config.yaml

# Parallel Frontier — single config (see SPEC.md)
# Everything lives here: sources, keywords, recipients, arXiv net + roster.

# ─────────────────────────────────────────────────────────────────────────────
# arXiv: broad category net → keyword gate → pinned lab/org roster
# ─────────────────────────────────────────────────────────────────────────────
arxiv:
  # 1) The net: categories queried over a rolling 24h window (~100–300/day).
  categories:
    - cs.AI
    - cs.CL
    - cs.LG
    - cs.CV
    - stat.ML
  window_hours: 36   # fetch margin; arXiv skips weekends. Digest still shows last-24h-new (dedup handles overlap)

  # 2) Roster-FIRST gate. The primary signal is the roster (tracked labs, below).
  #    A netted paper reaches the Research digest only if it is a roster hit OR its
  #    TITLE matches this tight topic allowlist. Everything else → archive.
  #    Keep this list SHORT and high-signal: it is the ONLY non-roster path in, so
  #    broad terms (LLM/agent/reasoning) do not belong here — they match nearly
  #    every cs.AI/cs.CL paper. Add sparingly, prune often.
  topic_allowlist:
    - technical report      # "X Technical Report" = a major model release
    - system card
    - model card
    - frontier model
    - 技术报告
    # NB: "foundation model" / "world model" were removed — too broad, they catch
    # plant/lunar/time-series/robotics papers. Keep this list release-signal only.

  # 3) Lab/org roster: ALWAYS pinned to the top of Research and shown in full,
  #    regardless of the keyword gate. A paper is a "roster hit" if it matches a
  #    roster author name OR an alias (model-family / lab name) in
  #    title / abstract / comments.
  #
  #    NOTE: `authors` are romanized surface names and collide easily (esp.
  #    common Chinese names). They only PIN (surface), never exclude, so a false
  #    positive costs one extra line — but verify/prune these as you soak.
  #    `aliases` (model & lab names) are the robust matchers.
  roster:
    - name: DeepSeek (深度求索)
      authors: []                     # team rarely lists stable individual authors; rely on aliases
      aliases: [DeepSeek, DeepSeek-V2, DeepSeek-V3, DeepSeek-R1, DeepSeek-Coder, DeepSeekMath, DeepSeek-VL]

    - name: Qwen / Alibaba (通义千问)
      authors: [Junyang Lin, Jingren Zhou, Binyuan Hui]   # pruned common names (An Yang/Bowen Yu/Chang Zhou collide)
      aliases: [Qwen, Qwen2, Qwen2.5, Qwen3, QwQ, Tongyi, 通义千问]

    - name: Zhipu AI / Tsinghua KEG (智谱)
      authors: [Jie Tang, Aohan Zeng, Zhengxiao Du, Yuxiao Dong]   # pruned Ming Ding (common)
      aliases: [GLM, ChatGLM, GLM-4, CogVLM, CogView, CogAgent, Zhipu]

    - name: Moonshot AI (月之暗面)
      authors: []
      aliases: [Kimi, Kimi-K2, Moonshot, 月之暗面]

    - name: Shanghai AI Lab (上海人工智能实验室)
      authors: [Jifeng Dai, Conghui He]   # pruned Kai Chen/Yu Qiao/Wenhai Wang (common)
      aliases: [InternLM, InternVL, Intern-, InternImage, OpenGVLab, 书生]

    - name: ByteDance Seed (字节豆包)
      authors: []
      aliases: [Doubao, ByteDance Seed, Seed-, BAGEL, 豆包]

    - name: 01.AI (零一万物)
      authors: [Kai-Fu Lee]
      aliases: [Yi-, Yi-34B, Yi-Lightning, 01.AI, 零一万物]

    - name: Baichuan (百川智能)
      authors: []
      aliases: [Baichuan, Baichuan2, Baichuan-, 百川]

    - name: MiniMax
      authors: []
      aliases: [MiniMax, MiniMax-01, ABAB]

    - name: StepFun (阶跃星辰)
      authors: []
      aliases: [StepFun, Step-1, Step-2, 阶跃星辰]

    - name: Tencent Hunyuan (腾讯混元)
      authors: []
      aliases: [Hunyuan, 混元]

    - name: Baidu ERNIE (文心)
      authors: []
      aliases: [ERNIE, Wenxin, 文心]

    - name: Tsinghua / OpenBMB
      authors: [Zhiyuan Liu, Maosong Sun]
      aliases: [MiniCPM, CPM-, OpenBMB]

    - name: Ant Group (蚂蚁)
      authors: []
      aliases: [Ling-, Bailing, Ant Group, 百灵]

    - name: SenseTime (商汤)
      authors: []
      aliases: [SenseNova, SenseChat, 日日新, 商汤]

    - name: Skywork / Kunlun (昆仑万维)
      authors: []
      aliases: [Skywork, 天工, Kunlun]

# ─────────────────────────────────────────────────────────────────────────────
# Other sources — stubs to be filled per SPEC build order.
# ─────────────────────────────────────────────────────────────────────────────
regulation:
  # WIRED & VERIFIED 2026-06-14 against live pages. Both sites reachable from
  # datacenter IPs (no WAF block). The comment lists are server-rendered.

  # TC260: every row on /portal/suggestion is a draft-standard comment period, so
  # the whole list is alert-worthy. DocketCollector hashes the normalized list
  # region (div.right); a new draft appearing changes the hash -> ALERT.
  # {name, url, selector}.
  tc260_pages:
    - name: "TC260 标准征求意见"
      url: "https://www.tc260.org.cn/portal/suggestion"
      selector: "div.right"

  # CAC: no dedicated comment column — solicitations are posted inline in the
  # 政策法规 feed. ListingCollector parses each row; AI items -> digest:regulation,
  # AI comment-periods -> alert. {name, url, row_selector}.
  cac_pages:
    - name: "CAC 政策法规"
      url: "https://www.cac.gov.cn/wxzw/zcfg/A093703index_1.htm"
      row_selector: "#loadingInfoPage a[href*='/c_']"

  # MIIT: collector ready (collectors/miit.py). Listing pages (政策文件
  # /zwgk/zcwj/, 最新政策 /xwfb/zxzc/) are JS-rendered SPAs (jQuery+layui+require.js)
  # — rows load via AJAX, so a row_selector on static HTML finds nothing. Wiring
  # needs the backing endpoint (browser devtools), deferred per SPEC anti-goal
  # (no source until it fires >=3 useful items). {name, url, row_selector}.
  miit_pages: []

  # AI keyword gate: promotes regulatory items to digest:regulation, and (for CAC
  # listing rows) to alert when combined with a comment-period marker.
  keywords: [人工智能, 算法, 大模型, 生成式, 智能体, 深度合成, 算力]

  # Drafting-unit (起草单位) watchlist — tc260_drafting.py parses each TC260 draft's
  # 编制说明 .docx and ALERTS when one of these orgs is on the drafting committee
  # ("the hardest response signal available"). These are distinctive Chinese org
  # names, so substring matching is safe. Seeded — prune/extend as you like.
  drafting_watchlist:
    # frontier labs / majors
    - 阿里            # Alibaba (阿里巴巴/阿里云/达摩院)
    - 百度            # Baidu
    - 腾讯            # Tencent
    - 华为            # Huawei
    - 蚂蚁            # Ant
    - 字节跳动        # ByteDance
    - 科大讯飞        # iFlytek
    - 商汤            # SenseTime
    - 快手            # Kuaishou
    - 智谱            # Zhipu
    - 月之暗面        # Moonshot
    - 深度求索        # DeepSeek
    - 百川            # Baichuan
    - 稀宇            # MiniMax (稀宇科技)
    - 阶跃            # StepFun (阶跃星辰)
    # key institutes / labs that co-draft
    - 中国信息通信研究院   # CAICT
    - 信通院
    - 中国电子技术标准化研究院  # CESI
    - 上海人工智能实验室   # Shanghai AI Lab
    - 之江实验室          # Zhejiang Lab
    - 智源               # BAAI
    - 鹏城实验室          # Pengcheng Lab
#
rss:                    # rss.py: lab blogs, 机器之心, 量子位, rival newsletters
  # → digest:discourse
  discourse:
    - name: 量子位 (QbitAI)
      url: https://www.qbitai.com/feed
    # 机器之心 (Jiqizhixin): no working public RSS — official /rss & /feed return HTML,
    # and the public rsshub.app instance 403s. Either self-host RSSHub and use the
    # route below, or keep it on the manual-skim list (see SPEC "stays manual").
    # - name: 机器之心 (Jiqizhixin)
    #   url: https://<your-rsshub-host>/jiqizhixin/article

  # → digest:rivals (titles only in digest, so you never duplicate them unknowingly)
  # Substack {name}.substack.com/feed is reliable; the rest marked VERIFY.
  rivals:
    # NOTE: *.substack.com feeds 403 from GitHub Actions datacenter IPs (IP-level
    # block — a browser User-Agent does NOT defeat it; confirmed in CI). These are
    # newsletters you can subscribe to by email directly. Re-enable if you ever run
    # the pipeline from a residential IP or behind a proxy.
    # - name: ChinAI (Jeffrey Ding)
    #   url: https://chinai.substack.com/feed
    # - name: AI Safety in China (Concordia AI)
    #   url: https://aisafetychina.substack.com/feed
    # GeopoliTech: feed is valid XML but EMPTY (0 entries as of 2026-06-14) —
    # dormant or placeholder. Re-enable once it's actually publishing
    # (SPEC anti-goal: no source until it fires >=3 useful items).
    # - name: GeopoliTech
    #   url: https://geopolitechs.substack.com/feed
    # - name: Recode China AI
    #   url: https://recodechinaai.substack.com/feed
    - name: Interconnected (Kevin Xu)
      url: https://interconnected.blog/rss/
    - name: ChinaTalk (Jordan Schneider)
      url: https://www.chinatalk.media/feed
    - name: CSET (Georgetown)
      url: https://cset.georgetown.edu/feed/
    - name: Concordia AI          # via its own site — the aisafetychina Substack is CI-blocked
      url: https://concordia-ai.com/feed/

# 安全内参 (secrss.com): governance/security trade press. No RSS, but the homepage
# is server-rendered — parsed as a listing, keyword-gated to the AI/governance cut.
secrss:
  url: "https://www.secrss.com/"
  keywords: [人工智能, 大模型, 算法, 智能体, 生成式, 深度合成, AIGC, 数据安全, 数据出境, 个人信息, 算力, 监管, 治理]

# AIIA (aiiaorg.cn) working-group rosters: DEFERRED — the main site and the
# aihub.caict.ac.cn/aiia_wgs 工作组 page are both JS-rendered SPAs (no static roster
# content). Needs the backing endpoint (browser devtools) to wire.

huggingface:            # huggingface_rel.py: model releases (HF handles ≠ GitHub handles)
  per_org: 8
  orgs:
    - deepseek-ai
    - Qwen
    - zai-org            # Zhipu / GLM (not THUDM on HF)
    - moonshotai
    - internlm
    - OpenGVLab
    - ByteDance-Seed
    - 01-ai
    - baichuan-inc
    - MiniMaxAI          # not MiniMax-AI on HF
    - stepfun-ai
    - tencent
    - baidu
    - openbmb
    - inclusionAI
    - Skywork

github:                 # github_rel.py: lab orgs — releases + model-card commits
  orgs:
    - deepseek-ai          # DeepSeek
    - QwenLM               # Alibaba Qwen
    - THUDM                # Zhipu / Tsinghua KEG (GLM, CogVLM)
    - MoonshotAI           # Kimi
    - InternLM             # Shanghai AI Lab
    - OpenGVLab            # Shanghai AI Lab (InternVL, InternImage)
    - ByteDance-Seed       # ByteDance Seed
    - 01-ai                # 01.AI (Yi)
    - baichuan-inc         # Baichuan
    - MiniMax-AI           # MiniMax
    - stepfun-ai           # StepFun
    - Tencent-Hunyuan      # Tencent Hunyuan
    - PaddlePaddle         # Baidu ERNIE lives under PaddlePaddle/ERNIE
    - OpenBMB              # MiniCPM / CPM
    - inclusionAI          # Ant Group (Ling / Bailing)
    - SkyworkAI            # Skywork / Kunlun

# ─────────────────────────────────────────────────────────────────────────────
# Russia edition (region=russia) — official/state feeds, title-gated to AI/tech.
# Renders in a 🇷🇺 Russia block in the same email. Gate mixes Russian stems +
# English (no bare "ИИ" — collides with "России"). Conservative: 3 official/state
# sources; validate before adding corporate (Sber/Yandex) or think tanks.
# ─────────────────────────────────────────────────────────────────────────────
russia:
  keywords:
    # Russian stems
    - искусственн          # искусственный интеллект (AI)
    - нейросет             # нейросеть (neural net)
    - машинное обучение
    - большая языковая модель
    - полупроводник        # semiconductor
    - микроэлектроник
    - суперкомпьютер
    - квантов              # quantum
    - беспилотн            # autonomous/UAV
    - гигачат              # GigaChat (Sber)
    - yandexgpt
    # tech-sovereignty / chip policy — the real Russia "frontier" signal
    - импортозамещени      # import substitution (chips)
    - технологическ сувере # technological sovereignty
    - цифровой суверенит    # digital sovereignty
    # English (for TASS en + English mentions)
    - artificial intelligence
    - machine learning
    - neural network
    - semiconductor
    - microelectronic
    - supercomputer
    - quantum comput
    - large language model
  sources:
    - name: Кремль (Kremlin)
      url: http://kremlin.ru/events/president/news/feed   # Russian-language feed
      category: regulation
    - name: TASS
      url: https://tass.com/rss/v2.xml                     # English state news
      category: discourse
    - name: Government.ru
      url: http://government.ru/all/rss/                    # Russian-language feed
      category: regulation

# ─────────────────────────────────────────────────────────────────────────────
# Delivery
# ─────────────────────────────────────────────────────────────────────────────
digest:
  # Resend's test sender only delivers to your Resend account email
  # (syntax606@gmail.com). Once you verify a domain, switch the sender below and
  # you can send anywhere (e.g. css216@nyu.edu).
  recipients: [syntax606@gmail.com]
  # Resend's test sender — no domain verification needed, but ONLY delivers to the
  # email your Resend account is registered under. Swap to a verified domain
  # (e.g. digest@yourdomain.com) for production / arbitrary recipients.
  sender: "Parallel Frontier <onboarding@resend.dev>"
  send_hour_utc: 6          # ~07:30 UK in summer; fixed UTC, accept DST drift (see SPEC)
  provider: resend          # resend | postmark  (API key in GitHub secrets: RESEND_API_KEY)
  capture_url: ""           # optional footer link to your manual capture file
# alert:
#   telegram_chat_id: ""    # optional: mirror alert-tier only, for true interrupts