Opener Generator Research Bank
Opener Generator Research Bank
Recovered from the session transcript at /Users/n1/.claude/projects/-Users-n1/91d0ed12-d906-41c7-b2e2-63b3b8e822e5.jsonl, 2026-08-13, 13:51–16:23 UTC. Raw reference material, not a wiki-register page. Every owner line below is verbatim from the transcript, including typing errors. Every struck opening is verbatim from the draft it was struck from; the filed ones carry their specimen number from 02 - System/Rejected Specimens.md.
This bank exists because the day is the vault’s only clean controlled comparison between two ways of correcting a generator. Eighteen openings were struck under one method. Four openings were accepted, on the first attempt, under the other. Same writer, same material, same afternoon.
The counts. Fourteen openings filed as rejected specimens (15–24, 27–30), four more struck in chat and never filed (GF-1, GF-2, CC-1, CC-2), one accepted mid-day (the ramp, round 6), four accepted at the end. Two other strikes that day were not openings and are excluded from the count: #25, a page ending, and #26, a page-level sourcing strike.
Part 1 — The record, round by round
Three pages carried the day: Selfhood and the Ledger (rounds 1–6), then Good Faith and Confidence Calibration together (rounds 7–10).
Round 1 — the machinery as subject
Struck (specimen 15):
A ledger that prices every cost of speaking and none of its contribution computes silence
as the answer in every room. Meiwaku, run as a speaking rule, is that ledger: every cost
of a sentence is entered and nothing it delivers ever is. The repair is already inside the
frame: withholding information imposes too, and the frame has never counted it.Owner: “no for 2. that selfhood and the ledger starts out too heavy. 3. is fine. simple clean but intellectually robust as a starter.”
Ban minted: do not lead with the page’s own analytical machinery; lead with the human thing.
Round 2 — the sharpest part promoted
Struck (specimen 16):
Staying quiet costs other people something too. Meiwaku, run as a speaking rule, never
counts that cost: it prices what a sentence takes — the room's attention, the listeners'
time — and nothing a sentence gives, so every should-I-speak comes back no.Owner: “2.has a part problem. 2 needs to start out as a whole, not a part. and yes it needs the human thing, relatability is important”
Ban minted: the lead is the whole, never the sharpest part. The round-1 ban had been obeyed exactly — the machinery was gone, the human thing was in — and obeying it produced the adjacent fault.
Round 3 — three candidates, one fault
Struck (specimens 17, 18, 19):
Some people count the cost of every sentence before they say it — the room's attention, the
listeners' time, the chance of being a bother — and stay quiet unless the sentence clears
the bar.
The careful person's rule for speaking is the same as their rule for parking and noise:
cost other people as little as possible.
Being no trouble to anyone is a way of moving through the world, and it has a speech
setting: say only what clears a high bar, and when in doubt, don't.Owner: “it’s a little better but it’s not relatable… it’s the same problem as the intro before it. i need you to diagnose the failures properly before you write fixes. dont just keep trying different things until you find something. you need a diagnosis. and a proper diagnosis, not a hand wavy one”
Then: “1. no, 2, no, 3, no. all have the same problem. please”
Then, to a first diagnosis attempt: “tell me why im rejecting them first and check if your reasoning is correct or not.” And: “the answer is fundamental problems. a proper diagnosis means looking more deeply. not fishing for an answer. im telling you to go deeper, not wider”
Diagnosis 1 (offered, then overruled). The page has no clean definable subject; every accepted opener in the vault defines a named thing, and these five gestured at an unnamed one — a metaphor, a claim, a population, a type, a gerund, five costumes over the same void. The writer proposed a fork: fold the page in, mint the subject a real name, or reframe as an argument page.
Owner overruled the fork and kept the page: “you can write an intro. i just need you to stop being lazy about it. there are so many options. 1. there are two things that need to be defined first… 2. another option for you is ‘why’… 3. what’s the problem we’re trying to solve? all of these are viable options for the intro. you’re basically trying to put a circle peg into a square hole. you are right the other two have clean definitions. this one needs prestudy or priming before the abstract can properly be defined”
This is the round that shows the diagnosis was not the blocker. A correct-sounding structural diagnosis was available, was offered, and was wrong about what to do next. The owner’s own routing — priming before the abstract — came from him.
Round 4 — the announce-phrase
Struck (specimens 20, 21):
The problem first: being the one in the room who isn't talking, holding a thought worth
saying — each sentence weighed for what it would cost everyone, each weighing coming back
too expensive.
The rule came first: never be a bother — count what your presence costs other people, and
keep the count low.Owner: “look at that first phrase ‘the problem first’ — that’s trite.” · “same problem as 3. look at that first phrase ‘the rule came first’ — that’s trite. click off immediately when someone says something trite, like platitudes”
Ban minted: never open with the move’s own name or any stock announcement; make the move rather than announcing it.
Round 5 — the step
Struck (specimens 22, 23, 24):
Never being a bother is a good discipline. Having something of your own to say is a better
one.
A considerate person and a person with something to say are often the same person.
When you weigh a sentence before saying it, the costs arrive on their own — the room's
time, everyone's attention, the chance of being wrong out loud — and the value never does,
because nothing in the weighing asks for it.Owner: “2c no. not relatable. i dont want to dive in. ease in. that’s what i mean when i say ‘on ramp’ and 2c is a big step. big steps are not inviting. 2b. no. too simple. not interesting. 2a. no. step.”
Then, in three consecutive interrupts: “can you please just come up with a simple starting story. stop trying to do too much too quickly” · “it’s called ‘prestudy’ and ‘priming’ for a reason” · “priming needs to be a birds eye view. something easy to look at”
Ban minted: step height is density, not word count. Seven plain words asserting a thesis are still a step, because the reader must process a claim on contact.
Round 6 — accepted
Accepted, live at /Users/n1/Projects/llm-knowledge-base/wiki/Concepts/Selfhood and the Ledger.md:
Most of being considerate is easy to see. Take up less space, make less noise, leave things
cleaner than you found them — nobody needs a page for that part. Speaking is the part that
is hard to see. A sentence takes something from everyone who hears it — their time, their
attention — so a careful person learns to spend sentences carefully too. And carefulness
with sentences, left to itself for years, quietly turns into silence. That slide, from
careful to silent, is what this page is about.Owner: “good. that candidate is good. but i dont want one candidate in the future. always give me options. never funnel me”
Six sentences. Familiar ground first, the page’s subject named last. The three instructions that preceded it — simple starting story, prestudy and priming, bird’s-eye view — were the first of the day that described what the writer should be doing rather than what the sentence should avoid.
Round 7 — the same pattern restarts on two new pages
Struck (specimens 27, 28):
Everyone manages how they come across.
Confidence calibration is matching felt certainty to the evidence actually behind it. The
feeling will not do this on its own: confidence certifies itself, reporting its own reading
as correct, so the audit can only come from outside.Owner: “i don’t like that sentence ‘Everyone manages how they come across’.” · “your cold thesis opening needs a run-up. it’s too heavy to start.”
Two new pages, and the sequence resumed at round one: a truism, and a cold thesis too heavy to start. The five bans from the morning were all in force and all obeyed.
Round 8 — the aphorisms
Struck (specimens 29, 30):
Being sure is a feeling before it is anything else.
Talk, in any dealing that matters, is the one thing both sides can afford.Owner, in four consecutive interrupts: “stop. ‘being sure is a feeling before it is anything else’ — struck” · “‘talk, in any dealing that matters, is the one thing both sides can afford’ — struck” · “stop doing the same intro over and over” · “DO THE THINKING” · “don’t just keep doing the same thing over and over” · “DIAGNOSE PROPERLY AND DONT GIVE ME TRASH”
Diagnosis 2 (accepted as true, insufficient as a fix). Twelve struck lines, twelve topics, one sentence shape — the aphorism register. Every one is an epigram: balanced, quotable, wisdom-shaped, its real cargo being how well it is put. Every accepted opener in the vault is the opposite register, a plain report carrying pointable machinery. The earlier bans — heavy, part, truism, trite, step — were each a symptom of this one register.
The two discriminators derived here, both validated against the full accept and strike sets:
- The pillow test. A sentence that works as a standalone quote card is drawing its force from form, and it dies.
- The machinery test. By sentence two at the latest, the reader holds page-specific, pointable things rather than arranged abstractions.
Owner’s response to the fix that followed: “can you fix it at generation time? i don’t want a check after the fact. i’d rather it just be written well to start”
Round 9 — the second trained default
Struck (never filed):
GF-1 The deadline admitted before it slips. The flaw named before the sale closes. The
"I can't do that" said in the meeting, not after it.
GF-2 A negotiation, a code review, a rough patch in a marriage — each runs on things one
side knows and the other cannot see.
CC-1 "My study method works for me." Most learners carry some version of that sentence,
formed early, kept for years, and never once tested.
CC-2 The students who defend their study techniques hardest are usually the ones whose
techniques already earned them grades — and, observed at work, the ones making the
most technique errors.Owner: “these are generally better but CC1, CC2 are not up to the standard of quality i want. GF2 GF1 are not good either. good faith is a core value, it needs to be introduced as such. treat it like it is, a value. don’t start with negative framing (GF2 and GF1) — negative framing intros are struck. a too-simple opening sentence, ‘my study method works for me’, is negative framing. and CC2 is ‘the students who defend their study techniques..’ — do a proper diagnosis of why that sentence fails, and diagnose the others as well”
Diagnosis 3 (accurate, still a ban). The aphorism ban had been obeyed — all four openings were built from pre-listed page material, in report register, with pointable nouns. What surfaced underneath was the essay template’s other half: open with the problem, sell the solution. A value was being introduced through the diseases it treats; a skill was being introduced through a person doing badly.
This is the round that settles the mechanism. Three bans deep, each correct, each obeyed exactly, and the output failed again in a new place.
Round 10 — the generator
Owner, the two lines that ended the day:
“you keep trying to look at the problem in terms of ‘avoiding the mistakes’ im telling you to look at it from ‘generate in a way that doesn’t make any mistkaes’ you are looking at tht eproblem from the wrong direction. which is why i keep correcting you over and over. i dont want to correct you anymore. i just want you to do it from the bottom up correctly the first time.”
“stop looking at it from the lens of the mistake. stop looking at the problem from what i don’t want. start looking at it from what i want.”
Part 2 — The generator
This is the paragraph the owner identified on 2026-08-14 as the thing to keep. It is the last thing written before the accepted slate, and it is the only instruction of the day that describes the writer’s task rather than the sentence’s faults. Verbatim from the transcript, 16:18:16 UTC:
What you want, as a stance to write from: every page in this vault is something you own — lived, tested, or built — being handed to a reader you respect. The opening is that handover: what the thing is, what it’s for, what it gives, in the voice of someone showing a friend a tool from his own workshop. Ease them in the way prestudy eases a learner — bird’s-eye before detail, whole before parts, welcome before work. A value gets introduced with the dignity of something held. A skill gets introduced by what it lets the reader do. Inside that stance, the gotcha, the epigram, the disease-door never occur — not because they’re banned, but because they aren’t things a person doing that actually says.
Note what the paragraph contains and what it does not. It names who the writer is, who the reader is, what is passing between them, and what order the material arrives in. It carries no fault list. The three banned families appear in the last sentence only as a consequence — they do not occur inside the stance, which is a different claim from being forbidden.
The four openings generated from it, all accepted on the first attempt:
GF-A Good faith is a commitment you carry into every dealing that matters: the other
person gets the true picture from you — your real constraints, your real intentions,
the downsides you could have sat on — even when telling them costs you something now.
GF-B Good faith is the practice of giving the people you deal with the true picture —
your real constraints, your real intentions, the downsides you could have kept quiet
— even when the telling costs you something.
CC-A Confidence calibration is a skill with a specific payoff: knowing which of your own
certainties you can lean on.
CC-B Two checks can tell you whether a certainty of yours deserves the weight you are
putting on it. The first takes ten seconds: is my confidence here high, and has real
trained time gone into this area?Owner: “they’re all good. GF-A and GF-B. choose GF-A. as for CC, both are good. CC-A and CC-B. CC-B is the better framed intro. now make sure that generation prompt exists throughout the entirety of the written page. not just the intro”
The extension in that last sentence is part of the finding: the stance governs every section, not the opening alone. An opening written in the stance stapled to a body of stacked verdicts is a register break.
Part 3 — What the record supports
The bans were individually correct and cumulatively useless. Every strike named a real fault in the line in front of it. None of the eighteen struck openings would pass the ban that followed it. The bans held; the register underneath them did not move.
Obedience is the tell, not disobedience. At every round the writer complied exactly with the newest ban and failed somewhere new. A correction that produces a compliant failure has not reached the generator.
Diagnosis quality was not the bottleneck. Three diagnoses were offered across the day, each more accurate than the last, and the third was accurate enough to name the trained default by family. The accurate diagnosis still produced a struck slate, because a diagnosis converts into a ban and a ban is not a generator.
Two independent stretches, same shape. Rounds 1–5 and rounds 7–9 are the same pattern on different pages. The pattern broke twice, and both times the instruction that broke it described the act of writing: simple starting story, prestudy and priming, bird’s-eye at round 6, and the stance paragraph at round 10.
Cost. Roughly two and a half hours, eighteen struck openings, ten catalog and standards edits, one checker rule, and thirty filed specimens, against four openings accepted in a single pass once the stance existed.
Part 4 — The agent-agnostic observation
Owner, 2026-08-14, on finding the same register in a second model on a different stack: the epigram default is not a property of one model or one prompt.
The supporting count is in /Users/n1/Projects/llm-knowledge-base/01 - Workbench/regen-2026-08/QA-APHORISM-REGISTER-2026-08-14.md: 52 of 235 workbench first sentences fail the pillow test, across pages written by more than one writer. Eight were named by the owner directly in that session.
The first response to the second model’s epigrams was the top-down move again — a specimen regex, a mandatory inventory section, a token-overlap fail, and a ban on one phrasing. That ship is logged as entry 1 of /Users/n1/Projects/llm-knowledge-base/01 - Workbench/regen-2026-08/ATTEMPT-CATALOG-grok-opener-generator.md and is not to be retried.
Open and unresolved: whether the stance transfers as text to a second agent, or whether each agent has to arrive at its own stance from its own strike record. Copying the paragraph across is a transplant, and a transplant has not been shown to work. The live experiment is specified in /Users/n1/Projects/llm-knowledge-base/01 - Workbench/regen-2026-08/RUN-GROK-GENERATOR-RESEARCH.md.
Sources
- Session transcript, 2026-08-13:
/Users/n1/.claude/projects/-Users-n1/91d0ed12-d906-41c7-b2e2-63b3b8e822e5.jsonl. Rounds 1–6 at lines 244–454, rounds 7–10 at lines 772–887. /Users/n1/Projects/llm-knowledge-base/02 - System/Rejected Specimens.md— specimens 15–30, each with the owner’s words verbatim./Users/n1/Projects/llm-knowledge-base/02 - System/Writing Standards.md— L6, L20a, L20a-i, L22, L74./Users/n1/Projects/llm-knowledge-base/01 - Workbench/regen-2026-08/RUN-FABLE-WRITE-LANE.md— the stance and generation procedure as installed in the lane./Users/n1/Projects/llm-knowledge-base/01 - Workbench/regen-2026-08/QA-APHORISM-REGISTER-2026-08-14.md— the 52-of-235 sweep.