alloevil · Oct 17, 2026 · Oct 17, 2026
diff --git a/README.md b/README.md
@@ -85,7 +85,7 @@ Every day it collects signals from 6 data sources, runs them through a smart sco
 | GitHub Search | New & rising | Repos created in the last 7 days with fast star growth |
 | [Hacker News](https://news.ycombinator.com/) | Community picks | GitHub repos from Show HN posts |
 | Rising Detection | Early signal | Repos created in the last 3 days with a fork/star ratio above 0.3 and at most 0.5 — forks as an early usage signal; higher ratios are discarded as fork farms |
-| AI/ML Keyword Sweep | AI focus | GitHub Search over an AI/ML keyword list, 5 keywords per day on a rotating schedule (inspired by [OSSInsight trending/ai](https://ossinsight.io/trending/ai), implemented against the GitHub Search API) |
+| AI/ML Keyword Sweep | AI focus | GitHub Search over an AI/ML keyword list, 5 keywords per day on a rotating schedule, plus 2 semiconductor / industrial-AI keywords per day over repos pushed in the last 30 days (inspired by [OSSInsight trending/ai](https://ossinsight.io/trending/ai), implemented against the GitHub Search API) |
 | [HF Daily Papers](https://huggingface.co/papers) | Research signal | GitHub repos linked from trending Hugging Face papers — we assume paper upvotes lead GitHub stars by days, but no committed measurement backs that lead time yet |
 
 ### Smart Scoring (100 points)
@@ -104,11 +104,22 @@ Every day it collects signals from 6 data sources, runs them through a smart sco
 - **Star fraud detection**: age ≤ 1 day with 1000+ stars, age ≤ 2 days with 2000+ stars, or 1000+ stars/day with an empty description → flagged
 - **Batch fraud detection**: same owner with ≥ 3 repos in one batch, or ≥ 2 repos under 7 days old already past 200 stars → flagged
 - **Content quality**: No description or no README → penalty
-- **Cross-day dedup**: 7-day window, no duplicate recommendations across days (a same-day rerun may repeat a pick)
+- **Cross-day dedup**: 7-day window for the daily report (a same-day rerun may repeat a pick)
+- **Picks never repeat**: a repo that has appeared in a picks issue (or was recommended in chat) is never picked again — see `data/picks_history.json`
+
+### Picks (every 3 days)
+
+Collection still runs daily; every `PICKS_EVERY_DAYS` (3) days `scripts/picks.py` publishes `data/picks-YYYY-MM-DD.json`:
+
+- 8–10 repos (`PICKS_MIN`/`PICKS_MAX`), at least `SEMI_MIN` (2) tagged semiconductor / industrial AI when the pool has them (`semi_min_met` says whether it did); AI-tagged repos rank ahead of untagged ones
+- every pick is re-read from `GET /repos/{owner}/{repo}` right before publishing; missing, archived or disabled repos are dropped
+- facts only: stars, forks, topics, dates and the sources that surfaced it; `growth` is a measured delta from the committed star snapshots, or `null` — never a lifetime average
+- no repeat performers, and no LLM in the pipeline; fewer than 8 verified candidates means no issue that day
+- `picks-never-repeat` and `picks-count-in-range` in `docs/claims.json` re-check every committed issue
 
 ### Email Subscription
 
-- Daily curated repos delivered to your inbox
+- Each picks issue (the same 8–10 repos) is delivered to your inbox; non-picks days send no email
 - Rendered dark-only, with `color-scheme` declared for dark-aware clients (Apple Mail / iOS)
 - Powered by Resend API
 
@@ -245,6 +256,7 @@ github-discovery/
 │   ├── quality.py           # Code quality detection
 │   ├── anti_spam.py         # Anti-spam scoring dimension
 │   ├── dedup.py             # Cross-day deduplication (7-day window)
+│   ├── picks.py             # Picks issue every 3 days (permanent dedup, facts only)
 │   ├── fraud_detection.py   # Batch fraud detection
 │   ├── snapshots.py         # Daily star snapshots (real growth)
 │   ├── verify_scoring.py    # Scoring verification / backtesting
@@ -324,7 +336,7 @@ have since moved.
 
 **Do I need any paid service?** Only for email. `RESEND_API_KEY` sends the digest through Resend; `FIRECRAWL_API_KEY` is optional and only makes GitHub Trending parsing more robust than regex-matching raw HTML. Everything else runs on the GitHub Actions free tier with the default `GITHUB_TOKEN`.
 
-**Will the same repo be recommended every day while it is hot?** No. Cross-day dedup blocks any repo recommended in the previous 7 days, using the history committed at `data/recommend_history.json`; records older than 30 days are cleaned up. The history has to be a committed file because every CI run starts from a fresh checkout with no local state.
+**Will the same repo be recommended every day while it is hot?** No. Picks never repeat a repo (permanent history in `data/picks_history.json`). The daily report's cross-day dedup blocks any repo recommended in the previous 7 days, using the history committed at `data/recommend_history.json`; records older than 30 days are cleaned up. The history has to be a committed file because every CI run starts from a fresh checkout with no local state.
 
 **Can I trust the numbers on the published page?** The page is a build artifact: `scripts/generate_site.py` renders `docs/index.html` from `docs/template.html` plus the committed reports, and the workflow publishes `docs/` to the `gh-pages` branch. The authoritative copies are the committed JSON under `data/`, which is what `verify_scoring.py` reads; the Markdown digests under `output/` are the human-readable copies, and the site reads the JSON first and falls back to those digests. So any claim on the page can be recomputed from a fresh clone.
 

diff --git a/README.zh-CN.md b/README.zh-CN.md
@@ -85,7 +85,7 @@ GitHub Discovery 告诉你什么**即将火起来**——增长曲线异常的
 | GitHub Search | 新晋上升 | 最近 7 天创建、星标增长迅猛的仓库 |
 | [Hacker News](https://news.ycombinator.com/) | 社区精选 | Show HN 帖子中出现的 GitHub 仓库 |
 | Rising Detection | 早期信号 | 最近 3 天创建、fork/star 比大于 0.3 且不超过 0.5 的仓库——fork 作为真实使用的早期信号；比率更高的按 fork farm 直接排除 |
-| AI/ML 关键词轮扫 | AI 方向 | 用一组 AI/ML 关键词扫 GitHub Search，每天轮换 5 个词（思路来自 [OSSInsight trending/ai](https://ossinsight.io/trending/ai)，实现走的是 GitHub Search API） |
+| AI/ML 关键词轮扫 | AI 方向 | 用一组 AI/ML 关键词扫 GitHub Search，每天轮换 5 个词，另加每天 2 个半导体/工业 AI 关键词（看近 30 天有推送的仓库）（思路来自 [OSSInsight trending/ai](https://ossinsight.io/trending/ai)，实现走的是 GitHub Search API） |
 | [HF Daily Papers](https://huggingface.co/papers) | 研究信号 | Hugging Face 热门论文关联的 GitHub 仓库——我们假设论文点赞领先 GitHub 星标几天，但仓库里还没有可提交的测量来支撑这个领先时间 |
 
 ### 智能评分（满分 100 分）
@@ -104,11 +104,22 @@ GitHub Discovery 告诉你什么**即将火起来**——增长曲线异常的
 - **刷星检测**：年龄 ≤ 1 天且 1000+ 星、年龄 ≤ 2 天且 2000+ 星，或日增 1000+ 星且描述为空 → 标记
 - **批量刷量检测**：同一作者一批里出现 ≥ 3 个仓库，或 ≥ 2 个上线不足 7 天的仓库已过 200 星 → 标记
 - **内容质量**：没有描述或没有 README → 扣分
-- **跨日去重**：7 天窗口内跨天不重复推荐（同日重跑可能重复选到同一个仓库）
+- **跨日去重**：每日报告 7 天窗口内跨天不重复推荐（同日重跑可能重复选到同一个仓库）
+- **推荐期永不重复**：进过推荐期（或在聊天里推过）的仓库永不再推，见 `data/picks_history.json`
+
+### 推荐期（每 3 天一期）
+
+采集仍是每天一次；每 `PICKS_EVERY_DAYS`（3）天，`scripts/picks.py` 发布一期 `data/picks-YYYY-MM-DD.json`：
+
+- 每期 8–10 个（`PICKS_MIN`/`PICKS_MAX`），候选里有的话至少 `SEMI_MIN`（2）个半导体/工业 AI 仓库（是否达标见 `semi_min_met`）；带 AI 标签的排在无主题的前面
+- 发布前逐个用 `GET /repos/{owner}/{repo}` 重新读取，不存在、已归档或已禁用的剔除
+- 只放事实：star、fork、topics、日期和发现它的数据源；`growth` 是按已提交 star 快照算出的实测增量，拿不到就是 `null`，绝不用终身平均值顶替
+- 不带 repeat，流程里不接大模型；核实后不足 8 个当天就不出刊
+- `docs/claims.json` 里的 `picks-never-repeat` 和 `picks-count-in-range` 会复核每一期
 
 ### 邮件订阅
 
-- 每日精选仓库直达你的收件箱
+- 每期推荐（同一份 8–10 个）直达你的收件箱；非推荐日不发邮件
 - 只渲染深色版本，并声明 `color-scheme` 供支持深色的客户端使用（Apple Mail / iOS）
 - 基于 Resend API 发送
 
@@ -198,6 +209,7 @@ github-discovery/
 │   ├── quality.py           # Code quality detection
 │   ├── anti_spam.py         # Anti-spam scoring dimension
 │   ├── dedup.py             # Cross-day deduplication (7-day window)
+│   ├── picks.py             # Picks issue every 3 days (permanent dedup, facts only)
 │   ├── fraud_detection.py   # Batch fraud detection
 │   ├── snapshots.py         # Daily star snapshots (real growth)
 │   ├── verify_scoring.py    # Scoring verification / backtesting
@@ -274,7 +286,7 @@ python scripts/verify_scoring.py --days 30
 
 **需要付费服务吗？** 只有邮件需要。`RESEND_API_KEY` 用于通过 Resend 发送日报；`FIRECRAWL_API_KEY` 是可选的，只是让 GitHub Trending 的解析比正则匹配原始 HTML 更健壮。其余部分都跑在 GitHub Actions 免费额度上，用默认的 `GITHUB_TOKEN` 即可。
 
-**一个仓库火着的时候会天天被推荐吗？** 不会。跨日去重会拦掉过去 7 天内推荐过的仓库，依据是提交在 `data/recommend_history.json` 里的推荐历史，超过 30 天的记录会被清理。这份历史必须是提交进仓库的文件，因为每次 CI 都是全新 checkout，没有任何本地状态。
+**一个仓库火着的时候会天天被推荐吗？** 不会。推荐期永不重复推同一个仓库（永久历史见 `data/picks_history.json`）；每日报告的跨日去重会拦掉过去 7 天内推荐过的仓库，依据是提交在 `data/recommend_history.json` 里的推荐历史，超过 30 天的记录会被清理。这份历史必须是提交进仓库的文件，因为每次 CI 都是全新 checkout，没有任何本地状态。
 
 **网页上的数字可信吗？** 网页是构建产物：`scripts/generate_site.py` 用 `docs/template.html` 加上已提交的报告渲染出 `docs/index.html`，工作流再把整个 `docs/` 发布到 `gh-pages` 分支。真正权威的是提交在 `data/` 下的 JSON，`verify_scoring.py` 读的就是它；`output/` 下的 Markdown 日报是给人看的副本，站点优先读 JSON、读不到才回退到日报。所以网页上的任何结论都能在全新 clone 上重算一遍。
 

diff --git a/data/picks_history.json b/data/picks_history.json
@@ -0,0 +1,15 @@
+{
+  "repos": {
+    "openai/math": {"first_picked": "2026-10-16", "source": "chat"},
+    "NandhaKishorM/laya": {"first_picked": "2026-10-16", "source": "chat"},
+    "morluto/rea": {"first_picked": "2026-10-16", "source": "chat"},
+    "zai-org/ZCode": {"first_picked": "2026-10-16", "source": "chat"},
+    "feder-cr/git-ai": {"first_picked": "2026-10-16", "source": "chat"},
+    "anthropics/oss-scanner": {"first_picked": "2026-10-16", "source": "chat"},
+    "rociiu/talorys": {"first_picked": "2026-10-16", "source": "chat"},
+    "chenwei810/Semiconductor_Manufacturing_Introduction": {"first_picked": "2026-10-16", "source": "chat"},
+    "hdl-tools/digital-chip-design-agents": {"first_picked": "2026-10-16", "source": "chat"},
+    "BLANK2077/xverif": {"first_picked": "2026-10-16", "source": "chat"}
+  },
+  "updated_at": "2026-10-16T17:00:00-07:00"
+}
diff --git a/docs/ROADMAP.md b/docs/ROADMAP.md
@@ -8,7 +8,7 @@
 - **100-point scorer** — acceleration (40) + quality (30) + anti-spam (30), with code-quality bonus, suspicious-star and batch-fraud penalties, and an explainable reason line on every card
 - **Cross-day dedup** with a 7-day window, deep checks budgeted to the top-K candidates after coarse ranking
 - **Delivery** — daily email digest via Resend (dark-mode HTML) + GitHub Pages site with date/language filters
-- **Current quality baseline** — 202 tests and 20 executable receipts; the site, feed and daily JSON are generated from committed reports. The presentation pass (light/dark themes, two-band header, merged report header, lead-card treatment) is complete; further UI work is gated by a user task or a measured defect.
+- **Current quality baseline** — 210 tests and 22 executable receipts; the site, feed and daily JSON are generated from committed reports. The presentation pass (light/dark themes, two-band header, merged report header, lead-card treatment) is complete; further UI work is gated by a user task or a measured defect.
 
 ## Long-term operating plan
 

diff --git a/docs/claims.json b/docs/claims.json
@@ -213,7 +213,9 @@
       "verified": "2026-09-26",
       "check": {
         "cmd": "python3 - <<'PY'\nimport glob, json\npaths = sorted(glob.glob('data/candidate-pool-*.json'))\nif not paths:\n    print('candidate pool: not-ready (no production pool yet)')\n    raise SystemExit(0)\npath = paths[-1]\ndoc = json.load(open(path, encoding='utf-8'))\nrows = doc.get('candidates') or []\nallowed = {'candidate', 'deduplicated', 'content_blocked', 'eligible', 'recommended'}\nnames = [r.get('full_name') for r in rows]\nproblems = []\nif len(names) != len(set(names)): problems.append('duplicate names')\nfor row in rows:\n    if not row.get('full_name'): problems.append('missing name')\n    if not (row.get('sources') or row.get('source')): problems.append('missing source')\n    if row.get('pool_status') not in allowed: problems.append('bad status')\nprint('candidate pool', path, 'rows', len(rows), 'problems', sorted(set(problems)))\nassert rows and not problems\nPY",
-        "expect": {"regex": "candidate pool: not-ready|candidate pool .* problems \\[\\]"},
+        "expect": {
+          "regex": "candidate pool: not-ready|candidate pool .* problems \\[\\]"
+        },
         "timeout": 60
       }
     },
@@ -228,7 +230,9 @@
       "as_of": "2026-09-26",
       "check": {
         "cmd": "python3 -c \"import re,subprocess,pathlib;out=subprocess.run(['python3','-m','pytest','--collect-only','-q'],capture_output=True,text=True).stdout;got=int(re.search(r'(\\d+) tests? collected', out).group(1));doc=int(re.search(r'(\\d+) tests', pathlib.Path('docs/ROADMAP.md').read_text()).group(1));print('pytest', got, 'roadmap', doc);assert got==doc\"",
-        "expect": {"regex": "pytest \\d+ roadmap \\d+$"},
+        "expect": {
+          "regex": "pytest \\d+ roadmap \\d+$"
+        },
         "timeout": 180
       }
     },
@@ -351,6 +355,40 @@
         },
         "timeout": 120
       }
+    },
+    {
+      "id": "picks-never-repeat",
+      "claim": "No repository appears in more than one picks issue, every pick is recorded in the permanent history, and none repeats a repository already recommended in chat.",
+      "value": "0 repeats",
+      "metric": "duplicate full_name across data/picks-*.json, plus picks missing from or colliding with data/picks_history.json",
+      "method": "scripts/picks.py filters candidates against data/picks_history.json, which is never pruned; this receipt re-derives the property from the committed files alone.",
+      "repro": "ls data/picks-*.json",
+      "evidence": "data/picks_history.json",
+      "as_of": "2026-10-16",
+      "check": {
+        "cmd": "python3 - <<'PY'\nimport glob, json\npaths = sorted(glob.glob('data/picks-*.json'))\nhist = json.load(open('data/picks_history.json', encoding='utf-8'))['repos']\nseen, problems = {}, []\nfor path in paths:\n    for p in json.load(open(path, encoding='utf-8'))['picks']:\n        k = p['full_name'].lower()\n        if k in seen: problems.append(f\"{p['full_name']} in {seen[k]} and {path}\")\n        seen[k] = path\n        if p['full_name'] not in hist: problems.append(f\"{p['full_name']} missing from history\")\n        elif hist[p['full_name']].get('source') == 'chat': problems.append(f\"{p['full_name']} repeats a chat pick\")\nprint('picks files', len(paths), 'repos', len(seen), 'problems', problems)\nassert not problems\nPY",
+        "expect": {
+          "regex": "problems \\[\\]$"
+        },
+        "timeout": 60
+      }
+    },
+    {
+      "id": "picks-count-in-range",
+      "claim": "Every picks issue holds 8 to 10 repositories with GitHub URLs, integer star counts, and a growth field that is either a measured snapshot delta or null.",
+      "value": "8-10 per issue",
+      "metric": "len(picks) per data/picks-*.json and the shape of its url, stars and growth fields",
+      "method": "scripts/picks.py publishes an issue only when at least PICKS_MIN verified repositories exist and never substitutes lifetime averages for growth; this receipt checks every committed issue.",
+      "repro": "ls data/picks-*.json",
+      "evidence": "data/picks-*.json",
+      "as_of": "2026-10-16",
+      "check": {
+        "cmd": "python3 - <<'PY'\nimport glob, json\npaths = sorted(glob.glob('data/picks-*.json'))\nproblems = []\nfor path in paths:\n    d = json.load(open(path, encoding='utf-8'))\n    n = len(d['picks'])\n    if not 8 <= n <= 10 or d['count'] != n: problems.append(f'{path}: count {n}')\n    for p in d['picks']:\n        if not p['url'].startswith('https://github.com/'): problems.append(f\"{p['full_name']}: url\")\n        if not isinstance(p['stars'], int): problems.append(f\"{p['full_name']}: stars\")\n        g = p['growth']\n        if g is not None and (g.get('source') != 'star_snapshots' or not isinstance(g.get('stars_gained'), int)):\n            problems.append(f\"{p['full_name']}: growth not measured\")\nprint('picks files', len(paths), 'problems', problems)\nassert not problems\nPY",
+        "expect": {
+          "regex": "problems \\[\\]$"
+        },
+        "timeout": 60
+      }
     }
   ]
 }
diff --git a/scripts/main.py b/scripts/main.py
@@ -27,6 +27,9 @@
 from fraud_detection import detect_batch_fraud, apply_fraud_penalty
 from snapshots import record_snapshots, get_growth
 from watchlist import add_discoveries
+import picks as picks_mod
+from sources import _gh_api
+from snapshots import _load as load_snapshots
 
 
 def get_subscribers() -> list[str]:
@@ -586,7 +589,24 @@ def main():
     # been sent (or benignly skipped). If the email fails, we exit non-zero
     # without writing the marker, so the next scheduled run retries.
     date_str = datetime.now().strftime("%Y-%m-%d")
-    email_result = send_digest_email(date_str, top_new)
+
+    # ── 推荐期（picks）：每 PICKS_EVERY_DAYS 天一期，邮件和聊天共用这一份 ──
+    # 非推荐日只采集（报告/站点/快照照常），不发邮件。
+    chosen = []
+    if picks_mod.is_picks_day():
+        ordered = picks_mod.select_candidates(scored)
+        chosen = picks_mod.verify_picks(ordered, _gh_api)
+        if len(chosen) < picks_mod.PICKS_MIN:
+            print(f"[Picks] only {len(chosen)} verified candidates (< {picks_mod.PICKS_MIN}); "
+                  "no picks issue today, no email.")
+            chosen = []
+    else:
+        print("[Picks] not a picks day — collect only, no email.")
+
+    if chosen:
+        email_result = send_digest_email(date_str, chosen)
+    else:
+        email_result = "skipped"
     if email_result == "failed":
         print("[ERROR] Email send failed — not writing today's report so the "
               "next run retries. Exiting non-zero to flag the failure.")
@@ -599,6 +619,11 @@ def main():
         f.write(md)
     print(f"\n[Saved] Report written to {out_path}")
 
+    if chosen:
+        picks_path = picks_mod.write_picks(date_str, chosen, load_snapshots())
+        picks_mod.record_picks([r for r, _ in chosen], date_str)
+        print(f"[Saved] Picks written to {picks_path}")
+
     # Persist source health only after ranking, dedup and recommendation selection are complete.
     set_source_recommendations([repo for repo, _scores in top_new + top_repeat])
     source_health = get_source_health()