NorthCUA Harness verification

Historical record. See the current verification for the later onboarding, skill, local batch and distribution audit.

Verification date: 2026-10-03. NexCUA repository, NorthCUA Harness candidate: 0.2.0, branch codex/northcua-clean-open-source. The latest audit follows candidate 6c7fa33f9f1deedf2326aa830db3ce68ecde3f61 and simplifies Desktop to one tool-returned-screenshot path. The clean implementation originated from main 315508f3c08a7dec7f753691f38a609c9b8d0601.

The local completion gates below passed after this simplification. Earlier test results and reviewer signoffs do not substitute for these runs. The matrix distinguishes real task execution, deterministic regression coverage and outstanding deployment acceptance.

Execution matrix

Path or gate	Latest result	What this establishes
Full default suite	2815 passed, 7 skipped in 180.83s	Runtime, provider streams, context/memory, evaluator, transport, batch/recovery and release regressions. The seven opt-in tests executed separately below. Two warnings come from intentional fork-safety tests.
Web: local application, CLI, local WebSocket server, controller in-process and controller process	5 passed in 25.94s	Actual Chrome and pinned public NexAU through a deterministic local HTTP model; shared actor/evaluator live page state, image budgets and working-memory projection.
The same five Web paths from final rebuilt wheels	5 passed in 25.51s	Execution outside the checkout in a separate environment. One warning concerns the copied test's unregistered integration marker; all five tests executed.
Native macOS Desktop	2 passed in 7.35s	Actual input and screenshots on a disposable Tk window, plus DesktopApplication → NexAU → note/type/keypress → observed text. Every actor model request must contain exactly one image at budget 1, including the request following a text-only note. The session ID contains Chinese and a slash. Evaluation is disabled in this native smoke test.
Four distributions	Four 0.2.0 wheels built	Packaged configs, LICENSE and NOTICE are present. Every packaged Python source matches the checkout; deleted observation modules are absent.
Wheel-only deployment	Passed before and after installing pinned public NexAU; final rebuild smoke also passed	Config generation, worker CLI/doctor, controller startup and runtime bundling outside the checkout. Dependency resolution is separate from the source lockfile.
Local Desktop WebSocket server	Controlled regression tests passed	Admission, physical-display exclusivity, ACK failure, cancellation, terminal errors and shutdown. No complete physical task through this server was accepted in this audit.
Remote Desktop / controller device input	Contract, transport and failure tests passed	The shared observation/command chain is covered; a complete physical task through each remote Desktop mode remains unverified.
Batch and OSWorld adapter	Deterministic regressions and fixtures passed	Claims, preparation, deadlines, retries, recovery, artifact custody, cleanup, task/oracle mapping and terminal behavior. No real cloud provisioning or upstream OSWorld VM acceptance.
Static and release checks	Passed	Ruff lint/format, configured Pyright scope, lockfile, workflow actionlint, whitespace, public package/config contracts, documentation links and paper/artwork hashes.

The source run used Python 3.12.13 on macOS arm64, the committed lockfile, NexAU c98f7fb73f41f338e462d0585c62c1e62362b693 and Chrome 145 / Playwright browser revision 1208. Installed-wheel dependencies were resolved separately. No hosted-model credentials or cloud accounts were used. The deterministic HTTP model checks real execution wiring and evidence delivery, not autonomous model capability. See testing for commands and opt-in requirements.

Single Desktop observation path

GUI actions return post-action screenshots. The runtime uses their clean images as the next state and projects history through the report's overlay/clean window. Initial capture supplies an image when no clean tool result exists; notes preserve the current grounding image. Optional accessibility recording and evaluator evidence remain available.

The removed code includes the pre-round probe policy, separate user-observation projection and factory, configurable observation chains, an empty window/control chain, the duplicate legacy collector, unused probe adapters, generic image allocation and export-only prompt interfaces. Orphaned driver comparison helpers and unused Desktop wait/clipboard settings were also removed. The configuration migration identifies obsolete keys, which now fail validation instead of being silently ignored.

Tests for deleted modes were removed; tests of surviving behavior were migrated and extended. Real-provider image assertions cover the note-only boundary. Native fixture focus is explicitly confirmed after application initialization, and the test waits for its own window's output before asserting the result. This addresses the focus/timing failures encountered during this audit without changing production clipboard handling.

Independent reviewer signoff

Both independent, read-only review agents examined this simplification:

• Runtime: scoped SIGNOFF. No unresolved P1/P2 in the reviewed image/state chain, removed APIs, local/remote observation adapters or source persistence. Independently passed 303 focused tests, followed by 19 configuration-snapshot/tool-surface checks with overlapping coverage. Verified the real Web and native execution logs.
• Release: scoped SIGNOFF. Independently passed 26 release/configuration/observation boundary checks. Source and packaged Desktop/Web configurations and the OSWorld example load successfully. The runtime import scan found no references to deleted modules; documentation distinguishes removed modes from retained capabilities.

The final full-suite, native and rebuilt-wheel results above satisfy the local completion conditions. The signatures cover this change and its stated checks, not all deployment platforms. Earlier lifecycle/packaging fixes and their evidence are recorded in the preceding candidate and changelog.

Remaining acceptance gates

Windows physical execution, hosted-model providers, real E2B creation/template builds and an upstream OSWorld VM run remain deferred at the project owner's request. Linux X11 runtime acceptance and remote GitHub Actions results were not observed in this local audit. Physical Desktop serve/remote modes need complete task runs as noted above.

GitHub Actions defines locked tests, Linux browser/X11 tasks, Linux/macOS wheel installations and Linux wheel-only Web tasks. Its static validation passed; a hosted run is still required. Pushes to main and the clean-candidate branch, pull requests and manual runs are configured. Release automation depends on CI and produces wheels, source, documentation and checksums. No package or GitHub release has been published.

The bundled 97-page report is a separately identified research draft. Its pending benchmark tables are not populated from repository smoke tests. See the paper map and clean scope for implementation boundaries.