NarcoScope — illicit-economy evidence atlas

io.github.beepboop2025由 GitHub 账号 beepboop2025 发布v1.6.0更新于 Oct 24, 2026

Read bounded illicit-economy evidence with provenance, privacy, availability, and federation intact.

发布方自己的服务:Red Atlas Api

已验证Streamable HTTP可网页运行金融知识与记忆数据与分析

概览

AI 生成的概览

只读的远程 MCP 服务器,开放 NarcoScope 的非法经济证据图集:毒品价格、缴获、公开制裁名单、市场指标与带引用的分析。

可以做什么
  • 查询毒品、武器与非正规经济指标的市场目录与分页观测,含单位、来源与明确的空缺值
  • 读取国家-年份的有组织犯罪与枪支溯源图集记录,附来源元数据与说明
  • 查询经批准的公开列名身份字段与有界汇总指标
  • 浏览证据新闻室:列出分析、获取报道的元数据、机器简报或带引用的档案
  • 获取 Palimpsest 的中国、一带一路、走廊与关联研究背景数据,附来源与披露限制
  • 列出该服务的能力、受众与安全边界
适用场景
适合需要带来源的汇总统计(毒品市场、缴获、过量死亡、污水监测、公开列名)或需要带引用与明确证据边界的分析时使用。适用于研究、教育与新闻背景工作,不适合行动层面或具体地点层面的问题。
运行要求
无需安装:它是位于 www.narcoscope.com/mcp 的远程 streamable HTTP 端点。未声明需要身份验证、环境变量或请求头。需要网络访问,返回结果有大小限制并分页。
安装前请注意
服务为只读,但会返回公开列名名单中的具名实体;列名是行政措施而非有罪认定,名单会变动,使用前应核对现行来源。部分数据为示意性或不可用,并已标注。不要把地理或时间上的重合当作行为者关系或因果关系的证据。

安装

在 SourceWeft 中

  1. 打开 控制台中的 NarcoScope — illicit-economy evidence atlas,将其添加到工作区。
  2. 为需要使用其工具的对话启用该服务。

Web executable,通过 Streamable HTTP。 远程服务在工作区中配置后即可从网页运行时运行。

交给智能体安装

把这段提示词粘贴到 Claude Code、Codex、Cursor 等智能体,或 SourceWeft 对话中。智能体会读取该服务器的安装指南,先向你展示发布者、端点和登录方式,经你同意后再把它添加到你的客户端。密钥不会写进任何命令。

阅读 https://sourceweft.com/mcp/io.github.beepboop2025%2Fnarcoscope/install.md,并按其中说明把这个 MCP 服务器接入我的客户端。修改任何配置之前,先告诉我它的发布者、端点和登录方式,等我确认。不要把 API 密钥写进命令或我会提交的文件;问我密钥应从哪里获取。

查看智能体所依据的安装指南

其他 MCP 客户端

把它添加到你客户端的 mcpServers 配置中。

{
  "mcpServers": {
    "narcoscope-illicit-economy-evidence-atla": {
      "type": "http",
      "url": "https://www.narcoscope.com/mcp"
    }
  }
}

README

[tests] [coverage]

🌍 NarcoScope

Live: narcoscope.com · evidence newsroom: narcoscope.com/#newsroom

Deployment status verified 2026-09-08: both narcoscope.com and www.narcoscope.com serve the Railway runtime. The Hetzner Fleet publisher deploys validated source from GitHub and verifies the exact commit on the provider and public origins. Registry metadata and agents use the apex MCP endpoint.

An educational, public-good data explorer that makes the world's drug-trade data legible. UNODC, INCB, and EUDA already publish street (retail) prices, precursor-chemical prices, and trafficking-flow/seizure data — but it's buried in dense PDFs and CSVs most people can't read. This app is a translation layer on top of that public data: clean charts, maps, and plain-English explanations.

Mission: democratize hard-to-read official drug data. Not a new data source — a way to understand the existing one.

Agent and API access

  • RFC 9727 API Catalog: https://narcoscope.com/.well-known/api-catalog
  • OpenAPI: https://narcoscope.com/openapi.json
  • MCP manifest: https://narcoscope.com/server.json
  • Streamable HTTP MCP: https://www.narcoscope.com/mcp
  • Machine-readable product context: https://narcoscope.com/.well-known/ai-catalog.json
  • Agent-readable guide: https://narcoscope.com/llms.txt
json
{  "mcpServers": {    "narcoscope": { "url": "https://www.narcoscope.com/mcp" }  }}

The MCP service is read-only and returns JSON Schema-backed structuredContent for every tool. Version 1.3 supports both the stateless MCP 2026-07-28 discovery lane and the legacy 2025-06-18 initialization lane. It preserves official, illustrative, stale, unavailable, and restricted states rather than turning missing evidence into a score.

The staged/live/Registry boundary and mandatory publication sequence are recorded in docs/MCP-1.3.0-RELEASE-GATE.md.

What it shows

  • Street Prices — retail price trends by country, with a purity-adjusted view and an affordability lens (price expressed as days of average local income).
  • Precursor Flows & Prices — trafficking corridors and precursor-chemical prices, with source hubs (notably China) highlighted.
  • Flow Map — an Equal-Earth world map of corridor arcs, animated over time.
  • Triangulation — reads seizures against modalities that interdiction agencies do not collect. Seizure volume moves with enforcement capacity at least as much as with trafficking volume, and nothing inside the seizure data separates the two; overdose mortality (vital-statistics registrars) and wastewater load (environmental chemists) can. Where the modalities agree the reading is credible; where they diverge, the divergence is the finding — seizures up with consumption flat reads as an enforcement effect, consumption up with seizures flat reads as expansion interdiction missed. Every verdict reports how many independent modalities backed it and whether it survives dropping any one of them. Validation case: Canadian cannabis 2019→2023 reads seizures −94% against measured wastewater consumption +69%. Canada legalised cannabis in 2018, so the seizure collapse is enforcement stopping, not demand stopping — and the tool says so rather than reporting a shrinking market.
  • Designations — the ~2,600 entities on the OFAC SDN list under the narcotics and transnational-crime authorities, searchable by name or by OFAC-published alias, plus a jurisdiction network built from the countries Treasury records each entity in. Betweenness and articulation-point analysis show which jurisdictions hold the designated networks together.
  • Evidence Newsroom — a deterministic, offline-capable publication pipeline with sentence-level citations and independent-source gates. Its first bounded analysis keeps lawful trade, selected official incidents and US harm data in separate evidence lanes, and states what cannot be joined or causally attributed from the public record.
  • Palimpsest BRI context — a pinned, bounded source-readiness and national economic context lane for CPEC, Gwadar, CMEC, Kyaukpyu and Balochistan. It preserves official, independent and modeled claim classes plus unavailable WDI coverage, but cannot enter drug-market, actor, route, guilt, political or causal inference.
  • Myanmar Focus — province-level (Golden Triangle) detail: production regions, civil-war conflict pressure, China/third-country precursor inflows, cross-border corridor towns, and seized volumes. The intelligence layer fuses multi-source evidence into per-region risk/confidence scores, flags cross-source disagreement, weights sources by reliability tier, computes a year-over-year risk trajectory (rising/falling/stable) so analysts see momentum, flags a geographic spillover watch when a calm region borders one that has already crossed the high-risk threshold, flags evidence staleness (current/aging/stale) when a region's freshest record predates the reporting year, discounting confidence accordingly, and scores precursor-corridor concentration with a Herfindahl-Hirschman Index (diversified/moderate/concentrated) to flag single-source supply dependency — both a fragility signal and an interdiction priority. Risk profiles and the evidence-graph ledger can be exported as CSV directly from the briefing for offline analyst review.

Every view carries an auto-generated "In plain English" sentence and hover tooltips that explain each figure in human terms.

Screenshots

Street Prices now ships official data: UNODC World Drug Report 2025, Statistical Annex 8.1 (retail per-gram prices + purities, 2019–2023, 208 records across 69 countries), with World Bank GDP-per-capita (2024) powering the affordability lens. Flow-map, precursor and Myanmar figures remain illustrative pending ingestion — the in-app badge states exactly which is which.

A dark, motion-led interface: a WebGL globe traces precursor corridors out of their source hubs (coral) toward transit and destination nodes (cyan), headings reveal letter-by-letter, and sections spring in as you scroll. The immersive layer is fully gated behind prefers-reduced-motion and falls back to a lightweight 2D canvas on mobile / WebGL-less devices.

[NarcoScope — WebGL hero globe]

Street Prices — price trends + affordability lens, with a plain-English summary:

[Street Prices]

Flow Map — Equal-Earth world map of precursor corridors, animated by year:

[Flow Map]

Myanmar Focus — province-level Golden Triangle detail:

[Myanmar Focus]

Ethical scope (please read)

This tool reports aggregate, published statistics — country-level, annual, and (for focus regions) province-level — strictly for awareness, education, and research. By design it does not provide point-of-sale, real-time, sub-street, or navigable location data, and the precursor layer stores logistics only (what, how much, where, control status) with no chemistry, synthesis routes, or yields. It is not, and must not be used as, a guide to obtaining any substance.

Private ScamShield signal

The Hetzner analyst environment can ingest ScamShield's privacy-minimized Telegram aggregate for private review. This does not enter the website or its Git-backed datasets. A strict schema firewall rejects raw messages, exact IOCs, source identifiers, universal-coverage claims, public-eligibility claims, and anything that drops the human-review requirement. The importer has no network access and retains only aggregate counts, a source-file hash, hourly snapshots, and an append-only receipt ledger under /var/lib/narcoscope-analyst/.

See deploy/private-import/README.md for the trust boundary and production service.

Named entities

The Designations tab names people and companies, which every other layer avoids. The line it holds:

  • A designation is a published act of a government — an entity placed on a list under a stated legal authority on a stated date. NarcoScope reports what the government did. It does not characterise what the entity did, carries no free-text allegation field, and a designation is not an adjudication of guilt. OFAC delists; check the live list before relying on any row.
  • Addresses, passport and national-ID numbers and dates of birth are present in the upstream OFAC file and are deliberately not extracted.
  • Journalism and crowd-sourced trackers are leads, not records. C4ADS, InSight Crime, EIA, GASO and the rest are registered in the source registry and read by the governed scraper into an analyst work queue. A named entity reaches a bundled dataset only if it also appears on an official designation list. The rule is written into scripts/scrape/myanmar-sources.json as verification_rule.
  • The designation graph has no entity-to-entity edges, because OFAC publishes none. Inventing them and running centrality over the result would produce confident rankings of private individuals from a graph this tool made up. These methods are outside the published aggregate-data scope.

Data provenance

Most bundled figures are now official extracts, regenerated by the pipeline from the sources below. What is still illustrative is named explicitly: the Myanmar region-level flow volumes and the precursor price series (INCB publishes no precursor prices). The in-app badge states which is which, and every generated dataset carries a provenance header naming its source, retrieval date and extraction rules.

The full registry spans automated datasets, manual-step datasets, and investigative reports that feed the analyst work queue rather than the app — in scripts/pipeline/sources.json. Each entry records its licence, cadence, automation tier, and what it feeds. Two entries are registered specifically so the reasoning is not relitigated: OpenSanctions publishes no licence on its bulk artifacts and is therefore a lookup pointer rather than a bundled dataset, and OCCRP Aleph's per-collection licences forbid redistribution.

Wastewater comes from Canada, not Europe. EUDA returns HTTP 403 to non-browser clients (on the current URL, the legacy one, and the copy linked from data.europa.eu) and ACIC ships PDFs, so neither of the famous programmes can be automated. Statistics Canada publishes the same measurement, at the same grain, in the same SCORE unit, through a keyless API under an open licence — so that is the bundled default, and Canada is the second fully-triangulated country. European and Australian coverage still needs a verified export through the CSV panel.

Load real data through the "Load official data (CSV)" panel in the footer; each file is parsed by src/lib/ingest.ts and bad rows are reported, not silently dropped. See src/lib/ingest-config-reference.md for the column mapping.

Myanmar conflict and precursor-flow source triage can be prepared with the Palimpsest-style governed scraper:

bash
npm run scrape:myanmar -- --out docs/sources/myanmar-observations.csv --pretty

That output is an analyst work queue with excerpts and content fingerprints, not direct app data; verify and code rows into the Myanmar civil-war / precursor CSV schemas before loading them.

A derived ontology of every entity and relation type actually present in the data is regenerated by npm run ontology into docs/ontology/. It is induced by observation over the bundled datasets rather than generated by a language model: the corpus is already typed and provenance-tagged, so the schema can be read off it directly, which removes the hallucination risk entirely. It is a draft for review and is never auto-applied to src/types.ts.

The new Enterprise Intel tab adds an event/entity evidence graph, regional risk scores, confidence/source-diversity indicators, and an evidence ledger for analyst review. See docs/ENTERPRISE_HARDENING.md for the paper-backed design.

The Evidence Newsroom tab reads a checked-in dossier and receipt generated without network access or model calls. Publication requires sentence and visual citations, active upstream-source independence for synthesis, a countercase, limitations and causal/culpability safety checks. See docs/EVIDENCE_NEWSROOM.md.

Evidence newsroom publication contract

The newsroom is visibly labelled automated evidence analysis, with humanReviewStatus: not_recorded. No generative model participates in its build, no expert or affected-person testimony is included, and it never simulates those human voices. Its initial article makes no named allegation, records right to reply as not_required, and publishes a correction/update history with stable revision and content hashes.

An official record used by the build can support an attributed observation. An analytical or methodological synthesis requires at least two independent, actively used official upstream groups; merely registering an available, capability-only or unavailable source contributes zero corroboration. When the inputs lack a lawful-trade denominator, record-level join, adjudicated outcome or defensible country allocation, the newsroom abstains. It does not turn origin labels, administrative designations or a separate mortality trend into guilt or causal attribution.

Publication surfaces are:

  • app route: /#newsroom;
  • standalone HTML: /news/china-linked-precursor-incidents-official-record.html;
  • machine brief: /news/china-linked-precursor-incidents-official-record.machine-brief.json;
  • cited dossier: /news/china-linked-precursor-incidents-official-record.dossier.json;
  • JSON Feed: /news/feed.json; and
  • Atom feed: /news/feed.xml.

Run npm run news:build to regenerate the offline bundle and npm run news:check to verify every checked-in byte. The production build runs the stale-artifact check before TypeScript and Vite.

Tech

React 18 · Vite 8 · TypeScript · Recharts · react-simple-maps (world-atlas bundled locally). The interface layer adds Three.js / React Three Fiber (a lazy-loaded hero globe with bloom post-processing — kept out of the initial bundle), @react-spring/web (physics-based letter/section reveals and animated counters), and Lenis (global smooth scroll) — all behind a prefers-reduced-motion guard. Runtime data store (src/lib/dataStore.ts) swaps sample → real data on load.

Develop

bash
npm installnpm run dev        # local dev servernpm run scrape:myanmar -- --prettynpm run ontology   # regenerate docs/ontology/draft-ontology.{json,md}npm run news:build # regenerate the deterministic evidence newsroomnpm run news:check # verify that checked-in newsroom artifacts are currentnpm run bridge:palimpsest-bri:check # verify the pinned BRI artifact + hashnpm run build      # type-check (tsc) + production build → dist/npm run preview    # preview the buildnpm run typecheck  # tsc --noEmitnpm test           # run unit tests (Vitest)

Deploy (Vercel)

The repo is Vercel-ready (vercel.json pins the Vite framework). Either:

  • Dashboard: import the Git repo at vercel.com — zero config, auto-detected.
  • CLI: npx vercel (preview) / npx vercel --prod (production).

Only api/mcp.mjs and api/v1.mjs are function entrypoints. API helpers and tests use underscore-prefixed paths so Vercel does not publish them as functions. npm run build checks this boundary against vercel.json before building.

Data pipeline

npm run data:refresh fetches the automatable open sources (UNODC WDR annexes, World Bank GDP, CDC VSRR mortality, OFAC SDN designations), regenerates the bundled datasets, and validates them against the test suite. Individual refreshes: npm run data:overdose, npm run data:designations. A quarterly GitHub Action does the same and opens a PR when the data changes. The full source registry (including the manual and API-key sources not yet wired in) lives in scripts/pipeline/sources.json; the playbook is docs/DATA_PIPELINE.md.

Current coverage limits

Price-per-pure-gram calculations require known purity. Street-price and qualified precursor-flow records retain their cited UNODC and INCB sources; a citable precursor-price series and the Myanmar dataset remain unavailable. Wastewater coverage is Canada. The Designations tab uses OFAC as its sole designating authority. These limits are not estimates of missing observations.

License

MIT covers the application code. Data retains its source-specific license and attribution requirements. UNODC World Drug Report values need written UNODC permission for commercial use, so they are withheld from the hosted product, its API and exports, and from this repository's tracked files; their source metadata remains. See docs/publication-rights.md.

Connected China and regional research

BRI, CPEC/Gwadar, Balochistan and Myanmar views include a shared Palimpsest research desk. Switch the research area to China for detailed NBS findings. Annual national debt, reserve, current-account and inflation comparisons keep reference years visible; they do not measure project or district outcomes.

The REST resource is /api/v1/connected-research; MCP exposes get_connected_research. Both serve the validated public/data/palimpsest-connected-research-v1.json snapshot with its SHA-256 sidecar. npm run bridge:connected:sync refreshes from the fixed Palimpsest URL; npm run bridge:connected:check verifies the packaged copy offline. The Hetzner wire publisher refreshes it during scheduled publication. Failed acquisition preserves the dated copy.

Snapshots can also include the China evidence observatory: deduplicated economic histories, EU-reported trade, official-document changes and historical methodology cases. China shows the complete findings; regional views show the separately attributed trade context. The optional observatory summary adds a fifth named input hash and a matching source clock. Its hash identifies the full Palimpsest evidence snapshot; the packaged-file SHA-256 also binds the summary served by REST and MCP. Older four-input snapshots remain readable.

The bridge admits four fixed public dataset routes and reviewed coverage counts. SAFE numeric records remain private; its card and download contain acquisition metadata only. Document absence, methodological revisions and trade disagreement do not establish concealment, culpability or causality. Official-announcement dates remain separate from collection dates.

NBS statistical-data attribution, World Bank CC BY 4.0 attribution and publisher metadata-only terms remain source-specific. Shared geography or timing does not establish actor relationships, culpability or causality. Detailed collector/analysis methods are in Palimpsest’s docs/CONNECTED-ECONOMIC-RESEARCH.md.

Global market research workspace

The shared dark research interface connects NarcoScope and Palimpsest. The data library queries source-native country, city, category, breakdown and period observations; it includes linked choropleth maps, comparison histories, geographic bars, source-record inspection and attributed CSV downloads. Existing seizure, organized-crime, regional, overdose and wastewater views remain accessible.

The new library contains 129,217 source cells (74,373 numerical observations and 54,844 explicitly unavailable cells) across 58 indicators. Sources include the WDR2026 drug annexes, monthly Statistics Canada wastewater measurements, World Bank informality estimates, and source-licensed WDI conventional-arms and firm-survey series. Armament transfer TIV is not a monetary value or a classification of illegal trade. Informality models do not measure the entire black economy. Each source retains its reference dates and restrictions.

npm run markets:check verifies the public source-bound catalog and contracts; npm run workspace:build produces the portable explorer used by Palimpsest. /api/v1/markets discovers coverage and valid filter combinations; /api/v1/market-observations returns paginated measurements. The MCP equivalents are get_market_catalog and query_market_observations. The source datasets remain external to the initial JavaScript bundle.

The MCP catalog uses narcoscope.global-market-catalog-page.v1: page and limit select up to four indicators in stable dataset/indicator order. Dataset coverage, hashes and total indicator/source counts remain complete; source details and reuse restrictions accompany the indicators on that page. Each indicator's geography, category, subgroup and observed-selection lists have separate dimensionPagination totals and next-page markers. Set dataset and indicator, then follow dimensionPage with the same dimensionLimit (default 25, maximum 50) to enumerate every value, including explicit blanks. Retain dataset hashes across pages and restart enumeration if they change. The full REST catalog and browser explorer retain their existing contract. MCP measures both protocol copies against a 1 MiB response budget and returns a bounded error with smaller-page guidance if future metadata exceeds it.

Connected research

The Seiche research_network tool and /api/v2/research-network endpoint connect Palimpsest's complete source catalog with separately completed funding context and explicit research steps into LiquiLens, Undertow and NarcoScope. Use topic, offset and limit; follow next_offset for remaining datasets. The human desk is https://seiche.info/#RESEARCH. Telegram readers can use /research china, /research regions or /research liquidity.

Catalog rows are references, not permission to redistribute source values or change a score. Source clocks, rights and missingness remain attached. The retrieved catalog hash is a byte identity, not a producer attestation.

来源:README.md,提交 75eb7de

工具

14

research_network

Connect global evidence with funding and institution research

Read Palimpsest’s complete source catalog by topic alongside Seiche funding context, with next steps into LiquiLens filings and Undertow liquidity. Source dates, rights and missingness stay attached; no actor, route, criminal or causal inference is implied.

输入参数limittopicoffset

list_capabilities

Discover NarcoScope

List the evidence explorer, newsroom, API, feeds, MCP tools, audiences, and safety boundaries.

无需输入参数。

get_overview

Read the official-data overview

Return bounded headline aggregates across official prices, seizures, overdose mortality, wastewater, and public designations.

无需输入参数。

get_atlas

Query the illicit-economy atlas

Return country-year organized-crime and firearms-tracing observations with source metadata, native caveats, explicit null fields, and stable cursor pagination. No records identify people or operational locations.

输入参数iso3yearlimitcursordomain

get_entities

Query privacy-minimized OFAC designation identities

Return approved public designation identity fields only. A designation is an administrative action, not an adjudication or proof of guilt; aliases, addresses, dates of birth, identity documents, and allegation text are withheld.

输入参数limitquerycursorcountryprogramentity_type

get_federation

Read Seiche and Palimpsest as separate evidence lanes

Return bounded public money-market, capital-market, BRI, or publication-rights context without cross-lane scoring, composites, causal inference, or culpability inference. Upstream unavailability stays typed.

输入参数lane

get_newsroom

Read the evidence newsroom

List current deterministic analyses with citations, correction metadata, publication gates, and explicit evidence limits.

输入参数limit

get_story

Read a newsroom story artifact

Fetch one story as metadata, a machine brief, or the cited dossier. Use the machine brief first for agent workflows.

输入参数slug必填artifact

get_palimpsest_bridge

Read the Palimpsest bridge

Return the official-only China aggregate shared with the Intelligence Commons, including disclosure and causal boundaries.

无需输入参数。

get_palimpsest_corridors

Read the Palimpsest corridor overlay

Return official country-level China, Pakistan, and Myanmar aggregates with missing-data, provenance, disclosure, and geography-and-time-only join rules.

无需输入参数。

get_palimpsest_bri_context

Read bounded Palimpsest Belt and Road context

Return pinned source readiness and national WDI coverage for CPEC, Gwadar, CMEC, Kyaukpyu, and Balochistan as a parallel context lane. Cross-lane drug, actor, route, guilt, political, project, causal, tactical, and navigable inference or use is prohibited.

无需输入参数。

get_connected_research

Read China and regional economic research

Return source-linked China economic findings, CPEC, Balochistan, BRI and Myanmar research questions, reporting metadata, and dated national indicators. Optional hash-bound observatory summaries add economic histories, EU-reported trade and official methodology cases. SAFE quantities remain private. Country and theme context only; gaps do not establish concealment, actor relationships or causality.

无需输入参数。

get_market_catalog

Discover granular global market evidence

Page drugs, arms and informal-economy indicators (page/limit, maximum 4), with native units, full coverage totals and licenses for sources on the page. Country/category/subgroup/selection lists have independent dimensionPagination; follow dimensionPage with the same dataset/indicator filters (dimensionLimit maximum 50). Retain dataset hashes across pages. Legacy atlas and prices remain distinct resources.

输入参数pagelimitdatasetindicatordimensionPagedimensionLimit

query_market_observations

Query granular market observations

Filter published observations by dataset, indicator, geography, category, subgroup and inclusive reference period with bounded stable pagination (default 200, maximum 500). Retain dataset hashes across pages. Units, source locators, estimates and unavailable values remain explicit. No market-size aggregation, unit conversion or causal inference.

输入参数togeofrompagelimitdatasetcategorysubgroupindicator

版本历史

1
  1. v1.6.0最新Sep 16, 2026

评分

已满足适用检查项的 100%

该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。

  • 近 30 天内可连通必需25 分
  • 已声明认证方式必需10 分
  • 已列出工具必需15 分
  • 描述必需15 分
  • 图标10 分
  • 文档15 分
  • 许可证10 分

其他 Atlas 集成