Paper → implementation → evidence

This NorthCUA repository provides the NorthCUA Harness component of the NexCUA Technical Report. The bundled 2026-10-03 draft, 97 pages describes a broader system: Harness, Experience Flywheel and Staged Learning. This repository releases the Desktop/Web execution and evaluation foundation described in §3.

The manifest records both the source PDF and distributed snapshot SHA-256. Distribution uses lossless PDF object deduplication and stream compression; page text and sampled renders were checked against the frozen source. The benchmark panels and release-result tables remain pending. No draft number is presented as a measured result of this release.

Harness mechanisms

Current report section	Role in the clean implementation	Code and evidence
§3.1 Unified Execution Across Desktop and Web	Common session/round/step progression with domain-specific observations and actions	northcua/agentkit/runtime/, northcua/agents/; hierarchy, runner and transport tests
§3.1.2 Tool Integration and Dispatch	Typed tool contracts, local/MCP dispatch and post-action feedback	northcua_runtime_shared/tool_contracts/, northcua/runtime/mcp/; schema and execution regressions
§3.1.3 Desktop and Web Instantiations	Native/OSWorld Desktop tools; Playwright screenshots, DOM/ref and tab state	Backend tests and opt-in real-browser/native checks
§3.2.1 Action-Aware Visual Context	Shared overlay/clean allocation; one total Web image budget including SoM	northcua/agentkernel/prompt/, Web prompt/visual_window/message_projector.py; final-message image-count tests
§3.2.2 Task-Preserving Context Compaction	Complete-exchange segmentation, pinned task/current state, summary projection and checkpoints	northcua/agentinfra/context/; context, history and overflow regressions
§3.2.3 Structured Working Memory	Idempotent durable note events, replacement by key, round-scoped active view	northcua/agentkernel/working_memory.py; persistence and real-provider projection tests
§3.3 Agentic Outcome Evaluation	Recorded-evidence retrieval, authorized live probes and a separate final report	Shared/Desktop/Web evaluator modules; lifecycle, round and same-session integration tests
§3.3.2 Task-Grounded, Multidimensional Rubrics	Structured criteria, outcome, execution assessment and evidence references	Evaluator contracts, final-report validation and prompt regressions
§3.4.1 Remote Execution with Centralized Agent Control	Controller runners with leased device tools; in-process and process executors	Controller/device/protocol packages; transport, process and real Web mode checks
§3.4.2 Isolated Environments with Pipelined Preparation	Fresh environments and task/attempt-bound prepared reservations	Batch preparation, assignment and cleanup failure tests
§3.4.3 Completion-Oriented Scheduling and Recovery	Claims, deadlines, backpressure, reconciliation, retries and artifact custody	Batch, controller and sandbox lifecycle regressions

The system tour follows this argument with the current paper's figures. Exact executed modes and reviewer scope are in verification. A test listed here is evidence for its checked contract; it is not a measurement of long-horizon capability or deployment throughput.

Research scope and release scope

Broader research component	Relationship to this repository
Experience Flywheel: grounding examples and intent/environment/verifier expansion	Supplies tasks and data in the research system; data production and generation pipelines are removed here.
Recursive skill-guided exploration and checkpoint-assisted collection	Proposed hard-task collection methods in the current draft; not bundled implementations. Ordinary skill loading remains available.
Mid-training, long-horizon SFT and agentic RL	Training systems and model checkpoints are outside this release. Generic external/managed task interfaces remain for integrations.
Benchmark/model capability panels	Results are pending in the draft. The repository's integration tests do not fill those tables.

Provenance and maintenance

The clean code originated from main 315508f3c08a7dec7f753691f38a609c9b8d0601. Historical implementation references in the research materials identify other revisions; they are not the release's execution evidence. The clean branch adds durable working memory, repairs evaluator resource ownership and makes Web image budgeting explicit. These changes should be attributed to the release rather than retroactively to historical research runs.

When changing a mechanism, update its operational guide, behavioral tests and this map. When reporting an experiment, retain the exact paper/code/NexAU revisions, task source, model/provider settings, observation/action profile, budgets, evaluator configuration and artifacts. Use CITATION.cff for citation metadata.