Practice — Landing the Six-Layer Model in Eight Industry Groups

1. Reading Guide and Positioning

This chapter is one of the two with the heaviest technical weight in the white paper, and it connects with Chapter 3 (Architecture): Chapter 3 answers "what a third-generation Harness looks like", while this chapter answers "when it lands in eight industry groups, how the center of gravity shifts, which layer the bottleneck falls in, and what practices have already been validated".

The eight industry groups are: AI Infra, Embodied Intelligence, Software Engineering, Hardware R&D, Knowledge Collaboration, Data Science, Creative Industries, and Risk & Compliance. All content is based on the research documents this project has already produced (each industry group's README and group-level AGENTS.md / SKILL.md); no new external retrieval is introduced. Every quoted hard data point matches its source document; where a source document is marked [To be verified], the same marking is used here.

A criterion running through the whole chapter (this document's judgment): the same six-layer model has a completely different center-of-gravity distribution in different industries; any scheme that tries to cover all industries with a single generic agent template will fail exactly where rigor matters most. This is precisely the significance of the AI Harness as an "engineered carrying layer" rather than a "generic framework".

2. Overview of the Eight Groups' Landing

2.1. Layer-Emphasis and Bottleneck Matrix

图 2-1｜六层模型六大行业组落地：层重点与瓶颈分布

数据来源：基于本文分析绘制的示意图。

The table below is a mapping analysis based on each group's research documents; it is this document's own summary, not a rating by any standards organization.

2.2. Comparison of the Eight Groups' Ground-Truth Forms

The form of the ground truth (the basis of judgment) is the first variable determining each group's Harness design differences (this document's judgment):

2.3. Comparison of the Eight Groups' Acceptance Anchors

The eight groups' group-level AGENTS.md each establish their own acceptance anchors and sources of truth; their differences are themselves the projection of "layer-emphasis differences" onto the governance end (quoted from each group's group-level AGENTS.md):

Three points of reading (this document's judgment):

1. Acceptance anchors are distributed along the spectrum of "executability → explainability → accountability": Software Engineering is at the far left of the spectrum (executable is trustworthy), Risk & Compliance at the far right (accountable is trustworthy), the other groups in between. This distribution is perfectly positively correlated with each group's L6 weight.
2. The difference in authority tiers of sources of truth determines that each group's first priority for L1 context assembly differs: Software Engineering loads executable evidence first, Risk & Compliance loads currently-in-force norms first (with effective dates and authority tiers), Creative Industries loads asset-anchored context first (must not be trimmed).
3. The L4 weight is proportional to "the reuse cycle of assets": the longer the asset reuse cycle (brand assets, project memory, experiment records), the higher the priority of building L4; the Software Engineering group, with the shortest reuse cycle, instead treats the codebase itself as external memory, with the lowest L4 weight.

3. The AI Infra Group: The Physical Foundation of the L2 Execution Layer

3.1. Layer Emphasis and Bottleneck

The AI Infra group (AI chips / servers and super-nodes / cluster interconnect and data centers / training frameworks / inference frameworks / operations and observability) covers the physical foundation of the L2 execution layer in the Harness six-layer model: chips and super-nodes set the physical ceiling of what the agent "can do", training and inference frameworks are the scheduling hub of L2/L3, and operations and observability correspond to L5 and L6 (source document's claim). This group forms a "design → run" closed loop with the Hardware R&D group in Chapter 6 — the design side's packaging (CoWoS, HBM) determines the runtime side's VRAM ceiling, and the runtime side's failure profile in turn drives the design side's reliability goals.

The group's structural tension is characterized by three sets of data (source document; tiers annotated):

Conclusion (this document's judgment): the entire engineering value of the AI Infra group lies in optimizing "fast compute", "affordable compute", and "stable compute" simultaneously — and none of the three lives inside the model layer; they can only be carried by the Harness's orchestration, observability, and governance layers.

3.2. Key Practices and Cases

Case 1: The "logical single-machine" of the super-node — the boundary of the L2 execution unit is redefined (Tier A, official criterion). Huawei's HC 2025 gave this group's most concise claim: "The super-node has become the dominant product form... the super-node is in fact a computer that can learn, think, and reason — physically composed of multiple machines, but logically one machine." This statement maps directly to the Harness's L2: when hundreds of accelerator cards present "single-machine" semantics through unified memory addressing, the model side's long-context carrying, KV sharing, and in-domain scheduling are all simplified into local operations — the boundary of the L2 execution layer is thereby moved up from "one card" to "one super-node".

Case 2: CloudMatrix 384's system-level benchmarking (Tier A, paper measurement). Under the constraint of process limits, the domestic route chose "the system compensates for the single point": 384 Ascend 910C cards interconnected via the UB network form a rack-level unit with system-dense BF16 of about 300 PFLOPS; the engineering details of PDC (Prefill-Decode decoupling) and EP320 expert parallelism are public in the paper (arXiv 2506.12708). This group's unified requirement: domestic parameter citations must state whether they are paper measurements or launch-event criteria — the two have different evidence levels.

Case 3: MegaScale's five-step fault-tolerance pipeline (Tier A, arXiv 2402.15627). The actual availability of ten-thousand-card training depends on "whether failures can be downgraded from global events to local events": heartbeat detection → self-check → isolation → replacement → checkpoint recovery. In echo, Llama 3 combines asynchronous checkpointing with hot-standby nodes, keeping effective training time at a high level. This pipeline is a direct embodiment of L3 orchestration and L4 state memory at the infrastructure layer — fault tolerance is not an operations action but an orchestration design.

Case 4: Form-factor selection driven by the PUE red line (Tier A, policy original text). The hard metric of PUE ≤ 1.25 / 1.2 turns power distribution and cooling from a "data-center specialty" into a front-loaded constraint on chips and complete machines: all-optical inter-rack interconnect, mid-plane designs without cables, water pipes, or fans, and mandatory liquid cooling are all products of this constraint. The GB200 NVL72 liquid-cooling scheme reduces cost and energy by 25× (vendor criterion, Tier A).

3.3. Landing Points

1. Downgrade failures to local events: fault-tolerant orchestration (detection → isolation → replacement → checkpoint recovery) is a first-class engineering capability; the measurement criterion is the effective-training-time ratio (ETTR), not peak compute.
2. Energy and compute have equal weight: energy budgets, PUE constraints, and change windows must be written into the group-level AGENTS.md red-line list, alongside the "forbidden list of irreversible actions".
3. Criterion marking is trustworthiness infrastructure: any performance number must be marked "official release / MLPerf / third-party measurement / vendor claim" plus the hardware generation; vague "performance improvement" statements are uniformly forbidden within the group.
4. Localization: the system compensates for the single point: under long-term process limits, the competitive focus shifts from single chips to system-level (super-node, interconnect, data format) innovation.
5. The most common failure path: treating Infra as "data-center matters" unrelated to the agent — in fact it determines how long a context L1 can hold, how complex a toolchain L2 can run, and whether L5 can obtain a trustworthy efficiency baseline.

4. The Embodied Intelligence Group: The Agent Runtime of the Physical World

4.1. Layer Emphasis and Bottleneck

The Embodied Intelligence group (cerebrum / cerebellum / servo motors and actuators / materials and sensors) is the group with the strongest explanatory power of the six-layer model and, at the same time, the highest cost of constraints (source document's claim). Its division has clear policy basis: MIIT's "Guiding Opinions on the Innovative Development of Humanoid Robots" (2023-11) divides the key technologies into three technology clusters — "cerebrum, cerebellum, limbs" — and the "Standard System for Humanoid Robots and Embodied Intelligence (2026 Edition)" further covers the entire industry chain with six major sections.

The group's core claim is the "cerebrum/cerebellum separation architecture" — a direct mapping of the Harness Loop into the physical world:

Evidence anchors: Figure Helix (2025-02) slow system 7–9 Hz, fast system 200 Hz; NVIDIA Isaac GR00T N1 (2025-03-18) System 2 is VLM reasoning planning, System 1 is a Diffusion Transformer generating continuous actions; the reverse route also exists — Galbot's self-developed "Galbot Star Brain" claims "cerebrum-cerebellum-neural control integrated into one model" (company criterion). The two routes are presented in parallel; this document does not presuppose a single conclusion.

Bottleneck identification (source document's claim): the cerebra and cerebella's bottleneck is at L5 (a structural gap exists between the 94%~97% on LIBERO and the real success rate in the open world; benchmarks such as HumanoidBench cover only the simulation side); the actuators' bottleneck is at L2 (once torque density, backlash, and response bandwidth are fixed, upper-layer algorithms cannot break through); the materials and sensors' bottleneck is at L1 (sensing data quality determines the physical ceiling of "what is seen"); the whole group's governance floor is at L6 — in the physical domain the cost of error is equipment damage and personal safety; ISO 10218:2025 (released 2025-02, in effect 2025-04-01, merged into the collaborative-safety requirements of ISO/TS 15066 including PFL) is this group's L6 anchor standard.

4.2. Key Practices and Cases

Case 1: Simulation first and the data pyramid (source document's engineering scheme). Real-machine data is extremely scarce (Mobile ALOHA body 75 kg, single-arm payload 750 g; teleoperation collection is extremely slow), so the industry adopts a "web video → synthetic data → real-machine teleoperation" data pyramid: GR00T N1's synthetic data brings a +40% performance gain and a 76.8% average real-machine success rate (Tier A); ALOHA-style low-cost master-slave arms + the ACT algorithm can collect about 50 data points in about 10 minutes, reaching an 80%~90% success rate; AgiBot World provides a million-scale shared dataset (1M+ trajectories, 217 tasks, 100 homogeneous robots). Simulation does not replace real machines; it amplifies the marginal value of real-machine data.

Case 2: Explicit distinction of coincidental tier numbers (source document's claim; mandatory across the group). The group standard T/CIE 298-2025 "Grading of Humanoid Robot Intelligence" adopts "four dimensions, five levels" (four dimensions: perceptual cognition, decision learning, collaborative interaction, execution performance; from L1 basic execution level to L5 general intelligence level), which coincides numerically with the Harness six-layer model's L1–L6 but means something entirely different. Mixed use in group writing is forbidden; citations must write out the full standard number — this is the concrete landing of "terminology governance" at the L6 layer.

Case 3: The actuator is a tool contract (source document's claim). Leaderdrive's triple-harmonic reducer precision is ≤10 arc-seconds, torsional stiffness improved 2–3× (Tier B); coreless motor efficiency is 70%–90%, the mainstream choice for dexterous-hand drive; the dexterous hand accounts for about 20%–30% of a humanoid robot's complete-machine cost. Once these hardware metrics are fixed, they directly bound the capability boundary of upper-layer strategies — the essence of actuator selection is writing a "tool contract".

Case 4: Parallel presentation of conflicting criteria (source document's claim). Harmonic-reducer global share has three widely divergent criteria of 40% / 75% / 80%; 2025 global humanoid-robot shipments have multiple criteria such as 16,000 units and nearly 18,000 units (+508% year-on-year). This group's unified requirement: key share and scale data must all be marked with statistical criterion and source nature, presented in parallel across multiple criteria, without merging.

4.3. Landing Points

1. Physical-world safety red line: simulation first + human-in-the-loop throughout real-machine operation; autonomous execution of irreversible physical actions is forbidden. This is a hard constraint essentially different from the Software Engineering group's "sandbox + approval" — the cost of error is physical damage and personal safety.
2. Safety norms are the L6 anchor: ISO 10218-1/2:2025 (including PFL power and force limits); actuator tests per GB/T 43200-2023; six-axis force calibration per GB/T 43199-2023; legged-robot performance tests per GB/T 44251-2024. Norm citations must write out the full standard number and edition.
3. Dual reporting of benchmarks and real machines: reporting only simulation benchmark scores (LIBERO / HumanoidBench) is insufficient to prove capability; real-machine acceptance results must be reported simultaneously; the sim2real transfer success rate is the core regression metric.
4. Asset-ization of data: real-machine trajectories, calibration records, and strategy checkpoints are long-term assets and must have unique identifiers, versions, and responsible persons; teleoperation data must be collected and used in compliance.
5. Beware tier-terminology confusion: T/CIE 298-2025's L1–L5 and the Harness's L1–L6 must be explicitly distinguished; this is the most easily made, and most easily spread, statement error in this group.

5. The Software Engineering Group: The Native Domain and the L5 Frontier

5.1. Layer Emphasis and Bottleneck

The Software Engineering group is the only conceptual native domain in the entire industry landscape of the AI Harness — the word Harness is borrowed directly from software engineering's Test Harness. The six layers develop in balance in this group, but L5 evaluation is at the frontier: code has compilers, tests, and CI, making this the only agent scenario at scale where "programmatic scoring" has landed.

The group's structural tension is characterized by three data points (Tier-A sources):

Conclusion (this document's judgment, following the source document): both capability and adoption are in place; the entire remaining value of the AI Harness in the software engineering field is almost concentrated in completing the measurement side.

5.2. Key Practices and Cases

Case 1: OpenAI's million-line code experiment (Tier A). First commit to an empty repository in late 2025-08; the scaffold was generated by Codex CLI + GPT-5; five months later about 1 million lines of code, about 1,500 PRs, 3 engineers (later expanded to 7), 3.5 PRs per person per day, time cost about 1/10 of hand-writing, 0 lines of human hand-written code, and the longest single task over 6 hours. The six key decisions (AGENTS.md about 100 lines as a table of contents, linter error messages written as repair instructions, doc-gardening agent, mechanical enforcement of the dependency chain, observability open to the agent, minimal-blocking merge) fall respectively on L1 / L6→L1 / L6 / L6 / L5→L1 / L3; the detailed breakdown is in Chapter 3, section 10.2.

Case 2: Anthropic's two-generation long-running Harness solutions (Tier A). From the Initializer + Coding Agent dual-Agent division of labor (handoff package + three opening steps + clean-state wrap-up) to the Planner / Generator / Evaluator three roles (Sprint Contract + independent Evaluator + hard thresholds), it demonstrates the coordinated design of L3 / L4 / L5.

Case 3: The measurement lesson of the METR controlled trial (Tier A). The divergence between measured 19% slower and self-rated 20% faster shows: self-rating measures the felt speed of "the code-writing segment", while measurement is end-to-end throughput; AI compressed the former while amplifying the cost of verification, review, and handling "almost-right-but-not-quite" products. In Stack Overflow 2025, 66% of respondents listed "AI solutions are almost right but not quite" as their biggest frustration — this is not a capability problem but a verifiability problem.

Case 4: Quantification of scaffold differences (Tier B/C). The same model scores 2~14 percentage points apart across different Harnesses; LangChain changed only the Harness (same model, same API) and rose from 52.8% to 66.5%, with all five changes involving no model swap.

5.3. Landing Points

1. Dual-layer gates: product gates (compilable, testable, deployable, rollback-able) and process gates (trajectory evaluation, step caps, permission-hit checks) coexist; neither can be missing.
2. Unify measurement to end-to-end: anchored on the four DORA metrics, forbid using lines of code or completion adoption rate as effectiveness conclusions; any efficiency conclusion must have a controlled measurement.
3. Acceptance anchors: compilable, testable, deployable, rollback-able; process evidence (Trace, scorecards, human-confirmation records) is archived with the change.
4. The most common failure path: skipping Eval and building Coding directly — without a regression set and trajectory evaluation, every prompt change is "by feel".

6. The Hardware R&D Group: L2 and L5 Heaviest, L6 Cost Highest

6.1. Layer Emphasis and Bottleneck

The Hardware R&D group is the touchstone best able to test "whether the Harness truly converts the model's uncertainty into engineering predictability", because it simultaneously has two extreme features (source document's claim):

1. The feedback closed loop is the most expensive: a single advanced-node tape-out is priced in hundreds of millions of dollars (IBS estimates 5nm~3nm at over $200M–$500M per chip, Tier B); design errors cannot be hot-fixed like software — this pushes L6 governance to the highest priority.
2. The ground truth is the hardest: simulators, formal verification engines, ATE test machines, and metrology equipment can all give deterministic pass-or-fail verdicts — this means L5 does not need to be "built extra" but should directly reuse coverage databases, regression farms, and formal proofs.

The core claim at the L2 level: EDA tools are irreplaceable; the agent's correct posture is not "replace the engine" but "orchestrate the engine." Cadence ChipStack orchestrates Xcelium (dynamic simulation) and Jasper (formal verification) inside an NVIDIA OpenShell sandbox; Siemens built its own MCP server exposing Questa One engine capabilities to agent frameworks, explicitly emphasizing "hooking into existing simulation engines rather than replacing them".

The core claim at the L5 level: formal verification and logic equivalence checking (LEC) are the only decidable ground truth. Any netlist after AI optimization must undergo a formal equivalence check against the golden RTL, confirming functional behavior is unchanged, before it can enter subsequent flows.

Per-direction bottleneck-layer distribution: 01 chip design at L5 (long design feedback cycles, proxy metrics decoupled from real performance); 02 chip verification at L2 (heterogeneous toolchains, licensed engines); 03 packaging at L1 (multi-physics, multi-vendor heterogeneous data hard to express uniformly); 04 testing at L6 (yield data is a core asset, forbidden from leaving the enterprise boundary); 05 AI Infra at L3 (at ten-thousand-card scale, failures become the norm).

6.2. Key Practices and Cases

Benchmark gaps and anti-cheating (source document's claim three): GPT-4o's spec-to-RTL pass@1 on VerilogEval v2 is 63%; but per compilations of public talks, scores drop to 10%~40% on the CVDP benchmark, which is closer to production, and the hardest RealBench has still had no AI system solve any single problem (Tier C). This group therefore requires: reporting only academic benchmark scores is forbidden; real in-project regression metrics must be reported simultaneously. In addition, no public EDA Agent benchmark covering the full "tool invocation + multi-round debugging + coverage convergence" flow has been found — this is a real and important industry gap, meaning this domain cannot compare different Harness implementations horizontally the way software engineering relies on SWE-bench, and can only rely on enterprises' internal private regression sets.

6.3. Landing Points

1. Take the EDA engine as the sole referee: the model only produces candidate solutions; all candidate solutions must pass a formal equivalence check before entering sign-off; the model must not self-judge "it should be fine".
2. Human-in-the-loop + strong sandbox + strong audit: irreversible actions such as tape-out, sign-off, and mass-production release require human confirmation + dual sign-off; agents have no direct trigger permission by default; autonomous agents must run in a sandboxed runtime (the NVIDIA OpenShell paradigm: policy control, isolation, controlled access to tools and design data).
3. Design data does not leave the boundary: RTL, netlists, layouts, test programs, and yield data are the enterprise's core IP, encrypted and managed per IEEE 1735-2023.
4. Coverage cannot be padded: distinguish code coverage from functional coverage and identify unreachable bins, otherwise "100% coverage but design flaws remain" false convergence occurs; SVA assertions generated by LLMs must be verified by formal tools.

7. The Knowledge Collaboration Group: Mainly L1 and L3

7.1. Layer Emphasis and Bottleneck

The first batch of real scenarios for enterprises landing large models are, in the vast majority, not code-writing but: finding a document, answering a question, drafting a material, processing a ticket, running a process through. The capability bottleneck of these scenarios is not in the model itself but in the carrying layer between the model and the organization's existing knowledge assets and process assets (source document's claim).

Seven directions (Projects / Works / Docs / RAG / Knowledge / Workflow / BPM) form a three-level progression of "information reachable → work orchestrable → processes governable", corresponding to the path of L1 moving up to L3, then to L6. There are two bottlenecks:

• L1 (retrieval and context engineering): retrieval quality sets the ceiling of generation quality; structured knowledge is required, not optional, for complex queries; long context cannot replace context engineering.
• L6 (permission-aware retrieval): permissions must land at the retrieval layer, not only at the application layer.

7.2. Key Practices and Cases

Case 1: The engineering payoff of Contextual Retrieval (Anthropic official, Tier A). Under the recall@20 criterion, the baseline retrieval failure rate is 5.7%; adopting Contextual Embeddings reduces it to 3.7%; adding Contextual BM25 reduces it to 2.9%; adding reranking reduces it to 1.9%, officially phrased as "reduce the number of failed retrievals by 49% and, when combined with reranking, by 67%". The one-time preprocessing cost is about $1.02 per million document tokens. This payoff comes from the engineering layer, not the model layer.

Case 2: The order-of-magnitude lift of structured knowledge (data.world's own benchmark, Tier-A criterion but with stated limitations). On 43 business questions, the bare LLM's average execution accuracy is only 16.7%, with the two high-schema-complexity categories "metrics and KPIs" and "strategic planning" at 0%; connecting a knowledge graph triples accuracy; adding OBQC (Ontology-based Query Check) and LLM Repair reaches 72.55%, a 4.2× improvement over pure SQL. Note: this benchmark has 43 questions, 13 tables, a single insurance domain, GPT-4 zero-shot, and has not been reproduced at scale by third parties; it should be viewed as an upper-bound reference of the self-built method.

Case 3: The physical limits of long context (academic paper, Tier A). Lost in the Middle (Liu et al., TACL 2024): multi-document QA shows a U-shaped curve; in some configurations, burying the relevant information in the middle performs even worse than closed-book — the retrieved context is a net negative. RULER (NVIDIA, COLM 2024): of 17 long-context models, only about half maintain satisfactory performance at 32K. NoLiMa (ICML 2025): of 13 models claiming ≥128K support, 11 fall below half their own short-context accuracy at 32K; GPT-4o drops from 99.3% to 69.7%.

Case 4: The double-edged sword of permission inheritance (GOV.UK official report, Tier A). The UK government's cross-department M365 Copilot experiment: the tool "adopts the permissions of the end user and will only retrieve documents that a user could normally access", while officials simultaneously warn that "Copilot's ability to search and query using a user's access may highlight issues when users have access to files they shouldn't". Conclusion: permission inheritance is not "safe once configured" but amplifies existing permission-configuration errors.

7.3. Landing Points

1. Retrieval first, generation after: assertions without recall support are never output ("no source, no answer"); set a Recall@K evaluation gate on the retrieval chain; retrieval-strategy changes are scored before going live.
2. Small-corpus exemption: when the knowledge base is under 200,000 tokens (about 500 pages), the full text can go directly into the prompt without RAG (Anthropic official original text); beyond that, the retrieval chain is mandatory.
3. Permission-aware retrieval: do result filtering at the retrieval layer; make the confidentiality level an index field rather than an after-the-fact label; periodically run red-team tests with "unauthorized retrieval use cases".
4. Standard absence is the norm; evaluation sets are the fill-in-the-gap means: of the seven directions, only Knowledge (ISO 30401:2018) and BPM (BPMN / CMMN / DMN / IEEE 1849 / OCEL 2.0) have formal standards; RAG, Docs, and Works have no corresponding ISO/IEEE formal standards, so an internal Golden Dataset and gate thresholds should be established.
5. The most common failure path: skipping the first layer (information reachability) and going straight to building the third layer (process governance) — without trustworthy retrieval and knowledge governance, process automation only scales up errors.

8. The Data Science Group: AI4S's ground truth is physical law

8.1. Layer Emphasis and Bottleneck

Data science group tasks have three structural characteristics (source document's claims): the product is not text but executable assertions (a SQL query, a simulation, a causal estimate all run on real systems and can be directly verified or falsified); errors are silent (a crashed pipeline immediately tells you it's broken, but a miscalculated pipeline tells you nothing); constraints come from the external world (permissions are determined by data classification, correctness is determined by physical law or business criteria, cost is determined by compute quota).

The six-layer emphasis differs greatly across the six directions: 01 Data Engineering at L1 + L6; 02 Data Analysis at L1 + L5; 03 Deep Research at L1 + L3; 04 Scientific Computing at L2 + L4; 05 High-Performance Computing at L2 + L3; 06 Scientific Discovery (AI4S) at L5 + L3. The most essential difference is at L5: AI4S's ground truth is not human annotation but physical law, experimental verification, and conservation-law constraints.

8.2. Key Practices and Cases

Case 1: Argonne National Laboratory's Aurora agent material screening — "the LLM never touches the scheduler". In the agent material screening work on Aurora, MCP tools do not run simulations directly but instead emit Parsl applications, with Parsl responsible for placement, concurrency, and fault tolerance. This is the best empirical evidence of the L2/L3 separation and establishes this domain's default architecture: the agent should not directly invoke the scheduler but should produce job descriptions that can be carried by a workflow engine.

Case 2: The cost of silent errors (source document's claims). A miscalculated pipeline does not report an error and appears with correctly formatted numbers and charts; a hallucinated review appears decoratively with complete citation formatting. The countermeasure is not a stronger model but: semantic regression sets + grain assertions + independent recomputation + result-set comparison, rather than relying only on row counts and freshness. Public reporting indicates (second-hand paraphrase) that only 9% of data engineers are satisfied with Airflow DAGs generated by general AI tools, the root cause being that the model lacks context such as the project DAG, execution environment, and existing conventions.

Case 3: Evaluation cost and layered evaluation. A single scientific simulation can take 1,600~4,400 seconds, making high-frequency regression impossible; the countermeasure is layered evaluation: cheap proxy metrics first, expensive verification triggered on demand, proxy metrics periodically calibrated against ground truth. Another research report states that scheduling 100 jobs via a cloud API service takes 1~2 hours, limiting real-time deployment — the countermeasure is forced separation of inference and constraints: the model is responsible for inference, and deterministic modules are responsible for feasibility and safety.

Case 4: The four-level reproducibility definition as the acceptance criterion. ACM gives a four-level terminology for reproducibility: Repeatability / Replicability / Reproducibility / Reusability; Harness's L4 should directly take these four levels as the acceptance criterion for memory-structure design — input data versions are traceable, execution environments are rebuildable, random seeds are controllable, and conclusions are recomputable.

8.3. Landing Points

1. Don't treat "looks right" as "right": make executable verification (SQL comparison, physical conservation-law checks, independent recomputation) a first-class citizen; pair analysis tasks with verified-queries regression sets, and look at Valid Efficiency Score (VES) in addition to Execution Accuracy (EX) — a query that returns correct rows via a full table scan is a production incident waiting to happen.
2. Data classification is a hard boundary: RLS/CLS, PII masking, and cross-border restrictions are enforced at L6; the model should never be the judge of permissions.
3. Reuse rather than reinvent orchestration: Airflow / dbt / Parsl / Slurm are data science's native orchestration forms; the value of the Harness lies in translating natural-language goals into job graphs these engines can digest.
4. Fallback plan when ground truth is missing: the deep research direction has no standard answers, so adopt a hard DoD of citation verifiability rate, external benchmarks, and reviewer-agent grading + human supervision.

9. The Creative Industries Group: L4 heaviest, consistency is cross-session state

9.1. Layer Emphasis and Bottleneck

The creative industries group's six directions (intelligent manufacturing / industrial internet / media / creative design / AI web dramas / AI animation) are divided into two categories by the difference in decidability of the deliverable (source document's claims):

• Content side (media, creative design, AI web dramas, AI animation): the deliverable is a content asset, and a large part of the judgment criteria is "is it still the same person, the same brand, the same tone" — this is a class of cross-session, cross-shot, cross-episode state-preservation problem, so the L4 Memory and State layer is the quality bottleneck.
• Industrial side (intelligent manufacturing, industrial internet): the deliverable is a state change in the physical world, and the judgment criteria are physical quantities such as yield, miss rate, and downtime, so L5 and L6 are the landing bottlenecks.

The two sides share the same L6 baseline: the "Measures for Labeling AI-Generated Synthetic Content" (Guoxinban Tongzi [2025] No. 2) takes effect from 2025-09-01, requiring AI-generated synthetic content to carry explicit and implicit labels; Article 9 requires that related log records be retained for no less than six months; Article 10 explicitly forbids malicious deletion, tampering, forgery, or concealment of labels. Compliance obligations have already sunk to the level of metadata fields and export scripts — "writing metadata", "applying corner marks", and "keeping logs" are no longer operations actions but the Harness's L2 tool contract and L6 governance red line.

9.2. Key Practices and Cases

Case 1: Coca-Cola's holiday marketing — the engineering chasm between "clip quality" and "final-film quality". In the 2025 holiday marketing, 5 AI experts spent 30 days fine-tuning over 70,000 video clips (per behind-the-scenes reporting, not an official disclosure) before pulling the generative video model up to broadcast standard; the 2024 version was widely criticized for distorted human figures. This shows that there is a huge engineering chasm between the clip quality and the final-film quality output by generative models, and it is precisely the Harness that fills this chasm.

Case 2: Scale crossing the critical point of human backstop. According to the China Netcasting Services Association's "Micro-Short-Drama Creation Guidelines" (cited from People's Daily, 2026), in Q1 2026 the whole industry launched about 128,000 micro-short-dramas, of which about 122,000 were AI micro-short-dramas, an share exceeding 95%. At this scale, any quality system relying on human item-by-item review will fail; consistency verification, label verification, and pre-review must be made into automatic gates on the pipeline.

Case 3: Scripting deterministic operations. Anthropic's description of Agent Skills gives an engineering principle: having the model generate sorting results token by token is far more expensive than directly running a sorting algorithm, and many applications need "what only code can provide" — deterministic operations must be scripted, not generated by the model.

9.3. Landing Points

1. Consistency is a state-preservation problem: character, brand, and tone consistency must be treated as cross-session state, managed via L4 (asset library + version + consistency verification), not left to the model's in-context memory.
2. Compliance obligations sink to tool contracts: label application, metadata writing, and log retention must be built into L2 tools and enforced at L6, not relying on human memory.
3. Scale forces automation: at the 100,000-works scale, any quality system relying on human item-by-item review will fail; consistency verification, label verification, and pre-review must be made into automatic gates on the pipeline.
4. Script deterministic operations: deterministic operations (sorting, formatting, batch processing) must be scripted, not generated by the model.

10. The Risk & Compliance Group: L6 heaviest, the output is the deliverable

10.1. Layer Emphasis and Bottleneck

The Risk & Compliance group is the group where L6 is heaviest: here L6 is not a constraint but a product — AI output is examined by third parties as evidence (source document's claims). The group's core claim is: the deliverable of risk and compliance work is not a "conclusion" but a "provable conclusion" — every conclusion must be accompanied by a restorable evidence chain.

The group's structural tension is characterized by three features (source document's claims):

• The judge is external: the acceptability of the deliverable is not determined by the model but by auditors, regulators, and courts — this pushes L6 governance to the highest priority.
• The evidence chain is the deliverable: every conclusion must be accompanied by a restorable evidence chain (data source, calculation process, model version, human sign-off) — this makes L4 memory and L5 observability a hard requirement rather than an option.
• Errors are costly and irreversible: a compliance error may lead to regulatory penalties, litigation losses, or reputational damage — this pushes L6's human-in-the-loop to a mandatory rather than optional position.

10.2. Key Practices and Cases

Case 1: The evidence-chain five-tuple as the acceptance criterion (source document's claims). The group's unified acceptance criterion is the "evidence-chain five-tuple": data source traceable, calculation process reproducible, model version identifiable, human sign-off present, timestamp verifiable. Every deliverable must satisfy all five items; any missing item means the deliverable is not acceptable.

Case 2: Human-in-the-loop as a mandatory gate (source document's claims). In risk and compliance scenarios, human sign-off is not an optional step but a mandatory gate: every conclusion that leaves the system must be signed off by a qualified human; the model's role is to provide supporting evidence and draft conclusions, not to make final judgments.

Case 3: Audit logging as a first-class requirement (source document's claims). Audit logging is not an afterthought but a first-class requirement: every data access, every calculation, every model invocation, and every human decision must be logged with full context; the logs must be immutable and retained for the required period.

10.3. Landing Points

1. The evidence chain is the deliverable: every conclusion must be accompanied by a restorable evidence chain (data source, calculation process, model version, human sign-off, timestamp); any missing item means the deliverable is not acceptable.
2. Human-in-the-loop is a mandatory gate: every conclusion that leaves the system must be signed off by a qualified human; the model provides supporting evidence and draft conclusions, not final judgments.
3. Audit logging is a first-class requirement: every data access, calculation, model invocation, and human decision must be logged with full context; logs must be immutable and retained for the required period.
4. Data does not leave the domain: sensitive data must not leave the controlled domain; model invocations must be within the authorized boundary.

11. Extracting Common Patterns Across Industries

11.1. The Universal Structure of the Six-Layer Model

Across the eight industry groups, the six-layer model exhibits a universal structure (this document's judgment):

1. L1 (Context Engineering) is the universal entry point: every group needs to decide "what the model sees"; the form of context assembly differs by group (executable evidence for software engineering, currently-in-force norms for risk and compliance, asset-anchored context for creative industries), but the need is universal.
2. L2 (Tooling and Execution) is the universal interface: every group needs to decide "what the model can do"; the form of tool contracts differs by group (EDA engines for hardware, sandboxed executors for software, workflow engines for data science), but the need is universal.
3. L3 (Orchestration and Control) is the universal coordinator: every group needs to decide "in what order to act"; the form of orchestration differs by group (plan-and-execute for software, real-time loops for embodied intelligence, deterministic workflows for data science), but the need is universal.
4. L4 (Memory and State) is the universal accumulator: every group needs to decide "what to remember"; the form of memory differs by group (codebase for software, asset library for creative industries, experiment records for data science), but the need is universal.
5. L5 (Evaluation and Observability) is the universal referee: every group needs to decide "how well it was done"; the form of evaluation differs by group (programmatic scoring for software, physical verification for embodied intelligence, evidence-chain restoration for risk and compliance), but the need is universal.
6. L6 (Governance and Safety) is the universal guardrail: every group needs to decide "what must not be done"; the form of governance differs by group (sandbox + approval for software, physical safety for embodied intelligence, human-in-the-loop for risk and compliance), but the need is universal.

11.2. The Differentiation Axis: Ground-Truth Form

The primary axis of differentiation across the eight groups is the form of the ground truth (this document's judgment):

The form of the ground truth determines the form of L5 evaluation, which in turn determines the form of L6 governance. This is the primary axis of differentiation across the eight groups.

11.3. The Differentiation Axis: Center-of-Gravity Distribution

The secondary axis of differentiation is the center-of-gravity distribution across the six layers (this document's judgment):

• L2-heaviest groups: AI Infra, Hardware R&D, Embodied Intelligence — the physical foundation and tooling are the primary constraints.
• L4-heaviest groups: Creative Industries, Knowledge Collaboration — state preservation and memory accumulation are the primary constraints.
• L5-heaviest groups: Software Engineering, Data Science — evaluation and observability are the primary constraints.
• L6-heaviest groups: Risk & Compliance, Embodied Intelligence — governance and safety are the primary constraints.

11.4. The Universal Failure Path

Across the eight groups, the most common failure path is the same (this document's judgment): skipping the foundational layer and going straight to the advanced layer. For example, skipping L1 (context engineering) and going straight to L3 (orchestration), or skipping L5 (evaluation) and going straight to L6 (governance). The foundational layers must be built first; the advanced layers depend on them.

12. Landing Path and Phased Recommendations

12.1. The Three-Phase Landing Path

Based on the eight groups' experiences, the landing path can be divided into three phases (this document's judgment):

1. Phase 1: Build the foundation (L1 + L2): establish context engineering and tooling interfaces. This is the prerequisite for all subsequent phases.
2. Phase 2: Build the loop (L3 + L4): establish orchestration and state management. This is the core of the Harness.
3. Phase 3: Build the guardrails (L5 + L6): establish evaluation and governance. This is the guarantee of reliability.

12.2. Phase-Specific Recommendations

12.3. Common Pitfalls and Countermeasures

1. Pitfall: skipping the foundation: going straight to advanced layers without building L1 + L2. Countermeasure: build the foundation first; the advanced layers depend on them.
2. Pitfall: over-engineering: building all six layers at once without prioritizing. Countermeasure: follow the three-phase path; build incrementally.
3. Pitfall: under-investing in L5: neglecting evaluation and observability. Countermeasure: L5 is not optional; it is the guarantee of reliability.
4. Pitfall: ignoring L6: neglecting governance and safety. Countermeasure: L6 is not a constraint but a product; it must be built from the start.

13. Conclusion

The six-layer model exhibits a universal structure across the eight industry groups, but the center-of-gravity distribution differs by group. The primary axis of differentiation is the form of the ground truth, and the secondary axis is the center-of-gravity distribution across the six layers. The landing path should follow the three-phase structure: build the foundation (L1 + L2), build the loop (L3 + L4), and build the guardrails (L5 + L6). The most common failure path is skipping the foundational layer and going straight to the advanced layer.

For the detailed landing practices of each industry group, see the corresponding chapters of this white paper. For the future evolution of the six-layer model, see Chapter 5 (Outlook).

14. Information Gap Statement

This document involves the following information gaps; they have been flagged in the corresponding places in the body and are summarized here:

1. Section 3.1: the national intelligent-compute scale (FP16) data (1590 EFLOPS at end of 2025, 1882 EFLOPS at end of March 2026) is per MIIT data (Tier B).
2. Section 3.1: the compute-center energy consumption data (196 billion kWh in 2025, +18.1% year-on-year; 2030 expected to exceed 500 billion kWh) is per CAICT criterion (Tier B).
3. Section 3.2: CloudMatrix 384's system-dense BF16 performance (about 300 PFLOPS) is per paper measurement (Tier A, arXiv 2506.12708).
4. Section 4.2: harmonic-reducer global share has three widely divergent criteria (40% / 75% / 80%); 2025 global humanoid-robot shipments have multiple criteria (16,000 units, nearly 18,000 units, +508% year-on-year).
5. Section 5.2: the METR controlled trial data (measured 19% slower, self-rated 20% faster) is per the original paper (Tier A).
6. Section 6.2: the AI-enabled EDA design cost savings (20%~30% forecast, 21.3% representative measurement) are per Semico Research (Tier B).
7. Section 7.2: data.world's own benchmark (43 questions, 13 tables, single insurance domain, GPT-4 zero-shot) has not been reproduced at scale by third parties.
8. Section 8.2: the 9% satisfaction rate for Airflow DAGs generated by general AI tools is per public reporting (second-hand paraphrase).
9. Section 9.2: Coca-Cola's holiday marketing data (5 AI experts, 30 days, over 70,000 video clips) is per behind-the-scenes reporting, not an official disclosure.
10. Section 10.2: the evidence-chain five-tuple as the acceptance criterion is this document's own summary, not a formal standard.

15. References

1. Effective context engineering for AI agents — Anthropic, 2025. https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents
2. Effective harnesses for long-running agents — Anthropic, 2025. https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents
3. Harness engineering: leveraging Codex in an agent-first world — OpenAI, 2026-02-11. https://openai.com/index/harness-engineering/
4. METR: Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity — METR, 2025-07-10. https://metr.org/blog/2025-07-10-early-2025-ai-experienced-osdevs/
5. DORA Report 2025: The State of AI-Assisted Software Development — Google DORA, 2025-11-12. https://dora.dev/research/
6. 2025 Stack Overflow Developer Survey — Stack Overflow, 2025-07-29. https://survey.stackoverflow.co/2025/
7. Lost in the Middle: How Language Models Use Long Contexts — Liu et al., TACL 2024. https://arxiv.org/abs/2307.03172
8. RULER: What's the Real Context Size of Your Long-Context Language Models? — Hsieh et al., COLM 2024. https://arxiv.org/abs/2404.06654
9. NoLiMa: A Simple Yet Complex-Benchmark for Long-Context LLMs — ICML 2025. https://arxiv.org/abs/2502.05167
10. Microsoft 365 Copilot: Cross-Department Trial Report — GOV.UK, 2025. https://www.gov.uk/government/publications/microsoft-365-copilot-trial-report
11. Agentic MOF Screening on Aurora ("the LLM never touches the scheduler") — supercomputing.news. https://www.supercomputing.news/hpc/agentic-mof-screening-aurora
12. AI 要给微短剧"洗牌"？ — 人民日报，2026。https://kpzg.people.com.cn/n1/2026/0511/c404214-40717117.html
13. 中注协提示会计师事务所在 2025 年年报审计中使用人工智能技术的风险防范 — 中国注册会计师协会，2026-03-05。https://cicpa.org.cn/
14. GB/T 45654—2025《网络安全技术 生成式人工智能服务安全基本要求》 — SAC/TC260，2025。https://www.tc260.org.cn/
15. 《人工智能生成合成内容标识办法》（国信办通字〔2025〕2 号） — 国家互联网信息办公室，2025。https://www.cac.gov.cn/
16. GB/T 43200-2023《机器人 工业机器人性能规范及其试验方法》 — SAC，2023。https://openstd.samr.gov.cn/
17. GB/T 43199-2023《机器人 六维力/力矩传感器 性能规范及试验方法》 — SAC，2023。https://openstd.samr.gov.cn/
18. ISO 10218-1/2:2025《Robotics — Safety requirements for industrial robots》 — ISO，2025。https://www.iso.org/standard/77447.html
19. T/CIE 298-2025《人形机器人智能化分级》 — 中国电子学会，2025。https://www.cie.org.cn/
20. 工信部《人形机器人创新发展指导意见》 — 工业和信息化部，2023-11。https://www.miit.gov.cn/