bits-bytes-nn · Oct 16, 2026 · Oct 16, 2026
diff --git a/README.ko.md b/README.ko.md
@@ -93,7 +93,8 @@ test/              _plugins/와 글 린트의 minitest 단위 테스트
 script/            lint-posts.rb (빌드 전 글 원본 규칙 검사)
                    validate-site.sh (빌드 후 검색 노출 검사)
                    check-plugin-output.rb (플러그인 훅이 글에 남긴 결과 검사,
-                   validate-site.sh가 실행)
+                   validate-site.sh가 실행), published-urls.txt (글이 게시된
+                   적 있는 모든 URL)
 _data/             post_conventions.yml (카테고리·표지·통제 태그·폐기 태그의
                    유일한 정본. scholar-lens도 읽음), covers.yml (표지마다 출처와
                    저작자 표시. 글 머리 이미지 아래에 표시)
@@ -177,7 +178,10 @@ _site/paper-reviews/language-models/2025/01/23/<slug>.html
 ```
 
 그래서 **이미 게시된 글의 카테고리나 날짜를 바꾸면 URL이 바뀝니다.** 외부에서 걸린 링크와
-검색 결과가 깨지므로, 한 번 정하면 그대로 둡니다.
+검색 결과가 깨지므로, 한 번 정하면 그대로 둡니다. 게시된 URL은 모두
+`script/published-urls.txt`에 있고, 그중 하나라도 빌드되지 않으면 `validate-site.sh`가 실패합니다.
+꼭 바꿔야 한다면 예전 URL을 글의 `redirect_from`에 넣습니다. 글이 머지되면
+`bundle exec ruby script/check-plugin-output.rb <site_dir> --record`로 그 URL을 기록합니다.
 
 예전 규칙으로 게시된 옛 URL(`/paper%20reviews/…`)은 계속 동작합니다.
 `_plugins/legacy_urls.rb`가 그 경로를 글의 `redirect_from`에 넣고, jekyll-redirect-from이 그

diff --git a/README.md b/README.md
@@ -95,7 +95,8 @@ test/              minitest unit tests for _plugins/ and the post lint
 script/            lint-posts.rb (post sources vs the conventions, before the build)
                    validate-site.sh (post-build discoverability checks)
                    check-plugin-output.rb (each plugin hook's mark on the posts;
-                   run by validate-site.sh)
+                   run by validate-site.sh), published-urls.txt (every URL a
+                   post has been published under)
 _data/             post_conventions.yml (categories, covers, controlled and
                    retired tags — the one copy, also read by scholar-lens);
                    covers.yml (each cover's source and credit, shown under the hero)
@@ -184,7 +185,11 @@ _site/paper-reviews/language-models/2025/01/23/<slug>.html
 ```
 
 So **changing the categories or date of a published post changes its URL**, which
-breaks inbound links and search results. Set them once and leave them.
+breaks inbound links and search results. Set them once and leave them. Every
+published URL is listed in `script/published-urls.txt`, and `validate-site.sh`
+fails when one is no longer built; if a change is unavoidable, add the old URL
+to the post's `redirect_from`. After a post merges, record its URL with
+`bundle exec ruby script/check-plugin-output.rb <site_dir> --record`.
 
 Posts first published under the old unslugified form (`/paper%20reviews/…`) keep
 working: `_plugins/legacy_urls.rb` adds that path to each post's `redirect_from`,

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
@@ -42,7 +42,7 @@ This post has one thesis: **Engineering rigor never disappeared. It relocated.**
 
 Each relocation was triggered by the previous era's failure. Let's trace that arc.
 
-<a href="/assets/images/three-era-timeline.svg" class="glightbox" data-gallery="agentic-patterns" data-glightbox="title: Evolution timeline of AI agentic patterns">
+<a href="/assets/images/three-era-timeline.svg" class="glightbox" data-gallery="post-images" data-glightbox="title: Evolution timeline of AI agentic patterns">
   <img src="/assets/images/three-era-timeline.svg" alt="Evolution timeline of AI agentic patterns — Prompt → Context → Harness Engineering" />
 </a>
 
@@ -285,7 +285,7 @@ The metric they called "the single most important metric" for production agents
 
 Here's how KV-cache works. When you send a prompt to an LLM API, the model computes attention Keys and Values for each token. This computation is expensive. But if the **prefix** of the current request matches the previous request, that portion can reuse cached results — no recomputation needed. On Claude Sonnet, a cache hit reduces cost by **10x**. The difference between an agent re-computing the system prompt across 30 turns versus leveraging the cache is enormous.
 
-<a href="/assets/images/kv-cache-mechanism.svg" class="glightbox" data-gallery="agentic-patterns" data-glightbox="title: How KV-cache works — the stable prefix determines cost">
+<a href="/assets/images/kv-cache-mechanism.svg" class="glightbox" data-gallery="post-images" data-glightbox="title: How KV-cache works — the stable prefix determines cost">
   <img src="/assets/images/kv-cache-mechanism.svg" alt="How KV-cache works — the stable prefix determines cost" />
 </a>
 

diff --git a/_posts/2026-04-05-evolution-of-ai-agentic-patterns.md b/_posts/2026-04-05-evolution-of-ai-agentic-patterns.md
@@ -42,7 +42,7 @@ translation_id: evolution-of-ai-agentic-patterns
 
 각 이동은 이전 시대의 실패가 촉발했습니다. 그 실패의 궤적을 추적해 보겠습니다.
 
-<a href="/assets/images/three-era-timeline.png" class="glightbox" data-gallery="agentic-patterns" data-glightbox="title: AI 에이전틱 패턴의 진화 타임라인">
+<a href="/assets/images/three-era-timeline.png" class="glightbox" data-gallery="post-images" data-glightbox="title: AI 에이전틱 패턴의 진화 타임라인">
   <img src="/assets/images/three-era-timeline.png" alt="AI 에이전틱 패턴의 진화 타임라인 — Prompt → Context → Harness Engineering" />
 </a>
 
@@ -285,7 +285,7 @@ Anthropic은 2025년 9월, "[Effective Context Engineering for AI Agents](https:
 
 KV-cache(Key-Value Cache)의 작동 원리를 설명하겠습니다. LLM API에 프롬프트를 보내면, 모델은 각 토큰의 어텐션 키(Key)와 값(Value)을 계산합니다. 이 계산은 비용이 큽니다. 그런데 이전 요청의 프롬프트 **접두어**가 현재 요청과 동일하면, 그 부분은 재계산할 필요 없이 캐시된 결과를 사용할 수 있습니다. Claude Sonnet 기준 캐시 히트 시 비용이 **10분의 1**로 줄어듭니다. 에이전트가 30번의 턴을 거치면서 매번 시스템 프롬프트를 재계산하는 것과, 캐시를 활용하는 것의 차이는 어마어마합니다.
 
-<a href="/assets/images/kv-cache-mechanism.png" class="glightbox" data-gallery="agentic-patterns" data-glightbox="title: KV-cache 작동 원리 — 안정 접두어가 비용을 결정한다">
+<a href="/assets/images/kv-cache-mechanism.png" class="glightbox" data-gallery="post-images" data-glightbox="title: KV-cache 작동 원리 — 안정 접두어가 비용을 결정한다">
   <img src="/assets/images/kv-cache-mechanism.png" alt="KV-cache 작동 원리 — 안정 접두어가 비용을 결정한다" />
 </a>
 

diff --git a/script/check-plugin-output.rb b/script/check-plugin-output.rb
@@ -3,16 +3,17 @@
 # Checks that the _plugins/ hooks that change a post's page left their mark,
 # that the search index parses and holds every post, that kramdown parsed the
 # post's block HTML, that no listing shows a translation beside its original,
-# that every cover is credited on its page, that no page repeats an id, and
-# that the feed has room for every post.
+# that every cover is credited on its page, that no page repeats an id, that
+# the feed has room for every post, and that no published URL went missing.
 #
 # test/ covers the plugins' functions, not the Jekyll hooks that apply them, and
 # with a hook gone the build still passes. Jekyll itself reads the posts here,
 # so which posts exist (published, future, folders, slug) is its answer, not a
 # second guess from the file names.
 #
-# Usage: bundle exec ruby script/check-plugin-output.rb <site_dir>
-# Prints one line per problem; exits 1 when there is any.
+# Usage: bundle exec ruby script/check-plugin-output.rb <site_dir> [--record]
+# Prints one line per problem; exits 1 when there is any. --record adds every
+# post's URL and redirect stub to script/published-urls.txt instead.
 
 require "jekyll"
 require "json"
@@ -45,6 +46,20 @@
 # of an old paper never reaches it.
 problems << "feed.posts_limit is #{limit} for #{posts.size} posts; raise it in _config.yml" if posts.size > limit
 
+# A post's URL comes from its `date:` and categories, so an edit to either moves
+# it, and legacy_urls.rb's stub moves with it: both indexed URLs then 404.
+# Every URL listed here must still be built, as a page or a redirect stub.
+published = File.join(root, "script/published-urls.txt")
+listed = File.file?(published) ? File.readlines(published, chomp: true).grep(%r{\A/}) : []
+if ARGV.include?("--record")
+  urls = posts.flat_map { |p| [p.url] + Array(p.data["redirect_from"]).map { |u| Jekyll::URL.escape_path(u) } }
+  File.write(published, File.read(published).lines.grep_v(%r{\A/}).join + (listed | urls).sort.join("\n") + "\n")
+  exit
+end
+listed.each do |url|
+  problems << "#{url}: published but no longer built; add it to the post's redirect_from" unless File.file?(built.call(url))
+end
+
 # search.json, with search_index.rb's plain_text: an index that does not parse
 # leaves js/search.js finding nothing, and saying nothing either.
 index = begin

diff --git a/script/published-urls.txt b/script/published-urls.txt
@@ -0,0 +1,68 @@
+# Every URL a post has been published under, so none goes missing when a post's
+# date or categories change (script/check-plugin-output.rb). After a post
+# merges, run: bundle exec ruby script/check-plugin-output.rb <site_dir> --record
+/insights/agentic-ai/2026/03/31/claude-code-architecture-analysis.html
+/insights/agentic-ai/2026/03/31/claude-code-source-map-leak-analysis-en.html
+/insights/agentic-ai/2026/03/31/claude-code-source-map-leak-analysis.html
+/insights/agentic-ai/2026/04/05/evolution-of-ai-agentic-patterns-en.html
+/insights/agentic-ai/2026/04/05/evolution-of-ai-agentic-patterns.html
+/insights/agentic-ai/2026/04/12/agentcore-harness-engineering-analysis.html
+/insights/data-architecture/2026/07/27/ai-ready-data-semantic-layer-knowledge-graph-en.html
+/insights/data-architecture/2026/07/27/ai-ready-data-semantic-layer-knowledge-graph.html
+/paper%20reviews/finetuning/2022/12/15/constitutional-ai-harmlessness-from-ai-feedback.html
+/paper%20reviews/finetuning/2023/05/23/qlora-efficient-finetuning-of-quantized-llms.html
+/paper%20reviews/finetuning/2023/05/29/direct-preference-optimization-your-language-model-is-secretly-a-reward-model.html
+/paper%20reviews/language-models/2023/02/27/llama-open-and-efficient-foundation-language-models.html
+/paper%20reviews/language-models/2023/07/18/llama-2-open-foundation-and-fine-tuned-chat-models.html
+/paper%20reviews/language-models/2023/10/10/mistral-7b.html
+/paper%20reviews/language-models/2024/01/05/deepseek-llm-scaling-open-source-language-models-with-longtermism.html
+/paper%20reviews/language-models/2024/01/08/mixtral-of-experts.html
+/paper%20reviews/language-models/2024/01/11/deepseekmoe-towards-ultimate-expert-specialization-in-mixture-of-experts-language-models.html
+/paper%20reviews/language-models/2024/01/25/deepseek-coder-when-the-large-language-model-meets-programming-the-rise-of-code-intelligence.html
+/paper%20reviews/language-models/2024/02/05/deepseekmath-pushing-the-limits-of-mathematical-reasoning-in-open-language-models.html
+/paper%20reviews/language-models/2024/05/07/deepseek-v2-a-strong-economical-and-efficient-mixture-of-experts-language-model.html
+/paper%20reviews/language-models/2024/07/31/gemma-2-improving-open-language-models-at-a-practical-size.html
+/paper%20reviews/language-models/2024/07/31/the-llama-3-herd-of-models.html
+/paper%20reviews/language-models/2024/11/22/tulu-3-pushing-frontiers-in-open-language-model-post-training.html
+/paper%20reviews/language-models/2024/12/27/deepseek-v3-technical-report.html
+/paper%20reviews/language-models/2025/01/22/deepseek-r1-incentivizing-reasoning-capability-in-llms-via-reinforcement-learning.html
+/paper%20reviews/language-models/2025/05/14/qwen3-technical-report.html
+/paper%20reviews/language-models/2025/07/28/kimi-k2-open-agentic-intelligence.html
+/paper%20reviews/language-models/2025/12/02/deepseek-v3.2-pushing-the-frontier-of-open-large-language-models.html
+/paper%20reviews/multimodal-learning/2023/12/19/gemini-a-family-of-highly-capable-multimodal-models.html
+/paper%20reviews/multimodal-learning/2024/10/09/pixtral-12b.html
+/paper%20reviews/multimodal-learning/2025/03/25/gemma-3-technical-report.html
+/paper%20reviews/multimodal-learning/2025/07/07/gemini-2.5-pushing-the-frontier-with-advanced-reasoning-multimodality-long-context-and-next-generation-agentic-capabilities.html
+/paper%20reviews/retrieval-augmented-generation/2024/01/31/raptor-recursive-abstractive-processing-for-tree-organized-retrieval.html
+/paper%20reviews/retrieval-augmented-generation/2024/04/24/from-local-to-global-a-graph-rag-approach-to-query-focused-summarization.html
+/paper%20reviews/retrieval-augmented-generation/2024/10/08/lightrag-simple-and-fast-retrieval-augmented-generation.html
+/paper%20reviews/retrieval-augmented-generation/2025/01/20/zep-a-temporal-knowledge-graph-architecture-for-agent-memory.html
+/paper-reviews/finetuning/2022/12/15/constitutional-ai-harmlessness-from-ai-feedback.html
+/paper-reviews/finetuning/2023/05/23/qlora-efficient-finetuning-of-quantized-llms.html
+/paper-reviews/finetuning/2023/05/29/direct-preference-optimization-your-language-model-is-secretly-a-reward-model.html
+/paper-reviews/language-models/2023/02/27/llama-open-and-efficient-foundation-language-models.html
+/paper-reviews/language-models/2023/07/18/llama-2-open-foundation-and-fine-tuned-chat-models.html
+/paper-reviews/language-models/2023/10/10/mistral-7b.html
+/paper-reviews/language-models/2024/01/05/deepseek-llm-scaling-open-source-language-models-with-longtermism.html
+/paper-reviews/language-models/2024/01/08/mixtral-of-experts.html
+/paper-reviews/language-models/2024/01/11/deepseekmoe-towards-ultimate-expert-specialization-in-mixture-of-experts-language-models.html
+/paper-reviews/language-models/2024/01/25/deepseek-coder-when-the-large-language-model-meets-programming-the-rise-of-code-intelligence.html
+/paper-reviews/language-models/2024/02/05/deepseekmath-pushing-the-limits-of-mathematical-reasoning-in-open-language-models.html
+/paper-reviews/language-models/2024/05/07/deepseek-v2-a-strong-economical-and-efficient-mixture-of-experts-language-model.html
+/paper-reviews/language-models/2024/07/31/gemma-2-improving-open-language-models-at-a-practical-size.html
+/paper-reviews/language-models/2024/07/31/the-llama-3-herd-of-models.html
+/paper-reviews/language-models/2024/11/22/tulu-3-pushing-frontiers-in-open-language-model-post-training.html
+/paper-reviews/language-models/2024/12/27/deepseek-v3-technical-report.html
+/paper-reviews/language-models/2025/01/22/deepseek-r1-incentivizing-reasoning-capability-in-llms-via-reinforcement-learning.html
+/paper-reviews/language-models/2025/05/14/qwen3-technical-report.html
+/paper-reviews/language-models/2025/07/28/kimi-k2-open-agentic-intelligence.html
+/paper-reviews/language-models/2025/12/02/deepseek-v3.2-pushing-the-frontier-of-open-large-language-models.html
+/paper-reviews/multimodal-learning/2023/12/19/gemini-a-family-of-highly-capable-multimodal-models.html
+/paper-reviews/multimodal-learning/2024/10/09/pixtral-12b.html
+/paper-reviews/multimodal-learning/2025/03/25/gemma-3-technical-report.html
+/paper-reviews/multimodal-learning/2025/07/07/gemini-2.5-pushing-the-frontier-with-advanced-reasoning-multimodality-long-context-and-next-generation-agentic-capabilities.html
+/paper-reviews/retrieval-augmented-generation/2024/01/31/raptor-recursive-abstractive-processing-for-tree-organized-retrieval.html
+/paper-reviews/retrieval-augmented-generation/2024/04/24/from-local-to-global-a-graph-rag-approach-to-query-focused-summarization.html
+/paper-reviews/retrieval-augmented-generation/2024/10/08/lightrag-simple-and-fast-retrieval-augmented-generation.html
+/paper-reviews/retrieval-augmented-generation/2025/01/20/zep-a-temporal-knowledge-graph-architecture-for-agent-memory.html
+/paper-summaries/agentic-ai/2025/11/26/toolorchestra-elevating-intelligence-via-efficient-model-and-tool-orchestration-summary.html