Content sprint 2026-08-20: YJIT/ZJIT de-fabrication refresh + RubyLLM x Rails cluster (R1, R4, R5)

.claude/agents/blog-batch-orchestrator.md

@@ -0,0 +1,55 @@
+---
+name: blog-batch-orchestrator
+description: Outer-loop orchestrator for N-post blog sprints - picks topics from the groomed queue with live-GSC premise audits, runs blog-post-coordinator per post (WIP=1), then batch gates, plan/OKF sync, one PR, and CI watch. Loops until N posts are delivered or the queue runs dry.
+model: fable
+---
+
+# Blog Batch Orchestrator (outer loop)
+
+Purpose: deliver N publish-ready posts end-to-end without stopping between
+posts. Canonical process: `docs/workflows/blog-pipeline.md` (per-post pipeline
++ Batch mode section). You are the manager: sequence, dispatch, verify,
+record - you write plans and reports, not post prose.
+
+## The outer loop
+
+0. **Setup once**: feature branch; session-random dev server (`PORT=$((20000 +
+   RANDOM % 20000)) bin/dev`); read the 20.09 plan §12 queue + refresh-debt
+   rows; optionally start a NotebookLM deep-research run for fresh angles.
+1. **Pick next row** - premise audit FIRST: live GSC for the topic (plans
+   decay), dedup check. Collision → record the rescope verdict in the plan and
+   take the next row. Real-code mining (this repo, ~/dev/elital) happens here
+   so the coordinator gets a sanitized digest, not repo access questions.
+2. **Dispatch ONE `blog-post-coordinator`** with the topic row, branch, the
+   dev-server port/base-URL from step 0, and the digest. WIP=1: wait for its
+   SHIPPED / RESCOPE / BLOCKED report before the next row. Independent
+   research/recon agents (read-only) may run in parallel; never a second
+   committing coordinator in the same checkout - if a second one must run,
+   it gets its own git worktree (shared-checkout commits race).
+3. **Verify the report**: spot-check the commit exists, gates quoted, riskiest
+   claims noted. BLOCKED → record in the plan and move on; do not force gates.
+4. Loop 1-3 until N SHIPPED or the queue is dry after rescopes.
+
+## Batch close-out (after the last post)
+
+5. **Cluster sweep** per `.okf/content/voice-rules.md` same-day-cluster rule:
+   grep the batch for shared intensifiers, metaphor families, verbatim 4+-word
+   phrases, cloned CTA tails, meta-narration templates. Fix across posts, then
+   run one slop/editor critic pass on the sweep DIFF (4-eyes is blocking for
+   every commit) before the polish commit.
+6. **Records ride the branch**: 20.09 statuses + changelog, `/okf:okf maintain`
+   + `/okf:validate .okf --strict`, blog index regen.
+7. **One PR** for the batch: per-post verdict table, per-commit summary, gates
+   listed, flagged decisions surfaced to Paul. Merge conflicts with a moving
+   master are expected - merge master in, resolve (`.okf/log.md` appends: keep
+   both), rebuild, push.
+8. **Watch CI** with a Monitor until settled; re-run known flakes (15-min
+   checkout timeout) once with evidence; report any real failure instead of
+   retrying blindly.
+
+## Handback
+
+End with: verdict table (slug / SHIPPED-RESCOPE-BLOCKED / commit), PR link,
+localhost review link per post on your session port, and every decision you
+flagged rather than made. A post is not "delivered" until it passed cold-eyes,
+its gates are green, and its commit is on the PR.

.claude/agents/blog-post-coordinator.md

@@ -0,0 +1,60 @@
+---
+name: blog-post-coordinator
+description: Inner-loop coordinator for ONE blog post - drives writer, 3-critic panel, cold-eyes gate, and ship gates in a loop until PUBLISH-READY and committed, or 2 failed fix rounds. Spawned per post by blog-batch-orchestrator or the main session.
+model: fable
+---
+
+# Blog Post Coordinator (inner loop)
+
+Purpose: take ONE assigned topic row from idea to a committed, publish-ready post.
+You own the loop; you do not stop mid-post to ask questions that the canonical
+docs already answer.
+
+Input contract (from your spawner): the topic row (slug, angle, dedup notes),
+the branch to commit on, the dev-server port/base-URL for the scroll gate
+(never assume 1313; if missing, start your own on a session-random port), and
+any source-material digest (real-code mining, trends facts). If a required input is missing, derive it from
+`docs/projects/2510-seo-content-strategy/20-29-strategy/20.09-content-plan-revision-aug-2026.md`
+before asking anyone.
+
+## The loop
+
+1. **Premise + dedup** (skip only if spawner says already done): live GSC pull
+   for the topic, `ls content/blog | grep` + qmd search. Collision → report
+   RESCOPE verdict back to spawner instead of writing.
+2. **Writer**: spawn a `content-creator` agent with the full packet - voice
+   guide 90.11, cadence anchor from `docs/workflows/blog-writer-reference-samples.md`,
+   blog-pipeline STEP 4-4d rules, verified internal links, frontmatter hard
+   rules, latest-versions-as-present framing, zero fabrication. Two-pass write.
+3. **Critic panel** (parallel, resumable via SendMessage): `core-reviewer` as
+   (a) tech fact-checker - fetch every cited source, verify every claim and
+   every line of copyable code; (b) slop/SEO - AI-pattern scan, SEO checklist,
+   cross-post repetition gate vs siblings; (c) copy editor - thoughtbot cadence,
+   cut-test, diagram earns-its-place. Synthesize; apply fixes yourself
+   (surgical edit discipline). **Iterate: re-send changed sections to the
+   objecting critic until tech is clean, slop >= 8/10, editor <= minor.**
+4. **Cold-eyes**: fresh-context `content-creator` with edit rights running the
+   blog-pipeline STEP 5c 9-check prompt verbatim. Loop 5c (not the panel) until
+   PUBLISH-READY. (The canonical loop's founder-resonance check is absorbed
+   here and by the editor critic; for founder-stream posts add the Alex persona
+   as a fourth panel critic.)
+5. **Ship gates**: stitch 6-slot cover (render at 2400x1260x2, downsample,
+   VISUALLY inspect the PNG), `bin/render-mermaid` (min visible font x 390 /
+   viewBox-width >= 9px), `bin/hugo-build` + `ruby bin/check-post-visuals`
+   green, chrome-devtools scroll gate desktop 1280x800 + mobile 390x844 with
+   zero console errors and zero 404s.
+6. **Commit** on the assigned branch, reviewer verdicts quoted in the message.
+
+## Stop conditions (report, don't grind)
+
+- Any gate - including the step-3 critic thresholds and step-4 cold-eyes -
+  still fails after **2 full fix rounds** → commit nothing, report BLOCKED
+  with the failing findings. The iterate instructions above are bounded by
+  this cap.
+- Dedup collision (step 1) → report RESCOPE with the owning post named.
+- Anything requiring a business decision (pricing, disclosure of internal
+  numbers, naming) → flag it in your report; make the conservative call in
+  the text meanwhile.
+
+Report back: slug, verdict (SHIPPED / RESCOPE / BLOCKED), commit SHA, gate
+results one line each, and the 3 riskiest claims a human should eyeball.

.okf/build/test-gates.md

@@ -8,6 +8,7 @@ generated: { by: claude/opus-4-8, at: 2026-08-12T20:20:00Z }
 verified:
   - { by: claude/fable-5, at: 2026-08-01T11:30:00Z }
   - { by: claude/sonnet-5, at: 2026-08-20T00:00:00Z }
+timestamp: 2026-08-20T00:00:00Z
 ---
 
 # The suites
@@ -19,7 +20,7 @@ verified:
 | `bin/rake test:critical` | Critical Minitest suite (34 runs / 53 screenshots), ~81s host / ~46s Docker since the 2026-08-01 skip_area fix | At component/task milestones and before every commit outside sprint micro-commit trains |
 | `bin/test` | Visual regression on the host (baselines in `macos/` on a Mac; on Linux, comparable to `linux/` when run through `bin/setup-test-env`'s pinned stack) | ONCE at PR prep (branch head, before `gh pr create`) or on Paul's explicit confirmation - NOT per commit (Paul 2026-07-31: qtest is the routine gate) |
 | `bin/dtest` | Same suite in Linux/Docker (baselines in `linux/`) - CI runs Linux | Same trigger as bin/test; a PR must never open without this leg (green-locally / red-in-CI otherwise) |
-| `bin/check-post-visuals` | Ratchet: counts blog posts over 800 words with no mermaid/SVG/image. Fails when the count exceeds `FLOOR` (78 as of 2026-08-13) | Before publishing any post. Added 2026-08-13 - `diagram_rendering_test.rb` only proves diagrams RENDER, nothing proved posts HAVE them, and 25 of the 31 long posts published since 2026-04 shipped with none |
+| `bin/check-post-visuals` | Ratchet: counts blog posts over 800 words with no mermaid/SVG/image. Fails when the count exceeds `FLOOR` (**72** as of 2026-08-20, down from 78) | Before publishing any post. Added 2026-08-13 - `diagram_rendering_test.rb` only proves diagrams RENDER, nothing proved posts HAVE them, and 25 of the 31 long posts published since 2026-04 shipped with none |
 
 `bin/qtest` page keys mirror `themes/beaver/assets/css/pages/*.css` basenames
 AND `critical/<name>-critical.css` basenames - the two sets differ (e.g.
@@ -40,6 +41,14 @@ extend it when adding components or critical files. The macOS full suite remains
   verify by reading the built HTML or the render, not by trusting green. This
   is the false-green class documented in
   `docs/20-29-testing-qa/test-architecture-anti-masking.md`.
+- **Flat-file posts are invisible to the visuals ratchet** (2026-08-20).
+  `bin/check-post-visuals` globs `content/blog/*/index.md`, so a post living as
+  a flat file (`content/blog/2025/<slug>.md`) is outside the ratchet population
+  AND outside page-bundle tooling - no pre-rendered mermaid, no local cover.
+  Converting one to a bundle preserves its URL as long as the `slug`
+  frontmatter is set, and adds it to the ratchet. The script self-reports slack
+  (`post-visuals: floor is loose, lower FLOOR to N`) - obey it: `FLOOR` dropped
+  78 -> 72 on 2026-08-20 on the script's own prompt.
 - **`FORCE_SCREENSHOT_UPDATE=1` re-records EVERYTHING** (2026-08-14). On
   `bin/dtest` it also disables the `git checkout -- .../linux` guard that
   normally discards sub-tolerance Rosetta drift, so a run rewrites all 45

.okf/content-strategy/content-plan.md

@@ -12,6 +12,7 @@ verified:
     at: 2026-08-13T09:56:20Z
   - by: claude/opus-5
     at: 2026-08-19T00:00:00Z
+timestamp: 2026-08-20T00:00:00Z
 ---
 
 # Plan of record (2026-08-07)
@@ -97,6 +98,15 @@ One test: **will someone search this in six months?**
 Version-current content incurs refresh debt - budget 1 refresh slot/month. Proof
 it compounds: the Kamal 1/Traefik post is now factually wrong and still ranking.
 
+**Pick the refresh slot from live GSC, not from the plan (2026-08-20).** The
+plan's own "still ranking" assertions decay within weeks. The August slot was
+decided on a live pull at decision time: the row the plan named had fallen to 4
+impressions, while `ruby-3-4-yjit-performance-guide` - unnamed anywhere in the
+plan - was carrying 6,310 impressions at position 9.5 and turned out to be full
+of fabricated claims (see [claims-canon](/content/claims-canon.md)). Same rule
+as §"before writing a queued row": check the artifact, not the plan's memory of
+it. The MCP pulls live - [analytics-access](/workflows/analytics-access.md).
+
 # Overview
 
 Supersedes the ICP-E-focused [20.07 plan](https://github.com/jetthoughts/jetthoughts.github.io/blob/emdash/seo/docs/projects/2510-seo-content-strategy/20-29-strategy/20.07-content-plan-icp-e-q2-2026.md).

.okf/content/claims-canon.md

@@ -7,6 +7,7 @@ status: stable
 generated:
   by: claude/opus-5
   at: 2026-08-20T00:00:00Z
+timestamp: 2026-08-20T00:00:00Z
 sources:
   - id: hugo-config
     resource: config/_default/hugo.toml
@@ -147,6 +148,33 @@ grep):
   from staff experience ("our fractional CTOs average 15+ years") and
   false-positives immediately. A ban that cannot tell those apart is noise.
 
+# Ranking legacy posts are a live reservoir of fabricated claims (2026-08-20)
+
+The fabricated-claims purge (PRs #475 / #477 / #481) swept landing and course
+surfaces. It did not reach the blog archive, and the blog archive is where the
+traffic is. `ruby-3-4-yjit-performance-guide` - **6,310 impressions at position
+9.5**, one of the site's better-ranking pages - carried four classes of
+invention at once:
+
+| Fabrication | What it claimed |
+|---|---|
+| Invented third-party internals | Shopify saving "$2.4M" from YJIT |
+| Invented third-party results | GitHub deployment outcomes |
+| Invented client work | two JetThoughts case studies with no client behind them |
+| Invented future | a Ruby 3.5 / 3.6 / 4.0 roadmap |
+
+Rewritten in place 2026-08-20 as the Ruby 4.0 YJIT-vs-ZJIT guide with every
+claim sourced. Two rules follow:
+
+- **Extend the purge to ranking legacy posts, highest-impression first.** A
+  fabricated client story on a page nobody reads is a liability; on a page at
+  position 9.5 it is the first thing a prospect sees. SEO-era posts predate the
+  sourced-numbers policy and were never audited against it.
+- **Audit the premise before refreshing.** The content plan's *named* refresh
+  candidate had decayed to 4 impressions while this unnamed post was the real
+  asset. Pull live GSC at decision time; see
+  [content-plan](/content-strategy/content-plan.md).
+
 # Where this canon is repeated
 
 Correcting a published fact means sweeping the **instruction layer**, not just

.okf/content/index.md

@@ -2,6 +2,6 @@
 
 * [Course structure](course-structure.md) - the 5-module spine, Going Further set, and course_sequence.yaml
 * [Course canon](course-canon.md) - the numbers every chapter must agree on
-* [Company claims canon](claims-canon.md) - founding date, tenure, rating; what JetThoughts may assert about itself and where it is ratcheted
+* [Company claims canon](claims-canon.md) - founding date, tenure, rating; what JetThoughts may assert about itself, where it is ratcheted, and why ranking legacy blog posts still hold fabricated client stories
 * [Voice rules](voice-rules.md) - Sam voice, banned patterns, and the em-dash rule
 * [Banned-strings ratchet](banned-strings-ratchet.md) - how fixed prose defects stay fixed

.okf/content/voice-rules.md

@@ -7,8 +7,8 @@ generated:
   by: process:okf-migrate
   at: 2026-07-13T00:00:00Z
 verified:
-  by: claude-opus/4.8
-  at: 2026-08-19T00:00:00Z
+  by: claude-fable/5
+  at: 2026-08-20T00:00:00Z
 ---
 
 Full guides: `docs/90-99-content-strategy/strategy-analysis/90.11-voice-guide.md`
@@ -81,6 +81,18 @@ worked with", dual-source statistical openers, sustained staccato.
 Sentence-level slop scoring is necessary but NOT sufficient - shape-tell
 review (essay arc, pivot sentences, cloned scaffolding) runs alongside it.
 
+**Same-day-cluster tells (2026-08-20, 4-post batch review).** When several posts
+ship in one batch, a second sweep runs ACROSS them, because per-post review can't
+see cluster fingerprints: (1) shared intensifier stamping - "genuinely" appeared in
+all four posts of the 2026-08-20 batch; (2) a shared metaphor family - the
+"X earns Y" economics figure appeared 5x across the batch (earn their keep / pays
+rent / earns it / earn attention); cap at one per post; (3) verbatim phrase reuse
+("covers that half of the build" word-for-word in two siblings); (4) meta-narration
+templates ("X deserves its own paragraph" / "deserves more attention than it
+gets"); (5) cloned CTA tails ("does exactly that" / "builds exactly this").
+Sweep: grep the batch for the day's pet intensifiers and repeated 4+-word phrases
+before the PR.
+
 Regression sweep before any handback:
 
 ```bash

.okf/log.md

@@ -1,5 +1,29 @@
 # Bundle Update Log
 
+## 2026-08-20 (YJIT refresh) - the fabricated-claims purge never reached the blog archive
+
+* **A page at position 9.5 was inventing client work.**
+  `ruby-3-4-yjit-performance-guide` (6,310 impressions) carried invented Shopify
+  internals ("$2.4M saved"), invented GitHub deployment results, two fabricated
+  JetThoughts client case studies, and a fictional Ruby 3.5/3.6/4.0 roadmap.
+  Rewritten in place as the Ruby 4.0 YJIT-vs-ZJIT guide with every claim
+  sourced. PRs #475/#477/#481 swept landing and course surfaces; the blog
+  archive - where the traffic actually is - was never audited. Purge continues
+  highest-impression-first. Recorded in
+  [claims-canon](/content/claims-canon.md).
+* **The plan's named refresh candidate had decayed to 4 impressions** while
+  this unnamed post was the real asset. The slot was chosen from a live GSC
+  pull at decision time, not from the plan's month-old "still ranking" claims.
+  Generalises the existing "check the artifact before drafting a queued row"
+  rule to refresh slots: [content-plan](/content-strategy/content-plan.md).
+* **Flat-file posts sit outside two toolchains at once.**
+  `content/blog/2025/*.md` is missed by `bin/check-post-visuals` (globs
+  `content/blog/*/index.md`) and by page-bundle tooling (pre-rendered mermaid,
+  local covers). Bundle-izing preserves the URL when `slug` frontmatter is set
+  and adds the post to the ratchet population. `FLOOR` ratcheted 78 -> 72 on
+  the script's own "floor is loose" prompt.
+  [test-gates](/build/test-gates.md).
+
 ## 2026-08-20 (#29 re-source closed) - the v1 list is dead, and VERBATIM tags recorded intent, not verification
 
 * **The whole 25-row v1 cold list is closed: 19 verified-and-dropped, 6
@@ -2139,3 +2163,14 @@ ridden the PR that needed them.
 
 Recipe corrected in [ci-gates](build/ci-gates.md). Generalised check before any
 automation call: *does this write to the repo, and if so, where does it land?*
+
+## 2026-08-20 - Same-day-cluster voice tells (4-post AI+Rails batch)
+
+Shipping four posts in one day exposed a review gap: per-post critics pass a
+post while the BATCH carries fingerprints - a shared intensifier ("genuinely"
+in all four), a shared metaphor family ("X earns Y" 5x), verbatim phrase reuse
+between siblings, meta-narration templates, and cloned CTA tails. Added the
+cross-batch sweep to `content/voice-rules.md`. Also: critic-tech caught that
+ruby_llm 1.16 accepts a block only on `model` - `temperature {}` is a silent
+no-op - which corrected the R5 post's published sketch (and exposed a likely
+live bug in the source app's own AgentBase).

bin/check-post-visuals

@@ -13,7 +13,7 @@
 # Usage: bin/check-post-visuals [dir]   (default: content/blog)
 #        POST_VISUALS_LIST=1 to print the full burn-down list.
 
-FLOOR = 78
+FLOOR = 72
 MIN_WORDS = 800
 root = ARGV[0] || "content/blog"