Skip to content

test(crv): add deterministic emulator host variant - #482

Draft
daniele21 wants to merge 10 commits into
devfrom
agent/crv-two-apk-emulator-host
Draft

test(crv): add deterministic emulator host variant#482
daniele21 wants to merge 10 commits into
devfrom
agent/crv-two-apk-emulator-host

Conversation

@daniele21

@daniele21 daniele21 commented Aug 28, 2026

Copy link
Copy Markdown
Owner

Scope

Adds a dedicated x86_64 emulatorE2e phone-host variant for remote two-APK Android E2E.

  • ordinary debug/release variants remain ARM64 and continue using the native llama.cpp backend;
  • the emulator variant keeps the real Harness app, Service, Binder host, control plane, activation/residency and RuntimeOrchestrator;
  • only model bytes and GGUF execution are replaced by an emulator-only pre-provisioned ModelStore and deterministic InferenceBackend;
  • deterministic output follows the real RedactGuard structured-analysis schema and is derived from the submitted selected type/segment identity;
  • no consumer-facing model identity or tuning is introduced;
  • Harness owns a reusable exact-ref workflow that builds and publishes the emulator host APK plus bounded source/SHA evidence for consumer E2E lanes.

Why

Prior emulator evidence exercised the Binder contract but not the real two-APK product path. The production phone APK is ARM64-only, so a GitHub x86_64 AVD cannot install it. This variant closes that REMOTE_AUTOMATED gap without pretending to validate native ARM64/GGUF execution, which remains CRV-110 REAL_ENVIRONMENT.

Cross-app evidence

RedactGuard #109 consumes the Harness-owned emulator artifact without checking out or building Harness source in the consumer repo. Its API 35 x86_64 AVD lane has passed Host-absent fail-closed behavior and the cross-process Binder/control-plane/readiness/runtime journey through Review on the exact Harness head used here.

Validation

Build logic/runtime composition/workflow changes require FULL exact-head remote preflight. Automatic FULL validation on this head is green; the explicit /preflight full replay is the final deterministic readiness confirmation. Physical ARM64 + real GGUF remains CRV-110 REAL_ENVIRONMENT and is not claimed by this PR.

Copy link
Copy Markdown
Owner Author

/preflight full

Copy link
Copy Markdown
Owner Author

/preflight full

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant