diff --git a/.github/workflows/jekyll.yml b/.github/workflows/jekyll.yml
index c98e0b8..98af82e 100644
--- a/.github/workflows/jekyll.yml
+++ b/.github/workflows/jekyll.yml
@@ -3,6 +3,12 @@ name: Build and deploy Jekyll site to GitHub Pages
 on:
   push:
     branches: [ "main" ]
+  # Runs the same four gates on the PR, so a change that breaks the build, the
+  # plugin tests, an internal link, or the sitemap/metadata checks is caught
+  # before it lands. Without this the gates only ran after a merge: a bad PR
+  # reached main and merely failed to deploy, leaving main red.
+  pull_request:
+    branches: [ "main" ]
   workflow_dispatch:
 
 permissions:
@@ -69,9 +75,12 @@ jobs:
         # this site actually shipped.
         run: script/validate-site.sh ./_site
       - name: Upload artifact
+        # Nothing to publish from a PR run; the gates above are the whole point.
+        if: github.event_name != 'pull_request'
         uses: actions/upload-pages-artifact@v3
 
   deploy:
+    if: github.event_name != 'pull_request'
     environment:
       name: github-pages
       url: ${{ steps.deployment.outputs.page_url }}
diff --git a/README.ko.md b/README.ko.md
index fa82300..c2ba49e 100644
--- a/README.ko.md
+++ b/README.ko.md
@@ -87,10 +87,10 @@ search.json        전체 본문 검색 색인 (simple-jekyll-search가 사용)
 test/              _plugins/ 로직의 minitest 단위 테스트
 script/            validate-site.sh (빌드 후 검색 노출 검사)
 sitemap-index.xml  사이트맵 인덱스 — Search Console에 제출할 URL
-.github/workflows/ CI: 빌드 → html-proofer → validate-site → main 푸시 시 배포
+.github/workflows/ CI: 테스트 → 빌드 → html-proofer → validate-site, main 푸시 시 배포
 ```
 
-**최상위 페이지:** `index.html`(홈), 네 개의 섹션 탭인 `paper-reviews.md`·
+**최상위 페이지:** `index.html`(홈), 네 개의 섹션 페이지인 `paper-reviews.md`·
 `paper-summaries.md`·`tech-guides.md`·`insights.md`, 그리고 `categories.html`·
 `tags.html`·`search.md`·`about.md`.
 
@@ -138,10 +138,34 @@ RSS `<summary>`에 실리는 문장입니다. 생략하면 `_plugins/post_descri
 카테고리는 **두 단계**입니다.
 
 - `categories[0]` — **유형**: `Paper Reviews`, `Paper Summaries`, `Tech Guides`,
-  `Insights` 중 하나. 글이 어느 내비 탭에 들어갈지를 정합니다.
+  `Insights` 중 하나. 글이 어느 섹션 페이지에 들어갈지를 정합니다. 현재 글이 있는 유형은
+  `Paper Reviews`와 `Insights`뿐이라 내비에도 이 둘만 있습니다 — `paper-summaries.md`와
+  `tech-guides.md`는 `main_nav` 없이 `sitemap: false` 상태이고, 첫 글이 들어오면 어떤 두
+  값을 되살리면 되는지가 각 파일 프런트매터에 적혀 있습니다.
 - `categories[1]` — **주제**: `Language-Models`, `Multimodal-Learning`,
   `Finetuning`, `Retrieval-Augmented-Generation`, `Agentic-AI` 등. 필요하면 자유롭게 추가합니다.
 
+### 태그: 세부 태그 위에 주제 태그를 하나
+
+태그는 자유 형식·하이픈 연결이고, 대부분 논문 하나의 기여를 서술합니다
+(`Fine-Grained-Expert-Segmentation`). 색인으로는 좋지만 글을 **잇지는** 못합니다 — 태그
+264개 중 220개가 단 한 편에만 붙어 있었고, `_plugins/related_posts.rb`는 **공유** 태그를
+요구하므로 11편이 관련 글을 하나도 얻지 못했습니다. 서로 남남이던 그래프 RAG 논문 네 편이
+거기 있었습니다.
+
+그래서 글마다 아래 통제된 주제 태그를 최소 하나 답니다.
+
+```
+Agentic-AI  Alignment  DeepSeek  Knowledge-Graph  Llama
+Mixture-of-Experts  Multimodal-Models  Reasoning-Models
+Retrieval-Augmented-Generation
+```
+
+기존 세부 태그는 **그대로 남깁니다.** 주제 태그가 담지 못하는 내용을 담고 있고, 지우면 이미
+살아 있는 `/tags/` 앵커가 움직입니다. 새 주제 태그를 만들기 전에 위 목록이나 이미 두 편
+이상이 쓰는 태그로 덮이는지 먼저 확인하세요. `Agentic-Architecture`·`Agentic-Patterns`·
+`Agentic-Infrastructure`가 한 주제의 세 이름이었던 것이 그 묶음에 링크가 없던 이유입니다.
+
 Jekyll은 이 둘과 날짜를 합쳐 출력 경로를 만듭니다.
 
 ```
@@ -192,7 +216,9 @@ CI도 같은 세 검사를 돌립니다. 로컬에서 미리 잡으면 배포 
 
 ## 배포
 
-`main`에 푸시하면 `.github/workflows/jekyll.yml`이 다음을 수행합니다.
+`.github/workflows/jekyll.yml`은 `main` 푸시뿐 아니라 **`main`으로 향하는 PR에서도**
+돕니다. 그래서 아래 네 게이트가 잘못된 머지를 보고만 하지 않고 막습니다. 1~4단계는 양쪽
+모두에서, 5단계는 PR에서 건너뜁니다.
 
 1. **`ruby test/run_all.rb`**로 `_plugins/` 단위 테스트를 돌리고
 2. `JEKYLL_ENV=production`으로 사이트를 빌드하고
diff --git a/README.md b/README.md
index 31f76f3..d6d98a4 100644
--- a/README.md
+++ b/README.md
@@ -89,11 +89,11 @@ search.json        Full-text search index (consumed by simple-jekyll-search)
 test/              minitest unit tests for the _plugins/ logic
 script/            validate-site.sh (post-build discoverability checks)
 sitemap-index.xml  Sitemap index — the URL to submit to Search Console
-.github/workflows/ CI: build → html-proofer → validate-site → deploy on push to main
+.github/workflows/ CI: tests → build → html-proofer → validate-site; deploys on push to main
 ```
 
 **Top-level pages:** `index.html` (home), plus `paper-reviews.md`,
-`paper-summaries.md`, `tech-guides.md`, `insights.md` (the four section tabs),
+`paper-summaries.md`, `tech-guides.md`, `insights.md` (the four section pages),
 `categories.html`, `tags.html`, `search.md`, and `about.md`.
 
 ---
@@ -142,11 +142,38 @@ share one.
 Categories are **two levels**:
 
 - `categories[0]` — the **type**: `Paper Reviews`, `Paper Summaries`,
-  `Tech Guides`, or `Insights`. This decides which nav tab the post appears under.
+  `Tech Guides`, or `Insights`. This decides which section page the post appears
+  on. Only `Paper Reviews` and `Insights` have posts today, so only those two are
+  in the nav — `paper-summaries.md` and `tech-guides.md` sit at `sitemap: false`
+  with no `main_nav`, and their front matter says which two keys to restore when
+  the first post lands in either.
 - `categories[1]` — the **topic**: `Language-Models`, `Multimodal-Learning`,
   `Finetuning`, `Retrieval-Augmented-Generation`, `Agentic-AI`, … (add new ones
   freely).
 
+### Tags: one topic tag on top of the specific ones
+
+Tags are free-form and hyphenated, and most describe one paper's contribution
+(`Fine-Grained-Expert-Segmentation`). Those are a good index but they connect
+nothing: 220 of 264 tags were used by exactly one post, and since
+`_plugins/related_posts.rb` requires a **shared** tag, 11 posts got no related
+reading at all — including four graph-RAG papers that share no tag with each
+other.
+
+So also give each post at least one tag from the controlled topic layer:
+
+```
+Agentic-AI  Alignment  DeepSeek  Knowledge-Graph  Llama
+Mixture-of-Experts  Multimodal-Models  Reasoning-Models
+Retrieval-Augmented-Generation
+```
+
+Keep the specific tags — they say something the topic tag does not, and dropping
+them would move live `/tags/` anchors. Before inventing a topic tag, check
+whether one of the above (or an existing tag used by 2+ posts) already covers it:
+`Agentic-Architecture`, `Agentic-Patterns` and `Agentic-Infrastructure` were three
+names for one thing, which is why that cluster had no links.
+
 Jekyll combines them with the date to build the output path:
 
 ```
@@ -198,7 +225,9 @@ changes every page on the site**, so add a case there before changing behaviour.
 
 ## Deployment
 
-Pushing to `main` triggers `.github/workflows/jekyll.yml`, which:
+`.github/workflows/jekyll.yml` runs on **pull requests to `main` as well as
+pushes to it**, so the four gates below block a bad merge instead of only
+reporting one. Steps 1–4 run on both; step 5 is skipped for pull requests. It:
 
 1. runs **`ruby test/run_all.rb`** (the `_plugins/` unit tests),
 2. builds the site with `JEKYLL_ENV=production`,
diff --git a/_config.yml b/_config.yml
index 1b8ae5c..bc452ee 100644
--- a/_config.yml
+++ b/_config.yml
@@ -13,8 +13,6 @@ feed:
 # breaks search indexing.
 exclude:
   - /vendor/
-  - circle.yml
-  - stackbit.yaml
   - README.md
   - README.ko.md
   - docs/
@@ -65,11 +63,10 @@ kramdown:
 paginate: 5
 paginate_path: /page/:num/
 
-simple_jekyll_search:
-  exclude:
-    - about.md
-    - search.md
-    - index.html
+# NOTE: no `simple_jekyll_search:` block. simple-jekyll-search is a client-side
+# script that reads search.json and knows nothing about Jekyll config, so the
+# `exclude:` list that used to sit here excluded nothing. search.json is built
+# from `site.posts` only, which already leaves out the pages it named.
 
 ga_tracking_id: G-MRH9Z5ETSQ
 
diff --git a/_includes/footer.html b/_includes/footer.html
index f1ac17f..c05a53e 100644
--- a/_includes/footer.html
+++ b/_includes/footer.html
@@ -61,13 +61,11 @@ <h2 class="footer-heading">{{ site.title }}</h2>
 
 <script src="{{ '/js/main.js' | prepend: site.baseurl }}" defer></script>
 
-{% if page.custom_js %}
-<!-- Custom page specific js files -->
-  {% for js_file in page.custom_js %}
-    <script src="{{ '/js/' | prepend: site.baseurl }}{{ js_file }}.js"></script>
-  {% endfor %}
-{% endif %}
+<!-- Removed from here: a `page.custom_js` loop that no page ever set (pages
+     needing an extra asset use `head_extra`, as search.md does), and a
+     conditional on site.ga_tracking_id whose entire body was a comment saying
+     the tracking lives in head.html. Both were inherited theme scaffolding.
 
-{% if site.ga_tracking_id %}
-<!-- Google Analytics 4 tracking is implemented in head.html -->
-{% endif %}
+     Note for whoever edits this comment: do not write Liquid tag delimiters in
+     here. Liquid tokenizes tags inside comments too — HTML and Liquid alike —
+     so a tag named in prose is opened for real and the build fails. -->
diff --git a/_includes/head.html b/_includes/head.html
index 93606ec..a9d4a33 100644
--- a/_includes/head.html
+++ b/_includes/head.html
@@ -44,20 +44,27 @@
 {%- endif -%}
 <head>
 	<meta charset="utf-8" />
-	<!-- Apply saved theme before paint to avoid a flash of the wrong colors -->
+	<meta http-equiv="X-UA-Compatible" content="IE=edge" />
+	<meta name="viewport" content="width=device-width, initial-scale=1" />
+	<!-- Light is the default; dark is opt-in via the toggle, not the OS setting.
+	     The matching `color-scheme` lives in _sass/_dark.scss, keyed off
+	     data-theme, because a meta tag cannot vary by theme. -->
+	<meta name="theme-color" content="#ffffff" />
+	<!-- Applies the saved theme before first paint, so a dark-mode visitor gets no
+	     flash of the light palette. Placed after the meta above, not before it:
+	     this also has to repaint the browser UI (the mobile address bar), and the
+	     tag has to exist to be found. #0d1117 is $bg in _sass/_dark.scss. -->
 	<script>
 		(function () {
 			try {
 				var t = localStorage.getItem('theme');
-				if (t === 'dark' || t === 'light') document.documentElement.setAttribute('data-theme', t);
+				if (t !== 'dark' && t !== 'light') return;
+				document.documentElement.setAttribute('data-theme', t);
+				var m = document.querySelector('meta[name="theme-color"]');
+				if (m) m.setAttribute('content', t === 'dark' ? '#0d1117' : '#ffffff');
 			} catch (e) {}
 		})();
 	</script>
-	<meta http-equiv="X-UA-Compatible" content="IE=edge" />
-	<meta name="viewport" content="width=device-width, initial-scale=1" />
-	<!-- Light is the default; dark is opt-in via the toggle, not the OS setting. -->
-	<meta name="color-scheme" content="light" />
-	<meta name="theme-color" content="#ffffff" />
 
 	<title>{{ page_title }}</title>
 	<meta name="description" content="{{ page_description | escape }}" />
@@ -98,7 +105,6 @@
 	<link rel="manifest" href="{{ '/assets/icons/manifest.json' | prepend: site.baseurl }}">
 	<meta name="msapplication-TileColor" content="#ffffff" />
 	<meta name="msapplication-TileImage" content="{{ '/assets/icons/ms-icon-144x144.png' | prepend: site.baseurl }}">
-	<!-- theme-color declared at top of <head> with prefers-color-scheme media queries -->
 
 	<!-- Facebook OGP cards -->
 	{% if site.protocols.fb_image %}
diff --git a/_layouts/archive.html b/_layouts/archive.html
deleted file mode 100644
index d9c4650..0000000
--- a/_layouts/archive.html
+++ /dev/null
@@ -1,26 +0,0 @@
----
-layout: default
----
-
-<div class="wrapper">
-	<div class="page">
-		<header class="post-header">
-			<h1 class="post-title">Posts in "{{ page.title }}"</h1>
-		</header>
-
-		{% include page_divider.html %}
-
-		<article class="post-content">
-			<ul class="posts-list">
-				{% for post in page.posts %}
-				<li>
-					<strong>
-						<a class="post-link" href="{{ post.url | prepend: site.baseurl }}">{{ post.title }}</a>
-					</strong>
-					<span class="post-date"> - {{ post.date | date: "%B %-d, %Y" }}</span>
-				</li>
-				{% endfor %}
-			</ul>
-		</article>
-	</div>
-</div>
diff --git a/_layouts/post.html b/_layouts/post.html
index 83004ac..4df7c38 100644
--- a/_layouts/post.html
+++ b/_layouts/post.html
@@ -60,18 +60,24 @@ <h1 class="title">{{ page.title }}</h1>
 
     {% include related_posts.html %}
 
-    {% if page.previous.url or page.next.url %}
+    {%- comment -%}
+      adjacent_older / adjacent_newer come from _plugins/related_posts.rb, not
+      Jekyll's page.previous / page.next. The built-ins walk raw date order, and
+      every translation pair here shares a date — so "← Previous" on the newest
+      Korean post was that same article in English.
+    {%- endcomment -%}
+    {% if page.adjacent_older or page.adjacent_newer %}
     <nav class="post-nav" aria-label="Adjacent posts">
-      {% if page.previous.url %}
-      <a class="post-nav-link prev" href="{{ page.previous.url | prepend: site.baseurl }}" rel="prev">
+      {% if page.adjacent_older %}
+      <a class="post-nav-link prev" href="{{ page.adjacent_older.url | prepend: site.baseurl }}" rel="prev">
         <span class="post-nav-dir">← Previous</span>
-        <span class="post-nav-title">{{ page.previous.title }}</span>
+        <span class="post-nav-title">{{ page.adjacent_older.title }}</span>
       </a>
       {% else %}<span></span>{% endif %}
-      {% if page.next.url %}
-      <a class="post-nav-link next" href="{{ page.next.url | prepend: site.baseurl }}" rel="next">
+      {% if page.adjacent_newer %}
+      <a class="post-nav-link next" href="{{ page.adjacent_newer.url | prepend: site.baseurl }}" rel="next">
         <span class="post-nav-dir">Next →</span>
-        <span class="post-nav-title">{{ page.next.title }}</span>
+        <span class="post-nav-title">{{ page.adjacent_newer.title }}</span>
       </a>
       {% else %}<span></span>{% endif %}
     </nav>
diff --git a/_plugins/related_posts.rb b/_plugins/related_posts.rb
index 20f98e3..962978e 100644
--- a/_plugins/related_posts.rb
+++ b/_plugins/related_posts.rb
@@ -1,13 +1,29 @@
 # frozen_string_literal: true
 #
-# Picks topically related posts for each post and exposes them as `page.related`.
+# Decides which other posts a post links to, and exposes them to the layout:
+#
+#   page.related          — topically related, ranked (see `rank`)
+#   page.adjacent_older   — the "← Previous" link in _layouts/post.html
+#   page.adjacent_newer   — the "Next →" link
+#
+# Both answers obey the same two eligibility rules (`eligible?`), which is why
+# they live together: never point a reader at this post's own translation, and
+# never cross languages. Jekyll's built-in `page.previous`/`page.next` know
+# neither rule, and all three translation pairs on this site share a date — so
+# the newest Korean post's "← Previous" was its own English version.
 #
 # A shared topic tag is required. Sharing only a category is not a reading
 # recommendation: "Paper Reviews / Language-Models" holds 17 posts, so category
 # overlap alone would put LLaMA and Llama 2 under DeepSeek-V3 with nothing
 # actually in common. Categories only break ties between posts that already share
-# a tag. An empty block is better than a wrong suggestion, so the 13 posts whose
-# tags are unique to them simply don't get one.
+# a tag, and an empty block is better than a wrong suggestion.
+#
+# That strictness only pays off if the tags can actually meet. They could not at
+# first: 220 of 264 tags were per-paper contribution phrases used by exactly one
+# post, and three near-synonyms (Agentic-Architecture / Agentic-Patterns /
+# Agentic-Infrastructure) split the one cluster the feature was built for — so 11
+# posts got nothing. The fix was a controlled topic layer in the posts' front
+# matter, not a looser rule here. All 35 posts now get suggestions.
 #
 # Jekyll's built-in site.related_posts is either "the 10 most recent posts" or
 # LSI, which needs the classifier gem and a slow indexing pass. This is cheaper
@@ -28,10 +44,7 @@ module RelatedPosts
   def self.rank(target, candidates, limit: LIMIT)
     scored = candidates.filter_map do |c|
       next if c[:id] == target[:id]
-      # A translation is the same article; the language switcher already links it.
-      next if translation_of?(target, c)
-      # Never suggest a Korean post to an English reader or the reverse.
-      next unless lang(c) == lang(target)
+      next unless eligible?(target, c)
 
       shared = shared_tags(target, c)
       next if shared.empty?
@@ -45,6 +58,28 @@ def self.rank(target, candidates, limit: LIMIT)
       .map { |c, _| c[:id] }
   end
 
+  # Nearest usable post in each direction, as [older, newer]. `ordered` must be
+  # sorted oldest-first. Unlike `rank` this never returns nothing for want of a
+  # shared tag — date order always has a neighbour — but it does skip past
+  # ineligible ones rather than dropping the link, so a post whose immediate
+  # neighbour is its own translation still gets the one beyond it.
+  def self.adjacent(target, ordered)
+    i = ordered.index { |c| c[:id] == target[:id] }
+    return [nil, nil] unless i
+
+    older = ordered[0...i].reverse.find { |c| eligible?(target, c) }
+    newer = ordered[(i + 1)..].to_a.find { |c| eligible?(target, c) }
+    [older, newer]
+  end
+
+  # A translation is the same article, and the language switcher already links
+  # it; a post in the other language is unreadable to whoever is here.
+  def self.eligible?(target, candidate)
+    return false if translation_of?(target, candidate)
+
+    lang(candidate) == lang(target)
+  end
+
   def self.shared_tags(a, b)
     normalise(a[:tags]) & normalise(b[:tags])
   end
@@ -99,10 +134,17 @@ def self.sort_time(post)
       }
     end
     by_url = docs.to_h { |d| [d.url, d] }
+    # Sorted here rather than trusting the collection's order, and by id as well
+    # as date because the three translation pairs share a date exactly.
+    ordered = meta.sort_by { |m| [RelatedPosts.sort_time(m), m[:id].to_s] }
 
     meta.each_with_index do |target, i|
       ids = RelatedPosts.rank(target, meta)
       docs[i].data["related"] = ids.map { |id| by_url[id] }
+
+      older, newer = RelatedPosts.adjacent(target, ordered)
+      docs[i].data["adjacent_older"] = older && by_url[older[:id]]
+      docs[i].data["adjacent_newer"] = newer && by_url[newer[:id]]
     end
   end
 end
diff --git a/_posts/2025-01-17-from-local-to-global-a-graph-rag-approach-to-query-focused-summarization.md b/_posts/2025-01-17-from-local-to-global-a-graph-rag-approach-to-query-focused-summarization.md
index 65cf2aa..02188d1 100644
--- a/_posts/2025-01-17-from-local-to-global-a-graph-rag-approach-to-query-focused-summarization.md
+++ b/_posts/2025-01-17-from-local-to-global-a-graph-rag-approach-to-query-focused-summarization.md
@@ -4,7 +4,7 @@ title: "From Local to Global: A Graph RAG Approach to Query-Focused Summarizatio
 date: 2024-04-24 18:38:11
 author: "Microsoft Research"
 categories: ["Paper Reviews", "Retrieval-Augmented-Generation"]
-tags: ["Graph-RAG", "Query-Focused-Summarization", "Community-Based-Summarization", "Entity-Knowledge-Graph-Extraction", "Map-Reduce-Summarization", "Modularity-Based-Retrieval", "Community-Detection-For-RAG", "Leiden-Algorithm-For-Text", "Multi-Level-Community-Indexing", "Global-Context-Understanding"]
+tags: ["Graph-RAG", "Query-Focused-Summarization", "Community-Based-Summarization", "Entity-Knowledge-Graph-Extraction", "Map-Reduce-Summarization", "Modularity-Based-Retrieval", "Community-Detection-For-RAG", "Leiden-Algorithm-For-Text", "Multi-Level-Community-Indexing", "Global-Context-Understanding", "Retrieval-Augmented-Generation", "Knowledge-Graph"]
 cover: /assets/images/retrieval-augmented-generation.jpg
 use_math: true
 ---
diff --git a/_posts/2025-01-27-llama-2-open-foundation-and-fine-tuned-chat-models.md b/_posts/2025-01-27-llama-2-open-foundation-and-fine-tuned-chat-models.md
index fb573bc..5ac5fb8 100644
--- a/_posts/2025-01-27-llama-2-open-foundation-and-fine-tuned-chat-models.md
+++ b/_posts/2025-01-27-llama-2-open-foundation-and-fine-tuned-chat-models.md
@@ -4,9 +4,7 @@ title: "Llama 2: Open Foundation and Fine-Tuned Chat Models"
 date: 2023-07-18 14:31:57
 author: "Meta AI"
 categories: ["Paper Reviews", "Language-Models"]
-tags: [
-    "Open-Foundation-and-Fine-Tuned-Chat-Models", "Reinforcement-Learning-with-Human-Feedback", "Ghost-Attention", "Rejection-Sampling-Fine-Tuning", "Grouped-Query-Attention", "Proximal-Policy-Optimization", "Red-Team-Safety-Testing", "Safety-Alignment", "System-Message-for-Multi-Turn-Consistency", "Safety-Context-Distillation"
-]
+tags: ["Open-Foundation-and-Fine-Tuned-Chat-Models", "Reinforcement-Learning-with-Human-Feedback", "Ghost-Attention", "Rejection-Sampling-Fine-Tuning", "Grouped-Query-Attention", "Proximal-Policy-Optimization", "Red-Team-Safety-Testing", "Safety-Alignment", "System-Message-for-Multi-Turn-Consistency", "Safety-Context-Distillation", "Alignment", "Llama"]
 cover: /assets/images/language-models.jpg
 use_math: true
 ---
diff --git a/_posts/2025-01-28-llama-open-and-efficient-foundation-language-models.md b/_posts/2025-01-28-llama-open-and-efficient-foundation-language-models.md
index 88bdfec..4d8b69b 100644
--- a/_posts/2025-01-28-llama-open-and-efficient-foundation-language-models.md
+++ b/_posts/2025-01-28-llama-open-and-efficient-foundation-language-models.md
@@ -4,9 +4,7 @@ title: "LLaMA: Open and Efficient Foundation Language Models"
 date: 2023-02-27 17:11:15
 author: "Meta AI"
 categories: ["Paper Reviews", "Language-Models"]
-tags: [
-    "Open-and-Efficient-Foundation-Language-Models", "Rotary-Positional-Embeddings", "SwiGLU-Activation-Function", "Scaling-Laws-for-Large-Language-Models", "Economical-Training", "Computation-Communication-Overlap", "Public-Dataset-Only Training", "RMSNorm-Pre-Normalization", "Memory-Efficient-Attention", "CommonCrawl-Quality-Filtering"
-]
+tags: ["Open-and-Efficient-Foundation-Language-Models", "Rotary-Positional-Embeddings", "SwiGLU-Activation-Function", "Scaling-Laws-for-Large-Language-Models", "Economical-Training", "Computation-Communication-Overlap", "Public-Dataset-Only Training", "RMSNorm-Pre-Normalization", "Memory-Efficient-Attention", "CommonCrawl-Quality-Filtering", "Llama"]
 cover: /assets/images/language-models.jpg
 use_math: true
 ---
diff --git a/_posts/2025-02-03-deepseek-llm-scaling-open-source-language-models-with-longtermism.md b/_posts/2025-02-03-deepseek-llm-scaling-open-source-language-models-with-longtermism.md
index d37c6af..6acb9c6 100644
--- a/_posts/2025-02-03-deepseek-llm-scaling-open-source-language-models-with-longtermism.md
+++ b/_posts/2025-02-03-deepseek-llm-scaling-open-source-language-models-with-longtermism.md
@@ -4,7 +4,7 @@ title: "DeepSeek LLM: Scaling Open-Source Language Models with Longtermism"
 date: 2024-01-05 18:59:13
 author: "DeepSeek AI"
 categories: ["Paper Reviews", "Language-Models"]
-tags: ["Scaling-Open-Source-Language-Models-with-Longtermism", "Compute-Optimal-Scaling", "Non-Embedding-FLOPs-per-Token", "Multi-Step-Learning-Rate-Scheduler", "Optimal-Model-Data-Scaling-Allocation", "Direct-Preference-Optimization", "Grouped-Query-Attention", "Language-Model-Scaling-Laws", "Efficient-Model-Training", "Depth-Over-Width-Scaling"]
+tags: ["Scaling-Open-Source-Language-Models-with-Longtermism", "Compute-Optimal-Scaling", "Non-Embedding-FLOPs-per-Token", "Multi-Step-Learning-Rate-Scheduler", "Optimal-Model-Data-Scaling-Allocation", "Direct-Preference-Optimization", "Grouped-Query-Attention", "Language-Model-Scaling-Laws", "Efficient-Model-Training", "Depth-Over-Width-Scaling", "DeepSeek"]
 cover: /assets/images/language-models.jpg
 use_math: true
 ---
diff --git a/_posts/2025-02-05-deepseekmoe-towards-ultimate-expert-specialization-in-mixture-of-experts-language-models.md b/_posts/2025-02-05-deepseekmoe-towards-ultimate-expert-specialization-in-mixture-of-experts-language-models.md
index 82bf713..2e7fbd3 100644
--- a/_posts/2025-02-05-deepseekmoe-towards-ultimate-expert-specialization-in-mixture-of-experts-language-models.md
+++ b/_posts/2025-02-05-deepseekmoe-towards-ultimate-expert-specialization-in-mixture-of-experts-language-models.md
@@ -4,7 +4,7 @@ title: "DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Expert
 date: 2024-01-11 17:31:42
 author: "DeepSeek AI"
 categories: ["Paper Reviews", "Language-Models"]
-tags: ["Ultimate-Expert-Specialization", "Fine-Grained-Expert-Segmentation", "Shared-Expert-Isolation", "Mixture-of-Experts-Architecture", "Parameter-Efficient-Language-Model-Scaling", "Sparse-Mixture-of-Experts", "Efficient-Transformer-Architecture", "Economical-Training", "Multi-Level-Load-Balancing", "Efficient-Inference"]
+tags: ["Ultimate-Expert-Specialization", "Fine-Grained-Expert-Segmentation", "Shared-Expert-Isolation", "Mixture-of-Experts-Architecture", "Parameter-Efficient-Language-Model-Scaling", "Sparse-Mixture-of-Experts", "Efficient-Transformer-Architecture", "Economical-Training", "Multi-Level-Load-Balancing", "Efficient-Inference", "Mixture-of-Experts", "DeepSeek"]
 cover: /assets/images/language-models.jpg
 use_math: true
 ---
diff --git a/_posts/2025-02-14-deepseek-coder-when-the-large-language-model-meets-programming-the-rise-of-code-intelligence.md b/_posts/2025-02-14-deepseek-coder-when-the-large-language-model-meets-programming-the-rise-of-code-intelligence.md
index cc580e4..1729a34 100644
--- a/_posts/2025-02-14-deepseek-coder-when-the-large-language-model-meets-programming-the-rise-of-code-intelligence.md
+++ b/_posts/2025-02-14-deepseek-coder-when-the-large-language-model-meets-programming-the-rise-of-code-intelligence.md
@@ -4,7 +4,7 @@ title: "DeepSeek-Coder: When the Large Language Model Meets Programming - The Ri
 date: 2024-01-25 14:17:53
 author: "DeepSeek AI"
 categories: ["Paper Reviews", "Language-Models"]
-tags: ["Repository-Level-Data-Construction", "Fill-in-the-Middle-Code-Completion", "Cross-File-Code-Completion", "Long-Context-Adaptation", "Repository-Level-Deduplication", "Dependency-Aware-Training", "Project-Level-Code-Understanding", "Cross-File-Dependency-Analysis", "Topological-Sorting-For-Code-Learning", "LeetCode-Competition-Benchmark"]
+tags: ["Repository-Level-Data-Construction", "Fill-in-the-Middle-Code-Completion", "Cross-File-Code-Completion", "Long-Context-Adaptation", "Repository-Level-Deduplication", "Dependency-Aware-Training", "Project-Level-Code-Understanding", "Cross-File-Dependency-Analysis", "Topological-Sorting-For-Code-Learning", "LeetCode-Competition-Benchmark", "DeepSeek"]
 cover: /assets/images/language-models.jpg
 use_math: true
 ---
diff --git a/_posts/2025-02-14-raptor-recursive-abstractive-processing-for-tree-organized-retrieval.md b/_posts/2025-02-14-raptor-recursive-abstractive-processing-for-tree-organized-retrieval.md
index 0f92da2..1738abd 100644
--- a/_posts/2025-02-14-raptor-recursive-abstractive-processing-for-tree-organized-retrieval.md
+++ b/_posts/2025-02-14-raptor-recursive-abstractive-processing-for-tree-organized-retrieval.md
@@ -4,7 +4,7 @@ title: "RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval"
 date: 2024-01-31 18:30:21
 author: "Stanford University"
 categories: ["Paper Reviews", "Retrieval-Augmented-Generation"]
-tags: ["Recursive-Abstractive-Processing", "Tree-Organized-Retrieval", "Hierarchical-Summarization", "Collapsed-Tree-Retrieval", "Dependency-Aware-Tree-Traversal", "Multi-Level-Abstraction-Retrieval", "Gaussian-Mixture-Model-Clustering", "Semantic-Similarity-Clustering", "Recursive-Summarization", "Contextual-Flexibility-Retrieval"]
+tags: ["Recursive-Abstractive-Processing", "Tree-Organized-Retrieval", "Hierarchical-Summarization", "Collapsed-Tree-Retrieval", "Dependency-Aware-Tree-Traversal", "Multi-Level-Abstraction-Retrieval", "Gaussian-Mixture-Model-Clustering", "Semantic-Similarity-Clustering", "Recursive-Summarization", "Contextual-Flexibility-Retrieval", "Retrieval-Augmented-Generation"]
 cover: /assets/images/retrieval-augmented-generation.jpg
 use_math: true
 ---
diff --git a/_posts/2025-02-28-deepseekmath-pushing-the-limits-of-mathematical-reasoning-in-open-language-models.md b/_posts/2025-02-28-deepseekmath-pushing-the-limits-of-mathematical-reasoning-in-open-language-models.md
index b321a65..7bf6ded 100644
--- a/_posts/2025-02-28-deepseekmath-pushing-the-limits-of-mathematical-reasoning-in-open-language-models.md
+++ b/_posts/2025-02-28-deepseekmath-pushing-the-limits-of-mathematical-reasoning-in-open-language-models.md
@@ -4,7 +4,7 @@ title: "DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Langu
 date: 2024-02-05 18:55:32
 author: "DeepSeek AI"
 categories: ["Paper Reviews", "Language-Models"]
-tags: ["DeepSeekMath-Corpus", "Group-Relative-Policy-Optimization", "Iterative-Reinforcement-Learning", "Code-Training-Benefits-Mathematical-Reasoning", "Supervised-Fine-Tuning", "Rejection-Sampling-Fine-Tuning", "Direct-Preference-Optimization"]
+tags: ["DeepSeekMath-Corpus", "Group-Relative-Policy-Optimization", "Iterative-Reinforcement-Learning", "Code-Training-Benefits-Mathematical-Reasoning", "Supervised-Fine-Tuning", "Rejection-Sampling-Fine-Tuning", "Direct-Preference-Optimization", "Reasoning-Models", "DeepSeek"]
 cover: /assets/images/language-models.jpg
 use_math: true
 ---
diff --git a/_posts/2025-04-04-mixtral-of-experts.md b/_posts/2025-04-04-mixtral-of-experts.md
index c35bcd3..8075b7b 100644
--- a/_posts/2025-04-04-mixtral-of-experts.md
+++ b/_posts/2025-04-04-mixtral-of-experts.md
@@ -5,7 +5,7 @@ subtitle: "Mixtral of Experts: 전문가 혼합 모델의 새로운 지평"
 date: 2024-01-08 18:47:34
 author: "Mistral AI"
 categories: ["Paper Reviews", "Language-Models"]
-tags: ["Sparse-Mixture-of-Experts", "Dynamic-Expert-Routing", "Two-Expert-Token-Processing", "Efficient-Inference-with-Reduced-Active-Parameters", "Routing-Network-Token-Selection", "Expert-Selection-Locality-Analysis", "Multilingual-Performance-Scaling", "Long-Context-Retrieval-Optimization", "Instruction-Fine-Tuning-with-Direct-Preference-Optimization"]
+tags: ["Sparse-Mixture-of-Experts", "Dynamic-Expert-Routing", "Two-Expert-Token-Processing", "Efficient-Inference-with-Reduced-Active-Parameters", "Routing-Network-Token-Selection", "Expert-Selection-Locality-Analysis", "Multilingual-Performance-Scaling", "Long-Context-Retrieval-Optimization", "Instruction-Fine-Tuning-with-Direct-Preference-Optimization", "Mixture-of-Experts"]
 cover: /assets/images/language-models.jpg
 use_math: true
 ---
diff --git a/_posts/2025-10-03-gemma-3-technical-report.md b/_posts/2025-10-03-gemma-3-technical-report.md
index 11f47ee..cbb950b 100644
--- a/_posts/2025-10-03-gemma-3-technical-report.md
+++ b/_posts/2025-10-03-gemma-3-technical-report.md
@@ -4,7 +4,7 @@ title: "Gemma 3 Technical Report"
 date: 2025-03-25 15:52:34
 author: "Google DeepMind"
 categories: ["Paper Reviews", "Multimodal-Learning"]
-tags: ["Alternating-Local-Global-Attention", "Long-Context-Adaptation", "Pan-&-Scan-Image-Processing", "Vision-Encoder-Token-Condensation", "Multimodal-Knowledge-Distillation", "RoPE-Positional-Embedding-Extension", "Grouped-Query-Attention", "Quantization-Aware-Training", "Sliding-Window-Attention-Optimization", "Efficient-Long-Context-Attention-Mechanism"]
+tags: ["Alternating-Local-Global-Attention", "Long-Context-Adaptation", "Pan-&-Scan-Image-Processing", "Vision-Encoder-Token-Condensation", "Multimodal-Knowledge-Distillation", "RoPE-Positional-Embedding-Extension", "Grouped-Query-Attention", "Quantization-Aware-Training", "Sliding-Window-Attention-Optimization", "Efficient-Long-Context-Attention-Mechanism", "Multimodal-Models"]
 cover: /assets/images/multimodal-learning.jpg
 use_math: true
 ---
diff --git a/_posts/2025-10-03-kimi-k2--open-agentic-intelligence.md b/_posts/2025-10-03-kimi-k2--open-agentic-intelligence.md
index 4c8c03f..e6fadfb 100644
--- a/_posts/2025-10-03-kimi-k2--open-agentic-intelligence.md
+++ b/_posts/2025-10-03-kimi-k2--open-agentic-intelligence.md
@@ -5,7 +5,7 @@ subtitle: "Kimi K2: 오픈 에이전틱 인텔리전스"
 date: 2025-07-28 05:35:43
 author: "Moonshot AI"
 categories: ["Paper Reviews", "Language-Models"]
-tags: ["MuonClip-Optimizer", "QK-Clip-Attention-Stabilization", "Large-Scale-Agentic-Data-Synthesis", "Multi-Stage-Reinforcement-Learning-with-Self-Critique", "Mixture-of-Experts-Sparsity-Scaling-Law", "Synthetic-Data-Rephrasing-for-Token-Efficiency", "Verifiable-Rewards-Reinforcement-Learning", "Agentic-Intelligence-Framework", "Computational-Efficiency-in-Large-Language-Models", "Efficient-Mixture-of-Experts-Architecture"]
+tags: ["MuonClip-Optimizer", "QK-Clip-Attention-Stabilization", "Large-Scale-Agentic-Data-Synthesis", "Multi-Stage-Reinforcement-Learning-with-Self-Critique", "Mixture-of-Experts-Sparsity-Scaling-Law", "Synthetic-Data-Rephrasing-for-Token-Efficiency", "Verifiable-Rewards-Reinforcement-Learning", "Agentic-Intelligence-Framework", "Computational-Efficiency-in-Large-Language-Models", "Efficient-Mixture-of-Experts-Architecture", "Mixture-of-Experts"]
 cover: /assets/images/language-models.jpg
 use_math: true
 ---
diff --git a/_posts/2025-10-05-deepseek-v2--a-strong--economical--and-efficient-mixture-of-experts-language-model.md b/_posts/2025-10-05-deepseek-v2--a-strong--economical--and-efficient-mixture-of-experts-language-model.md
index 1c90f6f..27b222e 100644
--- a/_posts/2025-10-05-deepseek-v2--a-strong--economical--and-efficient-mixture-of-experts-language-model.md
+++ b/_posts/2025-10-05-deepseek-v2--a-strong--economical--and-efficient-mixture-of-experts-language-model.md
@@ -5,7 +5,7 @@ subtitle: "DeepSeek-V2: 강력하고 경제적이며 효율적인 Mixture-of-Exp
 date: 2024-05-07 15:56:43
 author: "DeepSeek AI"
 categories: ["Paper Reviews", "Language-Models"]
-tags: ["Multi-Head-Latent-Attention", "DeepSeekMoE", "Low-Rank-Key-Value-Joint-Compression", "Decoupled-Rotary-Position-Embedding", "Device-Limited-Routing", "Auxiliary-Loss-for-Load-Balance", "Token-Dropping-Strategy", "Sparse-Mixture-of-Experts", "Expert-Parallelism", "Efficient-Long-Context-Attention-Mechanism"]
+tags: ["Multi-Head-Latent-Attention", "DeepSeekMoE", "Low-Rank-Key-Value-Joint-Compression", "Decoupled-Rotary-Position-Embedding", "Device-Limited-Routing", "Auxiliary-Loss-for-Load-Balance", "Token-Dropping-Strategy", "Sparse-Mixture-of-Experts", "Expert-Parallelism", "Efficient-Long-Context-Attention-Mechanism", "Mixture-of-Experts", "DeepSeek"]
 cover: /assets/images/language-models.jpg
 use_math: true
 ---
diff --git a/_posts/2025-10-06-tulu-3--pushing-frontiers-in-open-language-model-post-training.md b/_posts/2025-10-06-tulu-3--pushing-frontiers-in-open-language-model-post-training.md
index cfcfaef..0208b4d 100644
--- a/_posts/2025-10-06-tulu-3--pushing-frontiers-in-open-language-model-post-training.md
+++ b/_posts/2025-10-06-tulu-3--pushing-frontiers-in-open-language-model-post-training.md
@@ -5,7 +5,7 @@ subtitle: "Tülu 3: 오픈 언어 모델 사후 훈련의 새로운 지평"
 date: 2024-11-22 18:44:04
 author: "Allen Institute for AI"
 categories: ["Paper Reviews", "Language-Models"]
-tags: ["Reinforcement-Learning-with-Verifiable-Rewards", "Multi-Stage-Post-Training-Recipe", "Direct-Preference-Optimization", "Persona-Driven-Data-Synthesis", "Skill-Specific-Synthetic-Data-Generation", "Prompt-Decontamination", "Open-Language-Model-Evaluation-System", "Length-Normalized-Preference-Optimization", "Asynchronous-Reinforcement-Learning-Infrastructure", "Skill-Targeted-Model-Training"]
+tags: ["Reinforcement-Learning-with-Verifiable-Rewards", "Multi-Stage-Post-Training-Recipe", "Direct-Preference-Optimization", "Persona-Driven-Data-Synthesis", "Skill-Specific-Synthetic-Data-Generation", "Prompt-Decontamination", "Open-Language-Model-Evaluation-System", "Length-Normalized-Preference-Optimization", "Asynchronous-Reinforcement-Learning-Infrastructure", "Skill-Targeted-Model-Training", "Alignment"]
 cover: /assets/images/language-models.jpg
 use_math: true
 ---
diff --git a/_posts/2025-10-07-deepseek-r1--incentivizing-reasoning-capability-in-llms-via-reinforcement-learning.md b/_posts/2025-10-07-deepseek-r1--incentivizing-reasoning-capability-in-llms-via-reinforcement-learning.md
index 6b06109..41e0a2b 100644
--- a/_posts/2025-10-07-deepseek-r1--incentivizing-reasoning-capability-in-llms-via-reinforcement-learning.md
+++ b/_posts/2025-10-07-deepseek-r1--incentivizing-reasoning-capability-in-llms-via-reinforcement-learning.md
@@ -5,7 +5,7 @@ subtitle: "DeepSeek-R1: 강화학습을 통한 대규모 언어 모델의 추론
 date: 2025-01-22 15:19:35
 author: "DeepSeek AI"
 categories: ["Paper Reviews", "Language-Models"]
-tags: ["Large-Scale-Reinforcement-Learning-on-Base-Model", "Group-Relative-Policy-Optimization", "Reasoning-Oriented-Reinforcement-Learning", "Reinforcement-Learning-with-Cold-Start", "Distillation-of-Reasoning-Capability", "Multi-Stage-Reinforcement-Learning-with-Self-Critique", "Verifiable-Rewards-Reinforcement-Learning", "Rejection-Sampling-and-Supervised-Fine-Tuning", "Iterative-Reinforcement-Learning", "Unified-Paradigm-for-Reinforcement-Learning"]
+tags: ["Large-Scale-Reinforcement-Learning-on-Base-Model", "Group-Relative-Policy-Optimization", "Reasoning-Oriented-Reinforcement-Learning", "Reinforcement-Learning-with-Cold-Start", "Distillation-of-Reasoning-Capability", "Multi-Stage-Reinforcement-Learning-with-Self-Critique", "Verifiable-Rewards-Reinforcement-Learning", "Rejection-Sampling-and-Supervised-Fine-Tuning", "Iterative-Reinforcement-Learning", "Unified-Paradigm-for-Reinforcement-Learning", "Reasoning-Models", "Alignment", "DeepSeek"]
 cover: /assets/images/language-models.jpg
 use_math: true
 ---
diff --git a/_posts/2025-10-07-deepseek-v3-technical-report.md b/_posts/2025-10-07-deepseek-v3-technical-report.md
index 03a2896..388a820 100644
--- a/_posts/2025-10-07-deepseek-v3-technical-report.md
+++ b/_posts/2025-10-07-deepseek-v3-technical-report.md
@@ -5,7 +5,7 @@ subtitle: "DeepSeek-V3 기술 보고서"
 date: 2024-12-27 04:03:16
 author: "DeepSeek AI"
 categories: ["Paper Reviews", "Language-Models"]
-tags: ["Auxiliary-Loss-Free-Load-Balancing", "Multi-Token-Prediction", "Multi-Head-Latent-Attention", "DeepSeekMoE-Architecture", "FP8-Mixed-Precision-Training", "Efficient-Cross-Node-All-to-All-Communication", "Node-Limited-Routing", "Computation-Communication-Overlap", "Tile-Wise-Fine-Grained-Quantization", "Speculative-Decoding"]
+tags: ["Auxiliary-Loss-Free-Load-Balancing", "Multi-Token-Prediction", "Multi-Head-Latent-Attention", "DeepSeekMoE-Architecture", "FP8-Mixed-Precision-Training", "Efficient-Cross-Node-All-to-All-Communication", "Node-Limited-Routing", "Computation-Communication-Overlap", "Tile-Wise-Fine-Grained-Quantization", "Speculative-Decoding", "Mixture-of-Experts", "DeepSeek"]
 cover: /assets/images/language-models.jpg
 use_math: true
 ---
diff --git a/_posts/2025-10-07-the-llama-3-herd-of-models.md b/_posts/2025-10-07-the-llama-3-herd-of-models.md
index 3deddb6..ef529b5 100644
--- a/_posts/2025-10-07-the-llama-3-herd-of-models.md
+++ b/_posts/2025-10-07-the-llama-3-herd-of-models.md
@@ -4,7 +4,7 @@ title: "The Llama 3 Herd of Models"
 date: 2024-07-31 17:54:27
 author: "Meta AI"
 categories: ["Paper Reviews", "Language-Models"]
-tags: ["Direct-Preference-Optimization", "Scaling-Laws-for-Large-Language-Models", "Multimodal-Knowledge-Distillation", "Efficient-Long-Context-Attention-Mechanism", "Multilingual-Multimodal-Understanding", "Safety-Alignment", "Synthetic-Data-Generation-for-Mathematics", "Tool-Use-Emergence", "Rejection-Sampling-Fine-Tuning", "Multi-Stage-Post-Training-Recipe"]
+tags: ["Direct-Preference-Optimization", "Scaling-Laws-for-Large-Language-Models", "Multimodal-Knowledge-Distillation", "Efficient-Long-Context-Attention-Mechanism", "Multilingual-Multimodal-Understanding", "Safety-Alignment", "Synthetic-Data-Generation-for-Mathematics", "Tool-Use-Emergence", "Rejection-Sampling-Fine-Tuning", "Multi-Stage-Post-Training-Recipe", "Llama"]
 cover: /assets/images/language-models.jpg
 use_math: true
 ---
diff --git a/_posts/2025-10-08-qwen3-technical-report.md b/_posts/2025-10-08-qwen3-technical-report.md
index 7e635c6..714fb77 100644
--- a/_posts/2025-10-08-qwen3-technical-report.md
+++ b/_posts/2025-10-08-qwen3-technical-report.md
@@ -5,7 +5,7 @@ subtitle: "Qwen3 기술 보고서"
 date: 2025-05-14 13:41:34
 author: "Alibaba Group"
 categories: ["Paper Reviews", "Language-Models"]
-tags: ["Thinking-Budget-Mechanism", "Strong-to-Weak-Distillation", "Dynamic-Mode-Switching", "Mixture-of-Experts-Architecture", "Fine-Grained-Expert-Segmentation", "Global-Batch-Load-Balancing-Loss", "Long-Chain-of-Thought-Cold-Start", "Reasoning-Reinforcement-Learning", "Thinking-Mode-Fusion", "Multi-Stage-Post-Training-Recipe"]
+tags: ["Thinking-Budget-Mechanism", "Strong-to-Weak-Distillation", "Dynamic-Mode-Switching", "Mixture-of-Experts-Architecture", "Fine-Grained-Expert-Segmentation", "Global-Batch-Load-Balancing-Loss", "Long-Chain-of-Thought-Cold-Start", "Reasoning-Reinforcement-Learning", "Thinking-Mode-Fusion", "Multi-Stage-Post-Training-Recipe", "Mixture-of-Experts", "Reasoning-Models"]
 cover: /assets/images/language-models.jpg
 use_math: true
 ---
diff --git a/_posts/2025-10-10-pixtral-12b.md b/_posts/2025-10-10-pixtral-12b.md
index 9727028..8a4f0dc 100644
--- a/_posts/2025-10-10-pixtral-12b.md
+++ b/_posts/2025-10-10-pixtral-12b.md
@@ -4,7 +4,7 @@ title: "Pixtral 12B"
 date: 2024-10-09 17:16:22
 author: "Mistral AI"
 categories: ["Paper Reviews", "Multimodal-Learning"]
-tags: ["RoPE-2D-Positional-Encoding", "Block-Diagonal-Attention-Masking", "Flexible-Vision-Encoder-Architecture", "Natively-Multimodal-Transformer-Architecture", "Multi-Turn-Instruction-Tuning", "Sequence-Packing-Optimization", "Variable-Image-Resolution-Processing", "Break-Tokens-for-Image-Tokenization", "Standardized-Multimodal-Evaluation", "MM-MT-Bench-Benchmark"]
+tags: ["RoPE-2D-Positional-Encoding", "Block-Diagonal-Attention-Masking", "Flexible-Vision-Encoder-Architecture", "Natively-Multimodal-Transformer-Architecture", "Multi-Turn-Instruction-Tuning", "Sequence-Packing-Optimization", "Variable-Image-Resolution-Processing", "Break-Tokens-for-Image-Tokenization", "Standardized-Multimodal-Evaluation", "MM-MT-Bench-Benchmark", "Multimodal-Models"]
 cover: /assets/images/multimodal-learning.jpg
 use_math: true
 ---
diff --git a/_posts/2025-10-11-gemini--a-family-of-highly-capable-multimodal-models.md b/_posts/2025-10-11-gemini--a-family-of-highly-capable-multimodal-models.md
index 18e583e..96e7448 100644
--- a/_posts/2025-10-11-gemini--a-family-of-highly-capable-multimodal-models.md
+++ b/_posts/2025-10-11-gemini--a-family-of-highly-capable-multimodal-models.md
@@ -5,7 +5,7 @@ subtitle: "Gemini: 고도로 능력 있는 멀티모달 모델 패밀리"
 date: 2023-12-19 02:39:27
 author: "Google DeepMind"
 categories: ["Paper Reviews", "Multimodal-Learning"]
-tags: ["Natively-Multimodal-Transformer-Architecture", "Multimodal-Reasoning-with-Uncertainty-Routing", "Joint-Multimodal-Pre-Training", "Cross-Modal-Reasoning-Capabilities", "Efficient-Long-Context-Attention-Mechanism", "Multimodal-Safety-Evaluation-Framework", "Responsible-Multimodal-Model-Development", "Interleaved-Sequence-Processing", "Variable-Image-Resolution-Processing", "Unified-Multi-Modal-Generative-Model"]
+tags: ["Natively-Multimodal-Transformer-Architecture", "Multimodal-Reasoning-with-Uncertainty-Routing", "Joint-Multimodal-Pre-Training", "Cross-Modal-Reasoning-Capabilities", "Efficient-Long-Context-Attention-Mechanism", "Multimodal-Safety-Evaluation-Framework", "Responsible-Multimodal-Model-Development", "Interleaved-Sequence-Processing", "Variable-Image-Resolution-Processing", "Unified-Multi-Modal-Generative-Model", "Multimodal-Models"]
 cover: /assets/images/multimodal-learning.jpg
 use_math: true
 ---
diff --git a/_posts/2025-10-11-gemini-2.5--pushing-the-frontier-with-advanced-reasoning--multimodality--long-context--and-next-generation-agentic-capabilities.md b/_posts/2025-10-11-gemini-2.5--pushing-the-frontier-with-advanced-reasoning--multimodality--long-context--and-next-generation-agentic-capabilities.md
index 2a0c059..a8741c0 100644
--- a/_posts/2025-10-11-gemini-2.5--pushing-the-frontier-with-advanced-reasoning--multimodality--long-context--and-next-generation-agentic-capabilities.md
+++ b/_posts/2025-10-11-gemini-2.5--pushing-the-frontier-with-advanced-reasoning--multimodality--long-context--and-next-generation-agentic-capabilities.md
@@ -5,7 +5,7 @@ subtitle: "Gemini 2.5: 첨단 추론, 멀티모달리티, 긴 컨텍스트, 차
 date: 2025-07-07 17:36:04
 author: "Google DeepMind"
 categories: ["Paper Reviews", "Multimodal-Learning"]
-tags: ["Sparse-Mixture-of-Experts-Architecture", "Long-Context-Reasoning", "Dynamic-Thinking-Budget", "Multimodal-Tool-Use", "Natively-Multimodal-Transformer", "Reinforcement-Learning-from-Human-and-Critic-Feedback", "Agentic-Workflow-Generation", "Thinking-Mode-Fusion", "Advanced-Reasoning-Capabilities", "Automated-Red-Teaming"]
+tags: ["Sparse-Mixture-of-Experts-Architecture", "Long-Context-Reasoning", "Dynamic-Thinking-Budget", "Multimodal-Tool-Use", "Natively-Multimodal-Transformer", "Reinforcement-Learning-from-Human-and-Critic-Feedback", "Agentic-Workflow-Generation", "Thinking-Mode-Fusion", "Advanced-Reasoning-Capabilities", "Automated-Red-Teaming", "Reasoning-Models", "Multimodal-Models"]
 cover: /assets/images/multimodal-learning.jpg
 use_math: true
 ---
diff --git a/_posts/2025-10-19-direct-preference-optimization--your-language-model-is-secretly-a-reward-model.md b/_posts/2025-10-19-direct-preference-optimization--your-language-model-is-secretly-a-reward-model.md
index 7056093..95ea5b8 100644
--- a/_posts/2025-10-19-direct-preference-optimization--your-language-model-is-secretly-a-reward-model.md
+++ b/_posts/2025-10-19-direct-preference-optimization--your-language-model-is-secretly-a-reward-model.md
@@ -4,7 +4,7 @@ title: "Direct Preference Optimization: Your Language Model is Secretly a Reward
 date: 2023-05-29 17:57:46
 author: "Stanford University"
 categories: ["Paper Reviews", "Finetuning"]
-tags: ["Direct-Preference-Optimization", "Implicit-Reward-Modeling", "Bradley-Terry-Preference-Framework", "KL-Constrained-Reward-Maximization", "Binary-Cross-Entropy-Policy-Optimization", "Reward-Reparameterization", "Policy-Implicit-Reward-Function", "Closed-Form-Optimal-Policy-Extraction", "RL-Free-Preference-Learning", "Dynamic-Per-Example-Importance-Weighting"]
+tags: ["Direct-Preference-Optimization", "Implicit-Reward-Modeling", "Bradley-Terry-Preference-Framework", "KL-Constrained-Reward-Maximization", "Binary-Cross-Entropy-Policy-Optimization", "Reward-Reparameterization", "Policy-Implicit-Reward-Function", "Closed-Form-Optimal-Policy-Extraction", "RL-Free-Preference-Learning", "Dynamic-Per-Example-Importance-Weighting", "Alignment"]
 cover: /assets/images/finetuning.jpg
 use_math: true
 ---
diff --git a/_posts/2026-01-12-deepseek-v3.2--pushing-the-frontier-of-open-large-language-models.md b/_posts/2026-01-12-deepseek-v3.2--pushing-the-frontier-of-open-large-language-models.md
index 1e62185..b761ccb 100644
--- a/_posts/2026-01-12-deepseek-v3.2--pushing-the-frontier-of-open-large-language-models.md
+++ b/_posts/2026-01-12-deepseek-v3.2--pushing-the-frontier-of-open-large-language-models.md
@@ -5,7 +5,7 @@ subtitle: "DeepSeek-V3.2: 오픈 대규모 언어 모델의 새로운 지평"
 date: 2025-12-02 09:25:14
 author: "DeepSeek AI"
 categories: ["Paper Reviews", "Language-Models"]
-tags: ["DeepSeek-Sparse-Attention", "Scalable-Reinforcement-Learning-Framework", "Large-Scale-Agentic-Task-Synthesis-Pipeline", "Group-Relative-Policy-Optimization-Scaling", "Unbiased-KL-Estimate-for-RL", "Off-Policy-Sequence-Masking", "Keep-Routing-for-Mixture-of-Experts", "Thinking-Context-Management-for-Tool-Use", "Long-Chain-of-Thought-Cold-Start-Integration", "Multi-Stage-Agentic-Environment-Synthesis"]
+tags: ["DeepSeek-Sparse-Attention", "Scalable-Reinforcement-Learning-Framework", "Large-Scale-Agentic-Task-Synthesis-Pipeline", "Group-Relative-Policy-Optimization-Scaling", "Unbiased-KL-Estimate-for-RL", "Off-Policy-Sequence-Masking", "Keep-Routing-for-Mixture-of-Experts", "Thinking-Context-Management-for-Tool-Use", "Long-Chain-of-Thought-Cold-Start-Integration", "Multi-Stage-Agentic-Environment-Synthesis", "Mixture-of-Experts", "Reasoning-Models", "DeepSeek"]
 cover: /assets/images/language-models.jpg
 use_math: true
 ---
diff --git a/_posts/2026-01-16-constitutional-ai--harmlessness-from-ai-feedback.md b/_posts/2026-01-16-constitutional-ai--harmlessness-from-ai-feedback.md
index cfe0804..e6534a7 100644
--- a/_posts/2026-01-16-constitutional-ai--harmlessness-from-ai-feedback.md
+++ b/_posts/2026-01-16-constitutional-ai--harmlessness-from-ai-feedback.md
@@ -5,7 +5,7 @@ subtitle: "Constitutional AI: AI 피드백을 통한 무해성 구현"
 date: 2022-12-15 06:19:23
 author: "Anthropic"
 categories: ["Paper Reviews", "Finetuning"]
-tags: ["Constitutional-AI", "Reinforcement-Learning-from-AI-Feedback", "Self-Critique-and-Revision-Pipeline", "Chain-of-Thought-Reasoning-for-AI-Evaluation", "AI-Generated-Preference-Labels", "Scaling-Supervision", "Non-Evasive-Harmlessness-Training", "Model-Based-Feedback-Generation", "Constitutional-Principles-for-Behavior-Steering", "Preference-Model-Training-with-AI-Feedback"]
+tags: ["Constitutional-AI", "Reinforcement-Learning-from-AI-Feedback", "Self-Critique-and-Revision-Pipeline", "Chain-of-Thought-Reasoning-for-AI-Evaluation", "AI-Generated-Preference-Labels", "Scaling-Supervision", "Non-Evasive-Harmlessness-Training", "Model-Based-Feedback-Generation", "Constitutional-Principles-for-Behavior-Steering", "Preference-Model-Training-with-AI-Feedback", "Alignment"]
 cover: /assets/images/finetuning.jpg
 use_math: true
 ---
diff --git a/_posts/2026-03-25-lightrag--simple-and-fast-retrieval-augmented-generation.md b/_posts/2026-03-25-lightrag--simple-and-fast-retrieval-augmented-generation.md
index 46b57bf..abf1f5f 100644
--- a/_posts/2026-03-25-lightrag--simple-and-fast-retrieval-augmented-generation.md
+++ b/_posts/2026-03-25-lightrag--simple-and-fast-retrieval-augmented-generation.md
@@ -4,7 +4,7 @@ title: "LightRAG: Simple and Fast Retrieval-Augmented Generation"
 date: 2024-10-08 08:00:12
 author: "Beijing University of Posts and Telecommunications"
 categories: ["Paper Reviews", "Retrieval-Augmented-Generation"]
-tags: ["Graph-Based-Text-Indexing", "Dual-Level-Retrieval-Paradigm", "Low-Level-Entity-Retrieval", "High-Level-Relationship-Retrieval", "Graph-Enhanced-Entity-and-Relationship-Extraction", "LLM-Profiling-for-Key-Value-Pair-Generation", "Incremental-Knowledge-Base-Updates", "Graph-Vector-Hybrid-Retrieval", "Multi-Hop-Subgraph-Information-Extraction", "Deduplication-for-Graph-Optimization"]
+tags: ["Graph-Based-Text-Indexing", "Dual-Level-Retrieval-Paradigm", "Low-Level-Entity-Retrieval", "High-Level-Relationship-Retrieval", "Graph-Enhanced-Entity-and-Relationship-Extraction", "LLM-Profiling-for-Key-Value-Pair-Generation", "Incremental-Knowledge-Base-Updates", "Graph-Vector-Hybrid-Retrieval", "Multi-Hop-Subgraph-Information-Extraction", "Deduplication-for-Graph-Optimization", "Retrieval-Augmented-Generation", "Knowledge-Graph"]
 cover: /assets/images/retrieval-augmented-generation.jpg
 use_math: true
 ---
diff --git a/_posts/2026-03-31-claude-code-architecture-analysis.md b/_posts/2026-03-31-claude-code-architecture-analysis.md
index b276c49..b94be56 100644
--- a/_posts/2026-03-31-claude-code-architecture-analysis.md
+++ b/_posts/2026-03-31-claude-code-architecture-analysis.md
@@ -7,7 +7,7 @@ description: >-
   Its 8-layer security and 4-tier message compaction, read as architecture.
 date: 2026-03-31 12:00:01
 categories: ["Insights", "Agentic-AI"]
-tags: ["Claude-Code", "Agentic-Architecture", "Context-Compaction", "Multi-Agent-Orchestration", "Security-Architecture"]
+tags: ["Claude-Code", "Agentic-Architecture", "Context-Compaction", "Multi-Agent-Orchestration", "Security-Architecture", "Agentic-AI"]
 cover: /assets/images/insights.jpg
 use_math: false
 lang: en
diff --git a/_posts/2026-03-31-claude-code-source-map-leak-analysis.md b/_posts/2026-03-31-claude-code-source-map-leak-analysis.md
index 3ea99b8..d8bade0 100644
--- a/_posts/2026-03-31-claude-code-source-map-leak-analysis.md
+++ b/_posts/2026-03-31-claude-code-source-map-leak-analysis.md
@@ -7,7 +7,7 @@ description: >-
   8계층 보안과 4단 메시지 압축, 에이전틱 루프를 아키텍처로 읽습니다.
 date: 2026-03-31 12:00:00
 categories: ["Insights", "Agentic-AI"]
-tags: ["Claude-Code", "Agentic-Architecture", "Context-Compaction", "Multi-Agent-Orchestration", "Security-Architecture"]
+tags: ["Claude-Code", "Agentic-Architecture", "Context-Compaction", "Multi-Agent-Orchestration", "Security-Architecture", "Agentic-AI"]
 cover: /assets/images/insights.jpg
 use_math: false
 lang: ko
diff --git a/_posts/2026-04-05-evolution-of-ai-agentic-patterns-en.md b/_posts/2026-04-05-evolution-of-ai-agentic-patterns-en.md
index 9a742b0..90ffc7c 100644
--- a/_posts/2026-04-05-evolution-of-ai-agentic-patterns-en.md
+++ b/_posts/2026-04-05-evolution-of-ai-agentic-patterns-en.md
@@ -7,7 +7,7 @@ description: >-
   shifts from prompts to context to harnesses, traced by why each era failed.
 date: 2026-04-05 12:00:00
 categories: ["Insights", "Agentic-AI"]
-tags: ["Prompt-Engineering", "Context-Engineering", "Harness-Engineering", "Agentic-Patterns", "LLM-Architecture", "Vibe-Coding"]
+tags: ["Prompt-Engineering", "Context-Engineering", "Harness-Engineering", "Agentic-Patterns", "LLM-Architecture", "Vibe-Coding", "Agentic-AI"]
 cover: /assets/images/insights.jpg
 use_math: false
 lang: en
diff --git a/_posts/2026-04-05-evolution-of-ai-agentic-patterns.md b/_posts/2026-04-05-evolution-of-ai-agentic-patterns.md
index 7ccae7a..73e2b9c 100644
--- a/_posts/2026-04-05-evolution-of-ai-agentic-patterns.md
+++ b/_posts/2026-04-05-evolution-of-ai-agentic-patterns.md
@@ -7,7 +7,7 @@ description: >-
   컨텍스트에서 하네스로 패러다임이 세 번 바뀐 2022-2026년을 각 시대가 왜 실패했는지로 추적합니다.
 date: 2026-04-05 12:00:00
 categories: ["Insights", "Agentic-AI"]
-tags: ["Prompt-Engineering", "Context-Engineering", "Harness-Engineering", "Agentic-Patterns", "LLM-Architecture", "Vibe-Coding"]
+tags: ["Prompt-Engineering", "Context-Engineering", "Harness-Engineering", "Agentic-Patterns", "LLM-Architecture", "Vibe-Coding", "Agentic-AI"]
 cover: /assets/images/insights.jpg
 use_math: false
 lang: ko
diff --git a/_posts/2026-04-12-agentcore-harness-engineering-analysis.md b/_posts/2026-04-12-agentcore-harness-engineering-analysis.md
index 76f8e3b..36c6559 100644
--- a/_posts/2026-04-12-agentcore-harness-engineering-analysis.md
+++ b/_posts/2026-04-12-agentcore-harness-engineering-analysis.md
@@ -7,7 +7,7 @@ description: >-
   Build·Deploy·Assess 세 층으로 열어 하네스 체크리스트의 어디를 채우고 어디를 비워 두었는지 짚습니다.
 date: 2026-04-12 12:00:00
 categories: ["Insights", "Agentic-AI"]
-tags: ["AgentCore", "AWS-Bedrock", "Harness-Engineering", "Agentic-Infrastructure", "MCP", "Cedar-Policy", "Managed-RAG", "Agent-Registry"]
+tags: ["AgentCore", "AWS-Bedrock", "Harness-Engineering", "Agentic-Infrastructure", "Model-Context-Protocol", "Cedar-Policy", "Managed-RAG", "Agent-Registry", "Agentic-AI"]
 cover: /assets/images/insights.jpg
 use_math: false
 ---
diff --git a/_posts/2026-06-07-zep-a-temporal-knowledge-graph-architecture-for-agent-memory.md b/_posts/2026-06-07-zep-a-temporal-knowledge-graph-architecture-for-agent-memory.md
index 4de99e7..1c98bc8 100644
--- a/_posts/2026-06-07-zep-a-temporal-knowledge-graph-architecture-for-agent-memory.md
+++ b/_posts/2026-06-07-zep-a-temporal-knowledge-graph-architecture-for-agent-memory.md
@@ -5,7 +5,7 @@ subtitle: "Zep: 에이전트 메모리를 위한 시간 인식 지식 그래프
 date: 2025-01-20 16:52:48
 author: "Preston Rasmussen et al."
 categories: ["Paper Reviews", "Retrieval-Augmented-Generation"]
-tags: ["Temporally-Aware-Knowledge-Graph-Engine", "Bi-Temporal-Knowledge-Graph-Modeling", "Hierarchical-Knowledge-Graph-Construction", "Dynamic-Edge-Invalidation-for-Temporal-Reasoning", "Episodic-and-Semantic-Memory-Subgraphs", "Community-Detection-with-Label-Propagation", "Hybrid-Search-with-Breadth-First-Graph-Traversal", "Non-Lossy-Knowledge-Graph-Updates", "Multi-Hop-Entity-and-Relationship-Extraction", "Graph-Based-Memory-Retrieval-for-LLM-Agents"]
+tags: ["Temporally-Aware-Knowledge-Graph-Engine", "Bi-Temporal-Knowledge-Graph-Modeling", "Hierarchical-Knowledge-Graph-Construction", "Dynamic-Edge-Invalidation-for-Temporal-Reasoning", "Episodic-and-Semantic-Memory-Subgraphs", "Community-Detection-with-Label-Propagation", "Hybrid-Search-with-Breadth-First-Graph-Traversal", "Non-Lossy-Knowledge-Graph-Updates", "Multi-Hop-Entity-and-Relationship-Extraction", "Graph-Based-Memory-Retrieval-for-LLM-Agents", "Retrieval-Augmented-Generation", "Knowledge-Graph"]
 cover: /assets/images/retrieval-augmented-generation.jpg
 use_math: true
 ---
diff --git a/_sass/_dark.scss b/_sass/_dark.scss
index a7bfb8e..fdbf88f 100644
--- a/_sass/_dark.scss
+++ b/_sass/_dark.scss
@@ -190,4 +190,16 @@
 }
 
 // Opt-in only: dark applies solely when the visitor toggles it on.
-html[data-theme="dark"] { @include dark-theme; }
+// `color-scheme` is what tells the browser to render its OWN widgets — the
+// scrollbar, the search field's clear button, form controls — for a dark page.
+// It used to be a single unconditional `<meta name="color-scheme"
+// content="light">` in head.html, so those stayed light-rendered against
+// #0d1117. Declared here instead of in the meta tag because only this file knows
+// which theme is active, and the inline bootstrap in head.html sets data-theme
+// before first paint, so there is no flash.
+html { color-scheme: light; }
+
+html[data-theme="dark"] {
+  color-scheme: dark;
+  @include dark-theme;
+}
diff --git a/categories.html b/categories.html
index b966805..6dc053a 100644
--- a/categories.html
+++ b/categories.html
@@ -7,19 +7,18 @@
   Tech Guides, and Insights.
 ---
 
+{%- comment -%}
+  Category names are already written in display case in front matter
+  ("Agentic-AI", "Retrieval-Augmented-Generation"), so replacing the separators
+  is the whole job — which is what index.html and _layouts/post.html do.
+
+  This used to title-case each word through `capitalize`, which uppercases the
+  first letter and *lowercases the rest*: "Agentic-AI" shipped as "Agentic Ai"
+  on the live page while every other listing rendered it correctly.
+{%- endcomment -%}
 {% for category in site.categories %}
   {% capture cat %}{{ category | first }}{% endcapture %}
-  {% assign cat_temp = cat | replace: "-", " " | replace: "_", " " %}
-  {% assign words = cat_temp | split: " " %}
-  {% assign cat_display = "" %}
-  {% for word in words %}
-    {% capture capitalized_word %}{{ word | capitalize }}{% endcapture %}
-    {% if forloop.first %}
-      {% assign cat_display = capitalized_word %}
-    {% else %}
-      {% assign cat_display = cat_display | append: " " | append: capitalized_word %}
-    {% endif %}
-  {% endfor %}
+  {% assign cat_display = cat | replace: "-", " " | replace: "_", " " %}
   <h2 id="{{ cat | slugify }}">{{ cat_display }}</h2>
   <ul class="posts-list">
   {% for post in site.categories[cat] %}
diff --git a/css/main.scss b/css/main.scss
index 96c7894..d0f15cf 100755
--- a/css/main.scss
+++ b/css/main.scss
@@ -12,5 +12,5 @@
   "post",             // Post-specific styles
   "tags",             // Tags listing page
   "syntax",           // Rouge syntax-highlighting theme (server-side)
-  "dark"              // prefers-color-scheme: dark overrides (loaded last)
+  "dark"              // opt-in html[data-theme="dark"] overrides (loaded last)
 ;
diff --git a/docs/tech-doc.md b/docs/tech-doc.md
index 1e5012e..87212b4 100644
--- a/docs/tech-doc.md
+++ b/docs/tech-doc.md
@@ -52,14 +52,14 @@ Pages가 그대로 서빙.**
 |------|------|-----------|
 | 정적 사이트 생성기 | Jekyll 4.4 (`Gemfile`, Ruby 3.3+), kramdown(GFM 입력) | 마크다운 글을 HTML로 굽는 본체. `Gemfile`은 Ruby의 의존성 목록(= `package.json`에 해당) |
 | 플러그인(gem) | `jekyll-paginate`, `jekyll-sitemap`, `jekyll-feed` | 각각 목록 페이지 나누기, `sitemap.xml`(검색엔진용 지도), `feed.xml`(RSS 구독) 자동 생성 |
-| 로컬 플러그인(`_plugins/`) | `reading_time.rb`(한·영 읽기시간 계산), `lazy_images.rb`(`<img>`에 lazy-load 부여), `post_description.rb`(글의 검색 설명문 생성), `related_posts.rb`(공유 태그로 관련 글 선정), `scrollable_tables.rb`(넓은 표를 가로 스크롤 컨테이너로 감싸기) | 우리가 직접 만든 Ruby 확장. 이것 때문에 GitHub Pages 기본 빌드 대신 Jekyll을 직접 돌린다(§1 아래 참고) |
+| 로컬 플러그인(`_plugins/`) | `reading_time.rb`(한·영 읽기시간 계산), `lazy_images.rb`(`<img>`에 lazy-load 부여), `post_description.rb`(글의 검색 설명문 생성), `related_posts.rb`(공유 태그로 관련 글 선정 + 글 끝 이전/다음 링크), `scrollable_tables.rb`(넓은 표를 가로 스크롤 컨테이너로 감싸기) | 우리가 직접 만든 Ruby 확장. 이것 때문에 GitHub Pages 기본 빌드 대신 Jekyll을 직접 돌린다(§1 아래 참고) |
 | 신택스 하이라이팅 | Rouge(서버사이드, kramdown 내장) | 코드 블록에 색을 입히는 작업을 **빌드 때 미리** 한다(브라우저 부담 0). 색 테마는 `_sass/_syntax.scss` |
 | 수식 | kramdown `math_engine: mathjax` → MathJax 3, 포스트별 `use_math`로 로드 | 수학 기호를 브라우저에서 예쁘게 그려 주는 라이브러리. 수식이 있는 글에서만 불러온다 |
 | 스타일 | Sass(`_sass/`), 벤더링된 Bourbon + Neat 그리드 프레임워크 | "벤더링"은 외부 라이브러리를 저장소 안에 복사해 둔 것. `jekyll-sass-converter` 2.x(libsass)로 **고정** — 3.x(dart-sass)는 Bourbon/Neat의 구식 `/` 나눗셈 문법에서 에러 |
 | 자바스크립트 | 바닐라 JS(`js/main.js`, jQuery 없음) | "바닐라"는 프레임워크 없이 순수 JS만 쓴다는 뜻. 이미지 확대는 GLightbox. 외부 CDN 스크립트는 **SRI**로 무결성 검증(아래 설명) |
 | 검색 | `search.json`(전체 본문 색인) 위에 `simple-jekyll-search` | 서버 없이 브라우저에서 도는 검색. 미리 만들어 둔 색인 파일을 받아 클라이언트가 직접 찾는다 |
-| 다크모드 | 라이트가 기본, 토글로 opt-in(`_sass/_dark.scss`, `[data-theme="dark"]`) | OS의 다크모드 설정은 **따르지 않고**, 사용자가 버튼을 눌러 켜야 한다 |
-| 호스팅/CI | GitHub Actions로 GitHub Pages 배포(`.github/workflows/jekyll.yml`) | `main`에 푸시하면 자동으로 빌드·검사·배포 |
+| 다크모드 | 라이트가 기본, 토글로 opt-in(`_sass/_dark.scss`, `[data-theme="dark"]`) | OS의 다크모드 설정은 **따르지 않고**, 사용자가 버튼을 눌러 켜야 한다. 테마에 따라 바뀌어야 하는 두 값이 CSS 밖에 있다 — `color-scheme`(스크롤바 등 브라우저 위젯)은 `_dark.scss`가 `data-theme`으로 처리하고, `theme-color`(모바일 주소창)는 메타 태그라 `head.html`의 인라인 스크립트와 `main.js`의 토글이 함께 갱신한다 |
+| 호스팅/CI | GitHub Actions로 GitHub Pages 배포(`.github/workflows/jekyll.yml`) | `main`에 푸시하면 자동으로 빌드·검사·배포. 같은 검사가 `main`으로 향하는 PR에서도 돌아 머지 전에 막는다 |
 
 > **용어 — CDN과 SRI.** *CDN*은 인기 라이브러리를 전 세계 서버에서 빠르게 내려주는 공용
 > 배포망이다(예: cdnjs, jsDelivr). 문제는 그 외부 파일이 몰래 바뀌면 우리 사이트에
@@ -107,7 +107,9 @@ Pages가 그대로 서빙.**
 포스트는 프런트매터에 **2단계** 카테고리를 가진다: `categories: ["<유형>", "<주제>"]`.
 
 - **0단계 (유형)** — `Paper Reviews`, `Paper Summaries`, `Tech Guides`, `Insights` 중 하나.
-  글이 어느 내비 탭에 들어갈지를 정한다. 각각 전용 페이지를 가진다.
+  글이 어느 섹션 페이지에 들어갈지를 정한다. 넷 다 전용 페이지를 가지지만, 지금 글이 있는
+  유형은 `Paper Reviews`(28편)와 `Insights`(7편)뿐이고 내비에는 이 둘만 올라간다(아래
+  "내비게이션은 코드가 아니라 데이터가 만든다" 참조).
 - **1단계 (주제)** — `Language-Models`, `Multimodal-Learning`, `Finetuning`,
   `Retrieval-Augmented-Generation`, `Agentic-AI` 등. 세부 주제이며 필요하면 자유롭게 추가한다.
 
@@ -133,12 +135,18 @@ _site/paper reviews/language-models/2025/01/23/<slug>.html
 | nav_order | 페이지 | 소스 |
 |-----------|--------|------|
 | 1 | About | `about.md` |
-| 2 | Paper Summaries | `paper-summaries.md` — `site.categories['Paper Summaries']` 필터 |
 | 3 | Paper Reviews | `paper-reviews.md` — `site.categories['Paper Reviews']` 필터, 주제별 그룹화 |
-| 4 | Tech Guides | `tech-guides.md` — `site.categories['Tech Guides']` 필터 |
 | 5 | Insights | `insights.md` — `site.categories['Insights']` 필터 |
 | 6 | Search | `search.md` |
 
+`nav_order`의 2·4번이 빈 것은 실수가 아니다. `paper-summaries.md`(2)와
+`tech-guides.md`(4)는 **글이 0편이라 내비에서 내렸다** — 여섯 탭 중 둘이
+"this space will fill up soon"으로 이어지고 있었고, 같은 페이지가 사이트맵에도 올라가
+검색엔진에 빈 페이지를 내밀고 있었다. 두 파일은 `main_nav` 없이 `sitemap: false`로
+남아 있고(URL로는 계속 열린다), 첫 글이 그 유형으로 들어오면 프런트매터 주석이 지시하는
+대로 `main_nav: true`와 원래 `nav_order`를 되살리면 된다. 번호를 다시 매기지 않은 이유가
+이것이다.
+
 `categories.html`(`/categories/`)과 `tags.html`(`/tags/`)은 *모든* 카테고리/태그를 가로지르는
 전체 색인 페이지다. 메인 내비에는 없고, 각 포스트 하단의 메타데이터에서 링크된다.
 
@@ -152,7 +160,29 @@ _site/paper reviews/language-models/2025/01/23/<slug>.html
 > **함정 — 앵커 점프 깨짐.** 메타데이터의 카테고리/태그 링크는 모두 `slugify` 필터로 ID를
 > 만든다. 과거에는 링크 쪽은 `downcase`(공백 유지), 제목(H2)의 id는 원본 케이스(`Paper
 > Reviews`)를 써서 서로 어긋났고, 클릭해도 해당 위치로 점프하지 못했다(html-proofer가
-> 128건 적발). 지금은 양쪽 다 `slugify`로 통일했다.
+> 128건 적발). 지금은 양쪽 다 `slugify`로 통일했다 — `paper-reviews.md`가 마지막까지
+> `id="{{cat}}"`로 원본 케이스를 내고 있었는데, 그 앵커를 가리키는 링크가 없어 html-proofer가
+> 잡지 못했다. 새 목록 페이지를 만들 땐 `| slugify`를 빼먹지 않는다.
+
+### 태그: 세부 태그 + 통제된 주제 태그
+
+태그는 자유 형식이지만 **두 층**으로 쓴다.
+
+- **세부 태그** — 논문 하나의 기여를 서술한다(`Fine-Grained-Expert-Segmentation`,
+  `Dependency-Aware-Tree-Traversal`). 논문 색인으로는 정확하다.
+- **주제 태그** — 글을 서로 잇는다. `Agentic-AI`, `Alignment`, `DeepSeek`,
+  `Knowledge-Graph`, `Llama`, `Mixture-of-Experts`, `Multimodal-Models`,
+  `Reasoning-Models`, `Retrieval-Augmented-Generation`.
+
+세부 태그만 있으면 관련 글 기능이 굶는다. 태그 264개 중 220개(83%)가 단 한 편에만 붙어
+있었고, `related_posts.rb`가 공유 태그를 요구하므로 11편이 관련 글 0개였다 — 그중 넷은
+RAPTOR·GraphRAG·LightRAG·Zep, 즉 서로 붙어야 마땅한 그래프 RAG 논문들이었다. 한 주제를 세
+이름으로 부른 것(`Agentic-Architecture`/`Agentic-Patterns`/`Agentic-Infrastructure`)도 같은
+증상을 만들었다.
+
+주제 태그를 한 층 얹어 35편 전부가 관련 글을 얻게 했다. 고친 곳이 `related_posts.rb`의
+규칙이 아니라 **글의 프런트매터**라는 점이 중요하다 — 규칙을 느슨하게 풀면 공통점 없는 글이
+서로 추천된다. 세부 태그는 지우지 않았다(살아 있는 `/tags/` 앵커가 움직인다).
 
 ## 4. 수식 렌더링
 
@@ -206,8 +236,13 @@ MathJax 설정은 `head.html`에 있고 `{% if page.use_math %}`로 감싸 **프
   없이 단순 부분문자열 매칭을 한다. 즉 색인에 없는 글자는 못 찾는다. 예전에 snippet(앞
   40단어)만 색인했더니, "어텐션"·"트랜스포머"가 본문 중·후반에 18개 포스트나 있는데 발췌엔
   안 들어가 검색 결과가 **0건**으로 나왔다. `content`로 전체를 색인해 한글 재현율을 회복했다.
-  파일은 약 2.7 MB지만 gzip 압축 후 ~786 KB이고, `/search/` 페이지에서만 로드되므로 다른
-  페이지 속도엔 영향이 없다.
+  대가는 실측 **3.2 MB, gzip 991 KB**다(글이 늘면 같이 는다). `/search/`에서만 내려받으므로
+  다른 페이지 속도엔 영향이 없다.
+
+  > **알려진 한계 — 매치가 안 보일 수 있다.** 색인은 본문 전체인데 결과 카드에 찍히는 건
+  > 앞 40단어짜리 `snippet`뿐이다. 그래서 5만 번째 글자에서 걸린 검색어는 스니펫에 없고
+  > `<mark>` 강조도 안 걸려서, 맞는 결과가 엉뚱한 결과처럼 보인다. 고치려면 매치 위치
+  > 주변을 잘라 보여 주는 스니펫이 필요하다(`simple-jekyll-search`의 템플릿 밖 작업).
 - **`js/search.js`** 가 `simple-jekyll-search`(CDN 버전 **고정 + SRI**: `1.10.0`)를
   `search.md`의 `#search-input` 입력칸에 연결한다. 매칭된 키워드를 `<mark>`로 강조하고
   결과 개수를 라이브 상태줄에 표시한다. 강조 처리는 결과가 다 그려진 뒤 **디바운스**로 단
@@ -217,9 +252,10 @@ MathJax 설정은 `head.html`에 있고 `{% if page.use_math %}`로 감싸 **프
   > 마지막 입력 뒤 잠깐 멈출 때까지 기다렸다가 **딱 한 번만** 함수를 실행하는 기법이다.
 - `category`/`tags`도 색인에 들어가므로 제목·본문뿐 아니라 메타데이터로도 검색된다.
   색인에는 **글(`site.posts`)만** 들어간다 — `search.json`이 `site.posts`를 순회하므로
-  About·Search·index 같은 페이지는 애초에 후보가 아니다. (`_config.yml`에 남아 있는
-  `simple_jekyll_search.exclude`는 아무것도 하지 않는다. `simple-jekyll-search`는
-  브라우저에서 도는 JS라 `_config.yml`을 읽을 수 없다.)
+  About·Search·index 같은 페이지는 애초에 후보가 아니다. (`_config.yml`에 `about.md`·
+  `search.md`·`index.html`을 열거하던 `simple_jekyll_search.exclude` 블록이 있었는데,
+  `simple-jekyll-search`는 브라우저에서 도는 JS라 `_config.yml`을 읽을 수 없어 아무것도 하지
+  않았다. 지워졌고, 왜 없는지가 `_config.yml`의 주석으로 남아 있다.)
 
 ## 6. 스타일 (Sass)
 
@@ -266,6 +302,11 @@ Bourbon → base/ → Neat → _layout → _post → _tags → _syntax(Rouge 코
   `bundle exec htmlproofer ./_site --disable-external`로 미리 확인하면 배포 실패를 예방할 수
   있다.
 
+  같은 워크플로가 **`main`으로 향하는 PR에서도** 돈다(`deploy` 잡과 아티팩트 업로드만
+  `github.event_name != 'pull_request'`로 건너뛴다). 이게 없던 동안 네 게이트는 머지 **후에만**
+  돌았다 — 잘못된 PR이 main에 착지한 다음 배포에 실패해서, 사이트는 마지막 정상 배포를 계속
+  서빙하지만 main이 빨간 상태로 남았다. 지금은 머지 전에 걸린다.
+
   > **용어 — CI.** Continuous Integration. 코드를 올리면 정해 둔 검사·빌드·배포를 자동으로
   > 돌려주는 파이프라인이다. 여기서는 GitHub Actions가 그 역할을 한다.
 
diff --git a/insights.md b/insights.md
index f7a604c..5daa092 100644
--- a/insights.md
+++ b/insights.md
@@ -9,7 +9,15 @@ main_nav: true
 nav_order: 5
 ---
 
-<h2 id="insights">Insights</h2>
+{%- comment -%}
+  No heading here. _layouts/page.html already renders the front-matter `title` as
+  this page's top-level heading, so a level-2 heading repeating it printed
+  "Insights" twice in a row — which is what /paper-reviews/ never did.
+
+  A Liquid comment, not an HTML one: an HTML comment ships to the reader, and
+  spelling a heading tag inside it also trips the "exactly one h1 per page" check
+  in script/validate-site.sh, which counts occurrences in the served markup.
+{%- endcomment -%}
 <p class="desc"><em>The threads a single paper can't hold — agentic-AI architecture, industry shifts, and the design decisions underneath them.</em></p>
 
 {% include category-posts.html category="Insights" empty="The first piece is on its way — check back soon." %}
diff --git a/js/main.js b/js/main.js
index bf17548..d3095f5 100644
--- a/js/main.js
+++ b/js/main.js
@@ -7,21 +7,26 @@ document.addEventListener('DOMContentLoaded', function () {
   var themeToggle = document.getElementById('theme-toggle');
   if (themeToggle) {
     var icon = themeToggle.querySelector('i');
+    var themeColor = document.querySelector('meta[name="theme-color"]');
     function isDark() {
       return document.documentElement.getAttribute('data-theme') === 'dark';
     }
-    function syncIcon() {
+    function syncTheme() {
       if (icon) icon.className = isDark() ? 'fa-solid fa-sun' : 'fa-regular fa-moon';
       // aria-label is static, so without this a screen-reader user cannot tell
       // whether dark mode is currently on.
       themeToggle.setAttribute('aria-pressed', isDark() ? 'true' : 'false');
+      // Repaints the mobile browser's own chrome. The inline bootstrap in
+      // head.html does this on load; without it here the address bar stayed
+      // white until the next navigation. Keep in step with $bg in _dark.scss.
+      if (themeColor) themeColor.setAttribute('content', isDark() ? '#0d1117' : '#ffffff');
     }
-    syncIcon();
+    syncTheme();
     themeToggle.addEventListener('click', function () {
       var next = isDark() ? 'light' : 'dark';
       document.documentElement.setAttribute('data-theme', next);
       try { localStorage.setItem('theme', next); } catch (e) {}
-      syncIcon();
+      syncTheme();
     });
   }
 
diff --git a/paper-reviews.md b/paper-reviews.md
index 1865a90..6086e46 100644
--- a/paper-reviews.md
+++ b/paper-reviews.md
@@ -23,7 +23,13 @@ nav_order: 3
 
 {% for cat in topics %}
   {% assign cat_display = cat | replace: "-", " " | replace: "_", " " %}
-  <h2 id="{{cat}}">{{ cat_display }}</h2>
+  {%- comment -%}
+    slugify, like every other category anchor on the site (post metadata links,
+    /categories/). This emitted the raw value — id="Language-Models" — so it was
+    the one page whose heading ids did not match the convention documented in
+    docs/tech-doc.md §3.
+  {%- endcomment -%}
+  <h2 id="{{ cat | slugify }}">{{ cat_display }}</h2>
   <ul class="posts-list">
   {% for post in review_posts %}
     {% if post.categories contains cat %}
diff --git a/paper-summaries.md b/paper-summaries.md
index 790d78b..90b2cb1 100644
--- a/paper-summaries.md
+++ b/paper-summaries.md
@@ -5,11 +5,15 @@ permalink: /paper-summaries/
 description: >-
   Concise digests of AI/ML papers, posts, and talks — distilled to their core
   contributions and insights.
-main_nav: true
-nav_order: 2
+# No `main_nav` and `sitemap: false` while this category is empty: every post so
+# far is a Paper Review or an Insight, so the nav advertised a section whose only
+# content was "check back soon", and the sitemap offered search engines a page
+# with no posts on it. Restore both — `main_nav: true` and `nav_order: 2` — with
+# the first post filed under this category.
+sitemap: false
 ---
 
-<h2 id="paper-summaries">Paper Summaries</h2>
+{%- comment -%} No heading here — see insights.md. {%- endcomment -%}
 <p class="desc"><em>Concise digests of AI/ML papers, posts, and talks — distilled to their core contributions and insights.</em></p>
 
 {% include category-posts.html category="Paper Summaries" empty="Summaries are in the works — this space will fill up soon." %}
diff --git a/tech-guides.md b/tech-guides.md
index 63b86fe..6390f1e 100644
--- a/tech-guides.md
+++ b/tech-guides.md
@@ -5,11 +5,13 @@ permalink: /tech-guides/
 description: >-
   Hands-on technical guides and tutorials for the libraries, frameworks, tools,
   and platforms I work with.
-main_nav: true
-nav_order: 4
+# See paper-summaries.md: hidden from the nav and the sitemap while empty.
+# Restore `main_nav: true` and `nav_order: 4` with the first post in this
+# category.
+sitemap: false
 ---
 
-<h2 id="tech-guides">Tech Guides</h2>
+{%- comment -%} No heading here — see insights.md. {%- endcomment -%}
 <p class="desc"><em>Hands-on technical guides and tutorials for the libraries, frameworks, tools, and platforms I work with.</em></p>
 
 {% include category-posts.html category="Tech Guides" empty="Hands-on guides are on the way — check back soon." %}
diff --git a/test/test_related_posts.rb b/test/test_related_posts.rb
index 3a0cfd6..ffd9e25 100644
--- a/test/test_related_posts.rb
+++ b/test/test_related_posts.rb
@@ -145,46 +145,132 @@ def test_is_deterministic_for_equal_scores_and_dates
   end
 end
 
+class TestAdjacent < Minitest::Test
+  def ordered(*posts)
+    posts.sort_by { |p| [p[:date].to_time.to_i, p[:id]] }
+  end
+
+  def test_returns_the_neighbours_in_date_order
+    old = post("/old", date: "2024-01-01")
+    mid = post("/mid", date: "2025-01-01")
+    new = post("/new", date: "2026-01-01")
+    older, newer = RelatedPosts.adjacent(mid, ordered(old, mid, new))
+    assert_equal "/old", older[:id]
+    assert_equal "/new", newer[:id]
+  end
+
+  def test_nil_at_each_end_of_the_archive
+    old = post("/old", date: "2024-01-01")
+    new = post("/new", date: "2026-01-01")
+    list = ordered(old, new)
+    assert_nil RelatedPosts.adjacent(old, list)[0]
+    assert_nil RelatedPosts.adjacent(new, list)[1]
+  end
+
+  # The live defect: all three translation pairs share a date exactly, so
+  # Jekyll's page.previous pointed the newest Korean post at its own English
+  # version.
+  def test_skips_the_translation_twin_even_on_an_identical_date
+    ko = post("/ko", lang: "ko", translation_id: "t1", date: "2026-07-27")
+    en = post("/en", lang: "en", translation_id: "t1", date: "2026-07-27")
+    before = post("/before", lang: "ko", date: "2026-04-12")
+    older, newer = RelatedPosts.adjacent(ko, ordered(before, en, ko))
+    assert_equal "/before", older[:id]
+    assert_nil newer
+  end
+
+  # Skipping past an ineligible neighbour, rather than dropping the link: the
+  # reader still gets a previous post, just not the untranslated one.
+  def test_steps_over_an_other_language_neighbour
+    target = post("/target", lang: "ko", date: "2026-03-01")
+    english = post("/english", lang: "en", date: "2026-02-01")
+    korean = post("/korean", lang: "ko", date: "2026-01-01")
+    older, = RelatedPosts.adjacent(target, ordered(korean, english, target))
+    assert_equal "/korean", older[:id]
+  end
+
+  def test_an_english_post_navigates_english_posts
+    en_a = post("/en-a", lang: "en", date: "2026-01-01")
+    ko = post("/ko", lang: "ko", date: "2026-02-01")
+    en_b = post("/en-b", lang: "en", date: "2026-03-01")
+    older, newer = RelatedPosts.adjacent(en_b, ordered(en_a, ko, en_b))
+    assert_equal "/en-a", older[:id]
+    assert_nil newer
+  end
+
+  # Most paper posts declare no lang; they must not be treated as a third
+  # language and cut off from each other.
+  def test_posts_without_an_explicit_lang_are_neighbours
+    a = post("/a", date: "2024-01-01")
+    b = post("/b", date: "2025-01-01", lang: "ko")
+    older, = RelatedPosts.adjacent(b, ordered(a, b))
+    assert_equal "/a", older[:id]
+  end
+
+  def test_returns_nothing_when_the_target_is_not_in_the_list
+    assert_equal [nil, nil], RelatedPosts.adjacent(post("/x"), [post("/y")])
+  end
+
+  def test_a_lone_post_has_no_neighbours
+    only = post("/only")
+    assert_equal [nil, nil], RelatedPosts.adjacent(only, [only])
+  end
+end
+
 class TestTheActualHarnessCluster < Minitest::Test
-  # The real front matter of the five posts this feature was built for. The
-  # Claude Code posts share no tags with the harness chronicle — they are related
-  # only by category, which is deliberately not enough.
+  # The real front matter of the five Insights/Agentic-AI posts.
+  #
+  # These five used to share nothing at all: the Claude Code teardown said
+  # "Agentic-Architecture", the harness chronicle said "Agentic-Patterns", and
+  # AgentCore said "Agentic-Infrastructure" — three names for one topic, so the
+  # shared-tag requirement found no pair. The explicit "Agentic-AI" topic tag is
+  # what connects them, and it is a tag, not a category, that does it.
   def setup
     @evolution_ko = post("/evolution-ko", lang: "ko", translation_id: "evo", date: "2026-04-05",
                          tags: %w[Prompt-Engineering Context-Engineering Harness-Engineering
-                                  Agentic-Patterns LLM-Architecture Vibe-Coding],
+                                  Agentic-Patterns LLM-Architecture Vibe-Coding Agentic-AI],
                          categories: ["Insights", "Agentic-AI"])
     @evolution_en = post("/evolution-en", lang: "en", translation_id: "evo", date: "2026-04-05",
                          tags: @evolution_ko[:tags], categories: ["Insights", "Agentic-AI"])
     @agentcore = post("/agentcore", date: "2026-04-12",
                       tags: %w[AgentCore AWS-Bedrock Harness-Engineering Agentic-Infrastructure
-                               MCP Cedar-Policy Managed-RAG Agent-Registry],
+                               Model-Context-Protocol Cedar-Policy Managed-RAG Agent-Registry
+                               Agentic-AI],
                       categories: ["Insights", "Agentic-AI"])
     @claude_ko = post("/claude-ko", lang: "ko", translation_id: "cc", date: "2026-03-31",
                       tags: %w[Claude-Code Agentic-Architecture Context-Compaction
-                               Multi-Agent-Orchestration Security-Architecture],
+                               Multi-Agent-Orchestration Security-Architecture Agentic-AI],
                       categories: ["Insights", "Agentic-AI"])
     @claude_en = post("/claude-en", lang: "en", translation_id: "cc", date: "2026-03-31",
                       tags: @claude_ko[:tags], categories: ["Insights", "Agentic-AI"])
     @all = [@evolution_ko, @evolution_en, @agentcore, @claude_ko, @claude_en]
   end
 
-  # Harness-Engineering is the one tag they share, and it is the query cluster
-  # this feature exists to consolidate.
-  def test_agentcore_and_the_harness_chronicle_relate_to_each_other
-    assert_equal ["/evolution-ko"], RelatedPosts.rank(@agentcore, @all)
-    assert_equal ["/agentcore"], RelatedPosts.rank(@evolution_ko, @all)
+  # Two shared tags (Harness-Engineering and Agentic-AI) put these two ahead of
+  # the teardown, which shares only the topic tag. Depth of overlap still orders
+  # the list.
+  def test_agentcore_and_the_harness_chronicle_rank_first_for_each_other
+    assert_equal ["/evolution-ko", "/claude-ko"], RelatedPosts.rank(@agentcore, @all)
+    assert_equal ["/agentcore", "/claude-ko"], RelatedPosts.rank(@evolution_ko, @all)
+  end
+
+  # The taxonomy fix, pinned: this post had no related block on the live site.
+  def test_the_claude_code_teardown_now_connects_through_the_topic_tag
+    assert_equal ["/agentcore", "/evolution-ko"], RelatedPosts.rank(@claude_ko, @all)
   end
 
-  # Both are Insights/Agentic-AI, but they share no topic tag. "Also about agents"
-  # is not a reason to send a reader from one to the other.
-  def test_the_claude_code_teardown_is_not_linked_on_category_alone
-    assert_equal [], RelatedPosts.rank(@claude_ko, @all)
+  # Still true, and still the rule: strip the shared topic tag and shared
+  # categories alone must not resurrect the link.
+  def test_the_shared_category_alone_would_not_have_connected_them
+    bare_claude = @claude_ko.merge(tags: @claude_ko[:tags] - ["Agentic-AI"])
+    bare_all = @all.map { |p| p[:id] == bare_claude[:id] ? bare_claude : p }
+    assert_equal [], RelatedPosts.rank(bare_claude, bare_all)
   end
 
+  # Only three English posts exist and each is a translation, so an English
+  # reader's candidate pool is tiny — the topic tag is what keeps it non-empty.
   def test_an_english_post_only_sees_english_relatives
-    # /claude-en shares no tag with /evolution-en, and /agentcore is Korean.
-    assert_equal [], RelatedPosts.rank(@evolution_en, @all)
+    assert_equal ["/claude-en"], RelatedPosts.rank(@evolution_en, @all)
   end
 
   def test_no_post_is_offered_its_own_translation