OPEN-WEIGHT LLM · RAG · LORA · EVIDENCE GOVERNANCE 小懿 AI · 港航垂直行业生成式大模型研发项目 Xiaoyi AI · Open-Weight Maritime LLM with RAG, LoRA & Evidence Governance 研发作者：温家懿 · Research Author: Wen Jiayi 把港航问答从“模型说了什么”，提升为“证据来自哪里、适用于哪个辖区和日期、何时必须拒答、谁允许推进任务”的可审计决策链。 A local generative maritime model stack where retrieval, adaptation, jurisdiction, temporal applicability, refusal, and audit evidence remain independently reviewable. 129 登记文档 · 882 知识分块 · 68 官方来源 · 260 题六版本基准

核心指标 / Evidence · 能力全景 / Capabilities · 系统架构 / Architecture · 快速运行 / Quick start · 基准报告 / Benchmark · 决策保障 / Decision assurance · 证据对齐 / Evidence alignment · 日常问答 / Daily operations · 问题全集 / Question universe · 生成式模型栈 / Generative stack · 本机LoRA证据 / Local LoRA evidence · 升级路线 / Roadmap · RL证据 / RL evidence · 开源审计 / Audit

_{固定证据总览：同屏核对三套数据血缘、五算法矩阵、观测/动作/目标、小懿训练顾问和后端按钮联动。}

flowchart LR
  subgraph Sources["Knowledge & Runtime Sources / 知识与运行来源"]
    KB["129 docs / 882 chunks"]
    REG["Authority registry\n官方来源与辖区"]
    OPS["Sandbox or verified live adapter\n沙箱或验证实时适配器"]
  end
  subgraph Intelligence["Evidence Intelligence / 证据智能"]
    HYB["Hybrid Sparse + BM25"]
    ROUTE["Jurisdiction + effective date"]
    POLICY["Citation · refusal · privacy gates"]
    LLM["Local rules / optional compatible gateway"]
  end
  subgraph Workflow["Human-Gated Workflow / 人工门禁工作流"]
    CHAT["SSE chat + structured answer"]
    SOP["SOP task + report"]
    AUDIT["Receipt + audit persistence"]
  end
  KB --> HYB
  REG --> ROUTE --> HYB --> POLICY
  OPS --> POLICY
  POLICY --> LLM --> CHAT
  CHAT --> SOP --> AUDIT

小懿AI/
  app/
    main.py              # FastAPI服务入口 / service entry
    operations.py        # 运营、任务与报告API / operations, tasks, reports
    port_runtime.py      # 沙箱/生产数据适配器 / sandbox/live adapters
    rl_lab/              # 数据、4种RL、PID与评估 / data, 4 RL, PID, evaluation
    operator_assistant.py # 一线口语与对象追问 / frontline language and clarification
    xiaoyi.py            # 小懿回答引擎 / answer engine
    retrieval.py         # Hybrid Sparse与BM25对照 / retrieval and baseline
    prompts.py           # 港航专业提示词 / maritime prompts
    models.py            # 请求与响应模型 / request-response models
    config.py            # 路径与配置 / paths and configuration
  data/
    kb/                  # 港航知识库 / maritime knowledge base
    evaluation/          # 固定检索与证据评测 / pinned retrieval and policy set
    public/              # 公开RL数据与血缘 / public RL data and provenance
    port_profiles/       # 可替换港口参数与边界 / swappable port profiles
    rl_datasets.json     # 数据集目录 / dataset registry
    xiaoyi_index.json    # 检索索引 / retrieval index
  scripts/
    build_index.py       # 构建知识索引 / build knowledge index
    build_vector_index.py # 可恢复的1024维稠密索引 / resumable dense index
    build_lora_dataset.py # 来源隔离的LoRA数据 / source-isolated LoRA data
    train_lora.py        # PEFT LoRA训练与审计报告 / PEFT LoRA training
    export_lora_gguf.py  # 导出llama.cpp LoRA / export GGUF adapter
    fetch_public_rl_dataset.py # 下载并校验公开RL数据 / fetch and verify
    fetch_large_public_rl_dataset.py # 409,887行大规模基准 / large benchmark
    fetch_noaa_port_ais_dataset.py # NOAA港区AIS场景 / public port AIS
    run_rl_dataset_benchmark.py # 三数据集多种子RL证据 / multi-seed RL evidence
    run_rag_benchmark.py # 固定RAG基准与哈希报告 / benchmark and hash report
  web/
    index.html           # Web交互页面 / Web interface
  tests/
    test_retrieval.py    # 检索验证 / retrieval verification
    test_operations_api.py # 运营API与回归 / operations API and regression

推荐直接使用项目启动脚本；首次运行会创建 .venv、安装运行依赖、重建索引并启动服务：
Use the project launcher. On first run it creates .venv, installs runtime dependencies, rebuilds the index, and starts the service:

cd <xiaoyi-ai-repository>
.venv/bin/python scripts/local_model.py install-runtime
export XIAOYI_LOCAL_MODEL_PATH=/absolute/path/to/maritime-generation-4b-q4-k-m.gguf
export XIAOYI_EMBEDDING_MODEL_PATH=/absolute/path/to/maritime-embedding-0.6b-q8-0.gguf
bash run.sh

模型权重不进入公共仓库，部署时通过环境变量提供本地 GGUF 路径；公开注册表仅保留 参数规模、量化、字节数、上下文和可选 SHA-256 校验契约。模型存在且已安装 llama-server 时，run.sh 自动启用“4B 本地生成 + RAG + 回答后证据门禁”； 否则安全回退到确定性严格证据回答器。完整模型规模、LoRA 数据/训练和硬件边界见 生成式大模型方案。

Model weights are not committed. Supply local GGUF paths through environment variables; the public registry retains only parameter count, quantization, artifact size, context limits, and optional checksum contracts.

默认使用 4B 流畅生成档；显式验证本机 LoRA 档：

XIAOYI_LOCAL_MODEL_PROFILE=lora bash run.sh

也可以手动运行：
Or run manually:

cd <xiaoyi-ai-repository>
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.lock
python scripts/build_index.py
uvicorn app.main:app --reload --port 8010

打开：
Open:

http://127.0.0.1:8010

cd <xiaoyi-ai-repository>
python -m app.cli "岸电 THDi 超标告警应该先检查什么？"
python -m app.cli "小懿的核心能力是什么？" --mode expert

小懿是面向港航场景的本地生成式 AI 助手。它采用开源 4B 量化生成档、本地港航知识库和 RAG 检索，结合专业问答、SOP 生成、告警解释和运营建议；同时以可在本机完成反向传播的 1.7B 训练/推理档执行 LoRA，学习港航表达与回答逻辑。1.7B 适配器不会跨参数规模套到 4B。当前事实、私有资料和法规版本仍由 RAG 提供，生成结果继续通过引用、主张和数字完整性门禁。模型不可用或门禁不通过时自动回退到 local_rules 严格证据答案；只有远程模型才需要显式资料外发授权。

Xiaoyi is now an open-weight generative maritime model stack: 4B provides the default local dialogue layer, 0.6B-Embedding adds dense retrieval alongside Sparse/BM25, and an exact-base 1.7B profile supports reproducible LoRA engineering on this Intel Mac. RAG remains authoritative for current, private, and jurisdiction-sensitive facts; citation, claim-alignment, and numeric-integrity gates can reject a generated rewrite and retain the deterministic strict-evidence answer. The 1.7B adapter is never attached to the 4B model, and remote data egress still requires explicit authorization.

运营看板默认使用后端持续生成的运营沙箱事件流。它具有生产形态的业务对象、时间戳、质量码、延迟和来源适配器，但明确标记为非生产实绩。切换生产时保持 port-ops.v1 契约并配置只读数据网关；高风险任务仍返回 requires_human_confirmation: true，不会因接入真实数据自动获得写权限。

The operations dashboard consumes a continuously generated backend sandbox event stream. It uses production-shaped entities, timestamps, quality codes, latency, and source adapters while remaining explicitly labelled non-production. A live integration retains the port-ops.v1 contract and a read-only gateway. High-risk tasks still return requires_human_confirmation: true; real data never grants write authority automatically.

_{证据约束对话：现场数据未接入时明确拒绝生成运营实绩，同时保留自然语言研判、证据状态和人工确认入口。}

GET  /api/dashboard                 聚合概况、能耗、预警、任务 / aggregate operations, energy, alerts, tasks
GET  /api/runtime/status            模式、来源、质量与边界 / mode, provenance, quality, boundary
GET  /api/runtime/snapshot          船舶、泊位、设备等对象 / vessels, berths, equipment, yard, gates
GET  /api/operations/overview       运营概况 / operations overview
GET  /api/energy?range=today        能碳趋势 / energy-carbon trend (today / 7d / 30d)
GET  /api/alerts                    预警筛选 / alerts filtered by level/status
GET  /api/tasks/templates           任务模板 / executable task templates
POST /api/tasks                     创建沙箱任务 / create sandbox task
POST /api/tasks/{task_id}/next      推进任务步骤 / advance current step
POST /api/reports                   生成结构化报告 / generate structured report
GET  /api/reports/{report_id}       获取报告 / retrieve report
GET  /api/operator/scenarios        岗位问法与安全边界 / frontline prompts and safety limits

生产数据切换说明见 港口运营数据适配器。默认无需配置；接生产网关时设置 XIAOYI_PORT_DATA_MODE=live、XIAOYI_PORT_BASE_URL 和可选只读令牌即可。网关未返回 live_data_verified=true 或版本不是 port-ops.v1 时，服务会失败关闭，绝不把数据冒充生产实绩。

See the port-operations data adapter for live-data switching. No configuration is needed by default. For a production gateway, set XIAOYI_PORT_DATA_MODE=live, XIAOYI_PORT_BASE_URL, and an optional read-only token. The service fails closed unless the gateway returns live_data_verified=true and contract version port-ops.v1; unverified data is never presented as production performance.

运行测试：
Run tests:

python -m pip install -r requirements-dev.lock
pytest -q

本地模式仅用于回环开发。生产模式必须使用签名 JWT、显式主机名和 CORS 来源，否则服务拒绝启动。对话、任务、报告、自动化计划、反馈与审计保存在 SQLite；中断工作在重启后会安全标记为失败或取消，不会无声续跑。

Local mode is for loopback development only. Production requires signed JWTs, explicit hostnames, and CORS origins or the service refuses to start. Conversations, tasks, reports, automation plans, feedback, and audit evidence persist in SQLite. Interrupted work is safely marked failed or cancelled after restart rather than resuming silently.

GET  /health/live               进程存活 / process liveness
GET  /health/ready              深度就绪 / storage, index, RL, model, deployment readiness
GET  /metrics                   Prometheus指标 / metrics; restrict in production
GET  /api/system/info           安全与运维能力 / security and operations capability
GET  /api/models                模型、回退与熔断 / models, fallback, circuit breaker
GET  /api/conversations/{id}    持久对话 / identity-scoped conversation history
POST /api/chat/stream           SSE流式回答 / server-sent-event answer stream

部署、TLS/SSO、密钥、备份、模型数据外发和多实例边界见 部署指南；信任边界见 系统架构；开源审计结论见 真实性与工程化评估。

See the deployment guide for TLS/SSO, secrets, backups, model-data egress, and multi-instance boundaries; architecture for trust boundaries; and open-source readiness for the engineering audit.

完整分类问题库见：小懿可询问问题库。其中已区分当前运营沙箱、知识库问答和生产系统接入后实时问题。

See the categorized question library, which separates current sandbox questions, knowledge-base questions, and questions that require a verified live production connection.

RL板块不再回放其他项目的历史训练曲线。当前实现会在后台真实执行 Q-learning、SARSA、Expected SARSA 和 Double Q-learning，并用 PID 作为控制 理论基线。训练百分比来自已完成 episode 数，模型按算法保存并计算 SHA-256。

仓库同时保留三套数据：UCI Appliances 的 19,735 条原基准、UCI Household Power 聚合得到的 409,887 条五分钟级大规模基准，以及 NOAA Los Angeles–Long Beach 港区 710 个包含 AIS 消息的一分钟实测交通桶。前两套 证明算法规模与回归稳定性，不是港口实绩；AIS中的船舶数、航速、航行状态和 船型为公开观测，服务量、积压、等待和得分为校准沙箱代理。

The RL panel no longer replays historical curves from another project. It runs Q-learning, SARSA, Expected SARSA, and Double Q-learning in a real background worker, with PID as the control-theory baseline. Progress is derived from completed episodes; each algorithm artifact is saved and SHA-256 hashed.

The repository retains the original 19,735-row UCI benchmark, a 409,887-row five-minute UCI benchmark, and 710 observed minute buckets from public NOAA AIS messages in the Los Angeles–Long Beach port area. UCI data demonstrates algorithm scale, not port performance. AIS vessel count, speed, navigation status, and vessel class are measured; service, backlog, waiting, and score are calibrated sandbox proxies. Every dataset is split chronologically into train, validation, and sealed test segments. Training is headless and only post-training evaluation creates trajectories.

_{真实训练配置：三套公开数据、4 RL + PID、回合/时域/随机种子及训练无渲染、测试后回放边界均由后端接收。}

# 仓库已带数据；可从UCI重建 / bundled data; optionally rebuild from pinned UCI source
.venv/bin/python scripts/fetch_public_rl_dataset.py
.venv/bin/python scripts/fetch_large_public_rl_dataset.py
.venv/bin/python scripts/fetch_noaa_port_ais_dataset.py

# 复验固定三数据集×三种子报告 / verify the pinned multi-dataset report
.venv/bin/python scripts/run_rl_dataset_benchmark.py verify

# 运行全部测试 / run all tests
.venv/bin/python -m pytest -q

接口：

GET  /api/rl-lab/health
GET  /api/rl-lab/algorithms
GET  /api/rl-lab/datasets
GET  /api/rl-lab/contracts
GET  /api/rl-lab/evidence
POST /api/rl-lab/advisor
POST /api/rl-lab/runs
GET  /api/rl-lab/runs/{run_id}
POST /api/rl-lab/runs/{run_id}/cancel
POST /api/rl-lab/runs/{run_id}/evaluate

能源环境最少提供 timestamp,load_kw；港口作业环境最少提供 timestamp,vessel_count,anchored_vessels,avg_sog_knots，并设置 XIAOYI_RL_ENVIRONMENT_TYPE=port_operations 与港口 profile。字段名不同只需 配置 JSON 映射，不改训练器。完整说明见港口RL可替换数据契约 和落地执行清单。

The energy environment requires timestamp,load_kw. The port-operations environment requires timestamp,vessel_count,anchored_vessels,avg_sog_knots, XIAOYI_RL_ENVIRONMENT_TYPE=port_operations, and a port profile. Site columns are mapped through JSON without changing the trainer. See the swappable port RL contract and landing checklist.

_{港口环境契约：11项观测、5项动作、目标函数、硬约束、动作屏蔽和生产写入边界可直接与站点字段逐项核对。}

• 港口由哪些核心业务模块组成？ / What are the core operating domains of a port?
• 集装箱码头从船舶靠泊到离港的流程是什么？ / What is the container-terminal workflow from berthing to departure?
• TOS 系统在港口运营里负责什么？ / What does a TOS manage in port operations?
• 岸电 THDi 超标告警应该先检查什么？ / What should be checked first after a shore-power THDi alarm?
• 台风红色预警下港区要启动哪些安全流程？ / Which safety procedures apply under a red typhoon warning?
• 港口碳排放盘查需要保留哪些证据？ / What evidence must be retained for a port carbon inventory?
• 小懿的系统架构是什么？ / What is Xiaoyi’s system architecture?

项目已预留 TOS、PCS、EMS、EAM、VTS、AIS、气象海洋和国际贸易单一窗口连接器契约，但默认全部离线，未配置任何真实端点或凭据。安全配置、站点联调、写操作门禁、回滚和审计要求见 港口连接器接入手册，环境变量模板见 .env.connectors.example。

The project defines connector contracts for TOS, PCS, EMS, EAM, VTS, AIS, metocean services, and maritime single windows. All are offline by default, with no real endpoint or credential bundled. See the port-connector integration manual for secure configuration, site commissioning, write gates, rollback, and audit requirements, and .env.connectors.example for environment templates.

GET  /api/connectors                              目录与状态 / catalogue and status
GET  /api/connectors/{id}/field-mappings          字段映射 / field-mapping template
POST /api/connectors/{id}/health-check            真实健康探测 / real health probe
POST /api/connectors/{id}/write-preflight         写预检，不下发 / write preflight; no dispatch

POST /api/automation/plans                       自然语言转白名单步骤 / language-to-allowlisted steps
POST /api/automation/plans/{id}/next             回写并推进 / record result and advance
POST /api/automation/plans/{id}/confirm          人工确认或拒绝 / bind human approval or rejection
GET  /api/knowledge/status                       文档、来源与索引哈希 / docs, sources, index hashes
GET  /api/knowledge/catalog                      24类/96主题路线图 / 24-domain, 96-topic roadmap
GET  /api/knowledge/authority-coverage           权威覆盖与缺口 / authority coverage and gaps
POST /api/knowledge/search                       来源与双SHA-256检索 / search with source and dual hashes
GET  /api/knowledge/sources                      来源登记与等级 / source registry and verification level
POST /api/knowledge/intake                       待审资料暂存 / stage material for human review

专业资料全目录与当前覆盖状态见 港航专业知识总目录；知识来源分级、资料审核、发布、索引重建与拒答边界见 知识库治理说明。

一线调度、值班、设备、闸口、堆场和交接班人员的完整使用流程见 一线操作人员系统指南。

See the maritime knowledge catalogue for current coverage and the knowledge-governance guide for source tiers, review, publication, index rebuilds, and refusal boundaries. The frontline operator guide covers dispatchers, duty teams, equipment, gates, yard personnel, and shift handover.

小懿只承担港航知识、RAG、上下文识别、能力路由、结果解释和审计，不复制能碳驾驶舱、数字孪生平台、马六甲沙盘或航行模拟器的业务功能。跨系统能力默认 offline，不访问其他系统；dry-run 只检查契约，显式配置为 live 后也只允许登记的 GET 只读能力。

Xiaoyi owns maritime knowledge, RAG, context resolution, capability routing, result explanation, and audit. It does not duplicate the CarbonOps cockpit, digital-twin platform, Malacca sandbox, or sailing simulator. Cross-system capabilities default to offline; dry-run verifies only contracts, and even explicit live mode permits registered read-only GET capabilities only.

_{联动与评测闭环：7/7代表小懿平台能力已登记；标题同步给出4个系统与11项只读能力边界，并展示知识快照、检索对照、证据策略和持久审计。}

GET  /api/hub/systems                            四系统目录 / four-system catalogue
GET  /api/hub/capabilities                       只读能力 / read-only capabilities
POST /api/hub/capabilities/{id}/invoke           隔离预览或授权只读 / isolated preview or approved read
POST /api/context/resolve                        统一业务上下文 / unified business context
POST /api/evidence/fuse                          证据融合 / knowledge-system-simulation fusion
POST /api/orchestrator/run                       跨系统编排 / natural-language orchestration
GET  /api/governance/audit                       SQLite持久审计 / persistent audit
POST /api/evaluation/run                         固定RAG回归 / pinned RAG regression
POST /api/evaluation/feedback                    人工反馈 / human feedback
POST /api/evaluation/feedback/{id}/review        审核入待审区 / reviewed intake staging

网页左侧点击“智能联动中心”即可查看 7/7 完成度、系统能力注册表、跨系统安全预览、RAG 评测和反馈闭环。完整验证步骤见 智能联动中心说明。

Open “智能联动中心 / Intelligence Hub” from the left navigation to inspect 7/7 capability status, the system registry, safe cross-system previews, RAG evaluation, and the feedback loop. See the intelligence-hub guide for verification.

左侧“四系统联动”会实际校验并连接港口数字孪生、能碳驾驶舱、马六甲沙盘和航行模拟器，而不是只提供启动链接。网页系统采用后端健康接口与前端页面双重身份校验；缺少的组件按固定端口契约补启动，已验证在线的服务直接复用。每次业务调用返回 trace ID、耗时和 SHA-256 回执。航行模拟器采用单次消费桥接，执行完成后清除控制指令并恢复原航线，避免重复移动。完整流程见 四系统联动中心说明。

“Four-system linkage” validates and connects the port digital twin, energy-carbon cockpit, Malacca sandbox, and sailing simulator. It performs verified local calls, returns trace/duration/SHA-256 receipts, reuses identity-checked services, and fails closed on port conflicts.

仓库发布六套版本化固定基准，合计 260 题。v1 包含 40 题检索和 20 题证据 策略，其中 35 题构成固定发布测试分区；v2 新增 60 题困难集，分别检验多轮 上下文改写、复杂问题分治和对抗性证据边界；v3 新增 30 题决策保障集， 检验决策就绪度、证据冲突、新鲜度、风险级别和升级动作；v4 新增 20 题 主张—证据对齐集，检验引用角色、词面对齐及数字、日期、量值完整性；v5 新增 60 题日常运营集，覆盖能源、船舶泊位、堆场闸口、班组协同、单证和设备，并 单列 3 个模糊问法/实时边界用例；v6 建立 15 个业务域 × 26 种问题形式的 390 个意图单元与 780 条正式/日常矩阵问法，并抽取 30 题分域固定集，另列 5 题澄清、实时与官方全文边界。当前知识快照为 129 份文档、 882 个分块和 68 份官方核验来源。v1 测试分区的 24 个检索题上，Hybrid Sparse 的 Hit@1/3/5 为 100%/100%/100%，BM25-only 为 95.83%/95.83%/95.83%；MRR 为 1.0000 对 0.9583（+4.17 个百分点）。 11 个策略测试全部通过。v2 的多轮、复杂拆解与对抗边界分别为 20/20、20/20、20/20；v3 的真实问答链路与合成保障分别为 14/14、16/16； v4 的引用、对齐与数值门禁分别为 6/6、6/6、8/8；v5 六类日常问答均为 10/10，边界用例 3/3；v6 分域固定题 30/30，边界用例 5/5。

The repository publishes six versioned fixed benchmarks totaling 260 cases. v1 contains 40 retrieval and 20 evidence-policy cases, with 35 in the pinned release-test partition. v2 adds 60 challenge cases for dialogue rewriting, compound-question decomposition, and adversarial evidence boundaries. v3 adds 30 decision-assurance cases for readiness, conflict, freshness, risk, and escalation actions. v4 adds 20 cases for citation roles, deterministic lexical claim–evidence alignment, and numeric/date/value integrity. v5 adds 60 daily-operations cases across energy, vessel/berth, yard/gate, shift coordination, documents, and equipment, plus three separately counted clarification/live-data boundary checks. v6 defines 390 intent cells and 780 formal/daily matrix prompts across 15 domains and 26 question forms, then tests 30 pinned domain cases plus five separately counted safety-boundary cases. The current snapshot contains 129 documents, 882 chunks, and 68 officially verified sources. On the 24 v1 retrieval tests, Hybrid Sparse reaches Hit@1/3/5 of 100%/100%/100%, versus 95.83%/95.83%/95.83% for BM25-only; MRR is 1.0000 versus 0.9583 (+4.17 percentage points). All 11 v1 policy tests pass. The three v2 categories each pass 20/20; v3 passes 14/14 query-chain and 16/16 assurance cases; v4 passes 6/6 citation, 6/6 alignment, and 8/8 numeric-integrity cases; all six v5 daily categories pass 10/10 and the boundary checks pass 3/3; v6 passes 30/30 domain cases and 5/5 boundary cases.

# 快速哈希核验 / verify report-bound data, index, and policy-code hashes
.venv/bin/python scripts/run_rag_benchmark.py verify

# 深度复算60题 / rerun all 60 cases; may take minutes on one CPU core
.venv/bin/python scripts/run_rag_benchmark.py verify --deep

# 重新生成报告 / regenerate the report
.venv/bin/python scripts/run_rag_benchmark.py run

# 核验/复跑 v2 助手困难集 / verify or rerun the v2 assistant challenge set
.venv/bin/python scripts/run_assistant_benchmark.py verify
.venv/bin/python scripts/run_assistant_benchmark.py run

# 核验/复跑 v3 决策保障 / verify or rerun v3 decision assurance
.venv/bin/python scripts/run_decision_benchmark.py verify
.venv/bin/python scripts/run_decision_benchmark.py run

# 核验/复跑 v4 主张—证据对齐 / verify or rerun v4 claim-evidence alignment
.venv/bin/python scripts/run_alignment_benchmark.py verify
.venv/bin/python scripts/run_alignment_benchmark.py run

# 核验/复跑 v5 港口日常问答 / verify or rerun the v5 daily-operations set
.venv/bin/python scripts/run_daily_operations_benchmark.py verify
.venv/bin/python scripts/run_daily_operations_benchmark.py run

# 生成问题全集并核验/复跑 v6 / build the universe and verify or rerun v6
.venv/bin/python scripts/build_port_question_universe.py
.venv/bin/python scripts/run_question_universe_benchmark.py verify
.venv/bin/python scripts/run_question_universe_benchmark.py run

固定题集、JSON证据报告和 可读报告，以及 v2 困难集报告和 v3 决策保障报告、 v4 主张—证据对齐报告和 v5 日常运营问答报告、 v6 港口问题全集报告和 正式/日常问题清单 随仓库发布。六套题集都由项目维护、在开发中可见，不是“从未查看的独立留出集”或第三方 用户研究。上述数字是仓库基准上的检索与安全指标，不是港口生产收益、全球 知识覆盖率、法律意见或线上 SLA。

The fixed questions, JSON evidence report, and readable report ship with the repository. The partition is a v1 release-acceptance set that has exposed defects during engineering fixes; it is not a never-inspected independent holdout, and it is maintained by this project rather than a third-party user study. These are repository retrieval and safety metrics—not port-production benefits, global knowledge coverage, legal advice, or an online SLA.

简历可用声明与禁用表述见 简历指标口径。
See resume claim semantics for permitted and prohibited wording.