDPlayer: resolution variants 3/3 and original
monster/human/rock/empty/small-sprite samples 5/5 correct on each device. These are
small owned synthetic samples, not a real-game recognition guarantee. A single
pixel change was missed; arithmetic checks were only 2/4, and a numeric-output
request returned Yes, not 42. This is not a general maths solver or exact
pixel detector. Every expected/actual observation is included in the evidence.
The separate deterministic OCR + Math + Lua variable + focused input/paste + OCR
answer lookup + real tap pipeline passed on a controlled physical Samsung screen.
A real empty-hint-field paste rejection was narrowly fixed in the prepared app;
sensitive-input and selection protections remain. No model/OCR engine was changed
to disguise a failed capability.
UI follows 14 app languages and themes; equivalent model accuracy in 14 languages
is unproven. Each invocation analyzes one fresh region frame, in memory, without
screenshot-history files. Earlier fast-profile warm samples were 0.91–1.76 s on
Samsung and 0.39–1.29 s on LDPlayer; Samsung cold initialization was 3.66 s.
No universal three-second, 24-hour or arbitrary-screen guarantee is claimed.
Assets: descriptor, detailed guide, manifest, capability evidence, MIT descriptor/
documentation license and local-vision-SHA256SUMS. No APK, executable libraries
or model weights. Existing four official plugin v1.1.0 downloads are unchanged.