Grok needs its own aphorism gate (2026-08-14)

Claude’s 2026-08-13 stance works for Fable. It does not control Grok. Grok compresses the bank’s “operating claim” into a quote card, passes its own pillow test, and ships. The checker only counted words. That is why the class came back after we thought it was dead.

Owner named eight, then kept finding the same object. Fifty-two intros held, rewritten from inventory, promoted. The durable fix is not more stance prose in the write contract.

Correction, same day, owner: that gate is top-down. Fable was fixed by shifting the generator (chat voice is page voice; write from what he wants), after six or seven ban-iterations. The checker/inventory ship is the next trained default, not the Fable 1x. Discussion surface: /Users/n1/Projects/llm-knowledge-base/01 - Workbench/regen-2026-08/RUN-GROK-GENERATOR-RESEARCH.md. Do not grow the checker from here.

What had been installed (silent backstop only, not the fix):

• Live models: /Users/n1/Projects/llm-knowledge-base/01 - Workbench/regen-2026-08/GROK-OPENERS.md (struck → replacement).
• Assembly: droplog ## Inventory of five CARRY nouns before any page sentence; sentence one must use one. Operating claim is not sentence one.
• Checker: opener.aphorism_specimen (hard, validated on the eight named strikes / eight replacements) and opener.inventory (hard when a droplog exists).
• Skeletons may no longer write “the operating claim is [epigram].” They write s1-material: as five nouns.

Flip condition: if Grok starts stuffing inventory nouns into a still-shapely card (“The chair earns its place”), the inventory check is being gamed and the specimen list has to grow from a new named strike, not from reasoning.

Cost: one extra droplog section and a checker fail that Fable never needed. Cheaper than another 52-page rewrite.

Same day, four slates later. The working generator at /Users/n1/Projects/llm-knowledge-base/01 - Workbench/regen-2026-08/THE-GENERATOR-GROK.md added a stay-inside paragraph on top of the 16:18 stance: do not say what the page is doing; the name arrives when it is already obvious. That paragraph produced a third door — scene-walkthrough — that Fable never used. Fourth slate still opened “Look at 夫.” / “You want the thing.” Owner rejected it and handed Fable A/B: all good; two doors, not synonym swaps; definitional handover or payoff-or-ramp. GF-A is door A (“Good faith is a commitment you carry…”). Fable writes these intros. Catalog: /Users/n1/Projects/llm-knowledge-base/01 - Workbench/regen-2026-08/ATTEMPT-CATALOG-grok-opener-generator.md entries 2–4. Do not retry stay-inside.

Owner picks, same day, spliced live. Attention A, Decisional Delays A, How Chinese Characters Work A, Expectancy B, Environment A, Wanting Less A, Exit A, Empty A. Locked at /Users/n1/Projects/llm-knowledge-base/01 - Workbench/regen-2026-08/REPLACEMENTS-FROM-GENERATOR.md. Grok versions of those eight openers sit in /Users/n1/Projects/llm-knowledge-base/_archive/wiki-pre-fable-openers-2026-08-14/.

Pass 2, same day, replacement intros only. Owner: check aphorisms in the regen intros, because those are the replacements. Twelve promoted replacements still fail the pillow test (Reverse Causality, Techniques in School, Almanack of Naval, Deep Processing, Time Management hub, Declarative/Procedural/Conditional, Dimension Practice Tracks, Group Study, Writing with a Structure Engine, Ownership Cost, Four Things People Stay For, Learning Styles). Eight Fable picks pass. Held-cluster drafts listed separately and not promoted. Score: /Users/n1/Projects/llm-knowledge-base/01 - Workbench/regen-2026-08/QA-APHORISM-REGEN-INTROS-PASS2.md.

Opus 13, same day. Fable out of tokens. Owner: “these are also good quality.” Picks spliced onto workbench drafts only — Best-attempt B, Exam Technique A, AI-Assisted B, Reading & Retention A, Rules of Effective Memorization B, Rapid Skill A, Burnout A, Tonal Modulation A, Moral Core A, Iceberg A, Premise A, Setting A, Low Point A. Cluster holds stand. Do not promote. Lock: /Users/n1/Projects/llm-knowledge-base/01 - Workbench/regen-2026-08/REPLACEMENTS-OPUS-13.md.

Opus 12, same day, all Opus. Owner compared Grok A/B vs Opus A/B. Picks spliced onto already-promoted pages: Reverse Causality A, Techniques in School B, Almanack A, Deep Processing A, Deep Processing for Research A, Time Management A, Declarative A, Dimension Practice Tracks A, Group Study A, Writing with a Structure Engine B, Ownership Cost A, Learning Styles A. Grok slate held unused. Lock: /Users/n1/Projects/llm-knowledge-base/01 - Workbench/regen-2026-08/REPLACEMENTS-OPUS-12.md. Prior Grok-fail openers: /Users/n1/Projects/llm-knowledge-base/_archive/wiki-pre-opus-12-2026-08-14/.

Opus 24 held, 2026-08-15. Owner picked A/B (mostly A) on the remaining Story Craft / Learning Craft drafts. Clinical Learning System used an owner rewrite (pasted under the Microlearning label; text is the ward-as-syllabus clinical intro). Microlearning deferred for redo, then owner supplied the rewrite the same day — spliced onto the workbench draft. Lock: /Users/n1/Projects/llm-knowledge-base/01 - Workbench/regen-2026-08/REPLACEMENTS-OPUS-24.md.

37 held drafts promoted to live, 2026-08-15 15:52, on explicit owner accept. Story Craft, Learning Craft, Best-attempt Encoding, Exam Technique. LLM Knowledge Systems was briefly promoted in the same loop (not in the 37) and reverted the same minute. Meiwaku F2 and Grok Bot have no live original and stayed on the workbench.