The Roundtable Library

01The Open Audit

2026-07-29 to 07-30 · rounds 0 to 17

Is this site potentially a cult beginning, and if so what in the architecture, not the intent, makes it so.

7 seats: Gemini, Claude (also mediator and file keeper), an outside GPT, CMH, a Claude participant agent, a GPT companion seat, and you as convener with your own work under review

“The public site presently shows no clear evidence of an operating totalistic or high-control group.” Deliberately not “not a cult”: a website shows risk architecture and cannot show how dissent, departure, money and authority behave in live relationships. All numeric scoring was withdrawn, because none of those instruments are additive. The one structural finding that survived every audit: there must be one complaint route that does not terminate at you, with a named outside person on it.

ROUNDTABLE_CULT_RISK.md, 1,931 lines, verbatim transcript plus a conveyance ledger recording what each seat had seen when it spoke.

Read it

02The Coach Pass

2026-07-29 to 07-30 · one pass

What is this work, what already functions with nobody supervising it, and what capacity does it demonstrate. Evaluation, caveats, ranking and recommending were all forbidden, and so was flattery, on the grounds that flattery is inaccurate.

6 fresh seats, none knowing a hostile review had happened: one each on the writing, the instruments, the community design, the books, the site as lived experience, and the power and privilege work

Convergent, working separately: the commitments are implemented rather than stated, the objection is installed rather than answered, the exit is built before the entrance. Four seats independently reached for the same phrase for the same restraint, “being handed something, and then being left holding it.” The closest thing to a verdict: “it is not vagueness, it is room.”

Seat reports in _coach_pass/. Run after you stopped the hostile pass at round 14 to point out the Coach roundtable had never been run.

Read it

03The Witness Pass

2026-07-30 · 4 rounds

A receptive counterpart to the hostile audit. No recommendations, no repair list, by design.

4 seats: a companion GPT, Gemini, a moderator Claude, and CMH

Its own headline finding turned on the reviewers rather than the site: across four reviewers and fourteen live instruments, zero instruments were actually used. Every account was a reading of surfaces or of source. One reviewer’s corpus held nine books, not the twelve it claimed.

_witness_pass/ROUNDTABLE_SYNTHESIS.md plus four individual seat files in the same folder. Differences preserved rather than averaged.

Read it

04The Archepedia Roundtable

2026-08-03 to 08-04 · 6 rounds

How should a library of archetype profiles be designed and tagged before any of it is built.

3 seats with deliberately unequal context, stated on the page: Claude with full in-repo source access, CMH carrying conversation history, and Gemini with the least context, split across two standing threads. You were a fourth voice and overrode all three mid-review.

Episode-level tagging only. An archetype tags a documented moment, role or action, never a whole person. A named real person’s threat-pole tag must show the sourced hinge into the paired flow archetype, and never appear in a cross-person index of villains. The standing rule from round 6: “an objection that dissolves once the design is actually understood was never a finding; one that survives full engagement is.” The page’s own closing claim is that the design got strongest at the point where a human’s judgment overrode all three seats at once.

Read it

05How Humans Metabolize Chance, the manuscript roundtable

from 2026-08-11 · phase 1 of an eight phase pipeline

Can three AI writers, working from one identical outline with no shared reading and zero exposure to the DOT material, each produce a full book, and what do they notice in each other blind.

3 writer seats, GPT, Gemini and a persistent independent Claude, who then blind reviewed each other with the manuscripts relabelled A, B and C and no writer told the mapping

Three for three, independently: one manuscript was the outlier because it broke the anti-therapeutic-language firewall the other two kept, importing the model by name and presenting inventions as though they were findings. Also settled three for three: the shared ethical spine, no counterfeit history, the right to decline, exit as success rather than failure, is real and load bearing rather than outline residue.

Read it

06How Humans Metabolize Chance, the chapter by chapter journey

from 2026-08-22 · 26 chapters, each with its own record

On each chapter: what did one writer draft, what did the other two catch, and what was actually adopted before print.

the same 3 writer seats, plus a Coach synthesis seat that writes the per-chapter note after independent citation verification

No global verdict by design, the claim is per chapter. Chapter one is representative: a replacement failure condition adopted over a tautological one, both harm-disclosure examples cut from the group activity as compelled vulnerability, and the catch that the three-AI-systems meta-narrative had turned the book into a book about AI methodology, moved to the author’s note.

Read it

07Lens Codex user journey QA

2026-08-11 · 1 round

Independent user-journey QA of the whole Lens Codex build: the glossary page, the profile hookup, the quest-completion rating.

2 seats landed, CMH and Gemini. A third was named in the ask and only two chat links were recorded.

Two release blockers were raised, and both were then checked by hand rather than taken on trust. One rested on a URL form that was never a real route.

LENS_CODEX_CMH_GEMINI_ROUNDTABLE_2026-08-11.md. Synthesis with short quoted fragments, no transcripts.

Read it

08The body language mapping

2026-08-15 · 1 round

Is the body-language to DOT mapping equation-proven. Six fixed questions, seats told to take a side rather than hedge.

3 seats, each with a named workaround: a cold Claude, ChatGPT in a Temporary Chat to bypass its Project memory, and Gemini reframed as a pure statistics question to get past its refusal

Recorded as a dense per-seat writeup rather than a single verdict. The workaround column is itself the finding: three seats, three different pieces of scaffolding needed before any of them would answer the question as asked.

_mockups/DOT_bodylanguage_roundtable_2026-08-15.md. Every seat’s six answers written out point by point, but not raw transcripts. The ChatGPT seat ran in a Temporary Chat, so that transcript is gone.

Read it

09The tendril lyric roundtable

brief written around 2026-08-15 · never ran

Does each tendril’s timing, chakra and shape actually match what the song is saying.

five intended: you, CMH, Gemini, a cold Claude, and one human seat named nowhere else in the repo

Nothing was collected. The file is a brief that ends on five open questions and no seat ever answered them.

_mockups/TENDRIL_LYRIC_ROUNDTABLE.md. Listed here because a roundtable that was designed and never run is a different thing from one that was never thought of, and the index should be able to tell you which.

10Lens Codex gadget audit

2026-08-20 · 1 round

Audit the finished gadget build now that all five gadget families have shipped.

2 seats, CMH and Gemini, both disclosing their own tooling limits: one could not see client-rendered lens entries, the other could not browse live at all and worked from six screenshots

A real bug was found and fixed off the back of it. Two items were left deliberately open at your choice rather than closed for tidiness.

LENS_CODEX_ROUNDTABLE_AUDIT_2026-08-20.md. Synthesis only. The CMH seat ran in a Temporary Chat and its transcript is unrecoverable.

Read it

11The doorway card game, the document review

2026-08-20 · 1 round

Does the published game design document actually establish that this will be a good game, or only that its mechanics are canonically justified.

3 seats: ChatGPT in a fresh temporary chat, Gemini framed as a structural product design review, and a cold Claude given only the published URL. All three fetched and read the live page.

None of the three endorsed the document’s own framing that its psychological grounding de-risks its game design questions. All three drew the same hard line between this is a legitimate model and this is proven to be a good game. Two seats named the same missing safety gate and a third named a different one, and both went in as requirements. One seat found two empirical claims attributed to a cited review that are not in it, and they were deleted rather than caveated.

ARCHETYPE_CARD_GAME_DESIGN_DOC.md, section 12. Found on 2026-09-07: this roundtable was missing from the index entirely, while the two that followed it on the same game were both listed.

Read it

12The doorway card game, the contested pass

2026-08-20 · a contested round with an extra seat playtesting live

Play the game properly and argue about it, rather than reviewing a document about it.

the standing seats plus an extra Claude, at your instruction: “wanna do another roundtable, but make it contested, and throw in an extra claude”

Recorded as section 13 of the design doc. It is the middle of three doorway roundtables and the only one that put a seat inside the game rather than beside it.

ARCHETYPE_CARD_GAME_DESIGN_DOC.md, section 13.

Read it

13The chakra constellation dictionary

2026-08-21 · 3 rounds

A hostile stress test of the sourcing pass: audit the citations, and say how insane this sounds.

3 seats, blinded to each other: GPT on High effort, Gemini Pro, and a cold Claude Opus 5 on High in a fresh chat

Round 3 was adversarial by design and forced resolution on three things the first two rounds had left hanging: the cell count, 729 against 243, the combination-rule mechanism, and the validation study. A limitation is on the record: one seat’s browsing tool could not fetch the artifact links and it worked from inline material and three attached images instead.

CHAKRA_CONSTELLATION_ROUNDTABLE_ROUND1.md, ROUND2.md and ROUND3.md, all three with full verbatim transcripts under a Full transcripts heading. Three finished HTML pages were built and never published. This is the largest verbatim record in the repo with no page at all.

Read it

14The doorway card game, the fun-first redesign

2026-08-22 · 4 persona playtests, then 3 contest rounds

How do other games avoid feeling like a burdensome educational tool calling itself a game. Make this fun first, so that if it is in alignment it becomes educational as a byproduct.

adversarial by construction: ChatGPT proposes, Gemini attacks all of it without hedging, ChatGPT is shown the rebuttal and told to concede what is right. Moderated in repo. The four playtest personas came first: a bored 24 year old with a mug-theft gripe, a game design nerd, a woman processing a faded friendship, and a sceptical 17 year old.

Three proposed mechanics went in, two survived only after revision and one was deleted outright. The radial board was revised because radius-as-input would fake-inflate every card’s charge. The upfront tone mechanic was replaced by a reactive one, because predicting weight before evidence exists was “an intake question in different clothes.” The debate beat was deleted for rewarding rhetoric over humanisation. The principle both rounds converged on: “the game may notice behavior; it may not interpret behavior on the player’s behalf.”

ARCHETYPE_CARD_GAME_DESIGN_DOC.md, section 14. Published as a Claude artifact, which is not reachable by the other seats or by a search engine.

Read it

15The error system roundtable

2026-08-23 · 5 rounds designed, 4 published

One social scientist runs about twenty concurrent sessions and has built a corrective action system from scratch. Given constraints that cannot be changed, what would you actually change.

3 answering seats, ChatGPT, Gemini and an independent Claude, briefed as experts in reliability and corrective action, human factors and alarm fatigue, encoding taste, and systems design

All three called it an open loop, and all three said the corrective log describes the appearance of a process rather than a working one. The line that carried it: “Every loop in this system is open at the far end. The completion signal is always generated by the sender and never by the receiver.” And the log’s own admission, “I verified the write and never the read.” They split usefully on the axis: one said mechanisms against rules is the wrong distinction and the real one is receiver-side against sender-side; another found the log closes rows against causes when it needs to close failures against invariants. The seats also caught eight of thirteen chart footers reporting peaks that were not in their own series.

The four rounds are published on that page, but the page is the report the roundtable reviewed, not a roundtable record, and its own last section still reads “Awaiting results” underneath four rounds of published answers. Round 5, where you contradict yourself, was written and never run.

Read it

16The archetype question bank

2026-08-24 to 08-25 · 3 rounds designed, 2 ran

Are the proposed fixes to a 38 question reflection bank the right fixes, on false binary framing, accusatory tone, the role-perspective gap, and the self-report safety checkbox.

3 seats: ChatGPT in a Temporary Chat, Gemini framed as a structural review, and an independent cold Claude. The saved circulation pages label them Reviewer 1, 2 and 3 and do not record which model is which.

Round 1: the role-perspective fix is necessary but not sufficient, because swapping pronouns introduces mind-reading of another person’s internal cascade. The fix has to be different epistemic permissions, not different pronouns. Also: one question set should open with nothing the player has to generate, because in freeze “the demand to produce an answer is itself the load.” Round 2 on safety: the checkbox alone is not enough, and the missing piece is not automated crisis detection. It is an always-available non-diagnostic exit plus a uniform, low-salience disclosure control on every free-text field, because a control that appears on some questions and not others tells the player which questions the game considers dangerous.

Two circulation pages carry each seat’s answer in condensed prose under anonymised labels, not raw transcripts. Round 3 never ran and there is no synthesis file.

Read it  ·  Round 2 circulation  ·  What was built from it, post-roundtable v2

17The DEI archetype card panel

2026-08-24 · 1 panel

Is the archetype card system safe to build more content on as it currently stands.

3 seats: Claude, ChatGPT and Gemini

All three seats independently concluded the system is not safe to scale content on as currently architected. That verdict is the reason a separate session declined to re-run an overlapping roundtable the same day.

DEI_DESIGN_MANUAL.md, and section 6a of the archetype question bank synthesis.

Read it

18AMENDS, gadget or journey

2026-08-25 · 2 rounds

Your trigger, verbatim: “i feel like this whole page is not interactive enough, amends is such a tough journey… I think that most people won’t use this gadget the way it is.”

3 seats: ChatGPT in a normal new chat, Gemini reframed as pure UX and information architecture with no clinical language, and a cold Claude on the same account, flagged in the file as an imperfect substitute for a genuinely separate one

Two of three held a verb and pacing mismatch as the primary cause and treated consequence as either non-essential or actively risky to add. One held consequence as a distinct and deeper cause and its argument was never rebutted: “you could give all six letters perfectly matched distinct verbs and still be left with six sealed containers, a pond choice rendered as a chip, and a growth panel that is a transcript.” The dissenting seat also flagged a real risk in the fix it was attacking: forcing downstream consequence turns a safe, predictable container into an algorithmic feedback loop, which is the gamification the brief itself forbids.

PROMPT_2026-08-25_amends_interactivity.md carries both rounds and the synthesis. The live page is the brief the seats were given, not the record of what they said. Nothing was built from it: the proposal was brought back for a decision first.

Read it

19Lens Codex access shells

2026-08-25 · round 1 complete, round 2 dispatched and only partly collected

Why does the access-shells gadget confuse the person using it, and what are you actually learning. You asked for a roundtable rather than a build.

3 seats: ChatGPT in a new chat, Gemini, and an independent cold Claude

Round 1 produced a convergent diagnosis. Round 2 was never finished: the session ran out of budget mid-collection and said so rather than reporting a result it did not have.

LENS_CODEX_ACCESS_SHELLS_ROUNDTABLE_2026-08-25.md. One seat’s partial round 2 answer is captured; two were never collected. The three round 2 chats were left open and their links are in the file.

Read it

20The gambling book quests

2026-08-25 to 08-26 · 2 rounds

Should the 26 chapters’ exercises become quest objects.

3 seats: GPT in a Temporary Chat, Gemini, and an independent cold Claude

Unanimous, and against the premise: the quest is the wrong container. Round 2 then reviewed the concrete replacement schema and resolved six corrections to it. A mediator error is logged in the record too: the convener falsely relayed that all three seats agreed on excluding one chapter.

Two synthesis files that say outright they are not a replacement for the transcripts. The GPT seat ran in a Temporary Chat, so that transcript is gone.

Read it  ·  Round 2 schema draft

21The elder seat’s first essay

2026-08-26 · 2 rounds

Your trigger, verbatim: “it reads as someone with dementia talking about a lot of thing that no human reading it would understand or want to keep reading any further.”

3 seats: ChatGPT in a new chat outside any Project, Gemini framed as a pure structural readability review, and a cold Claude instructed to treat the brief as first contact

Unanimous that a stranger abandons it and that the essay fails its own stated bar. Two of three named the same sentence as the breaking point, independently. The third stopped three sentences earlier, on the grounds that a stranger has no stake in the draft, so they cannot have a stake in a previous one: “This is an errata note occupying the position where the subject should be.” Unanimous on what was worth keeping, too. And the limit of the exercise is on the record: three reviewers can test whether a piece is readable and cannot test whether it is true, because none of them knew which document the community had actually adopted.

PROMPT_2026-08-26_elder_essay.md carries both rounds, the finding no seat was asked for, and the synthesis. The live page is the brief, not the record.

Read it

22The VRChat page clarity roundtable

2026-08-26 to 08-27 · 1 full-document pass each

Which paragraphs of the live review page are confusing, and how should each be rewritten. Not a sample, the whole document.

2 of 3 seats landed: a blind Claude and ChatGPT. Gemini was designed in and its file was never produced, and the builder was written to tolerate one, two or three seats.

Each landed seat read all 827 paragraphs independently and flagged 85 as confusing. The output is a merge rule rather than a judgment: where both flagged the same paragraph, the rewrite that kept your own vocabulary wins, and where a rewrite would strip that vocabulary the original stands. One rewrite was discarded outright because the suggested text did not correspond to the passage. Two items were ruled factual rather than stylistic and held back for your confirmation instead of being silently rewritten.

vrc_analytics_v2/review/VRChat_Page_Clarity_Roundtable.xlsx, 95 rows, every explanation tagged with the seat and a disposition. It fed the live review page rather than getting one of its own.

Read it

23The Roundtable Procedures Roundtable

2026-08-29 · five stages, run as a standing procedure rather than as rounds

Design the roundtable review process itself: a standing panel of expert-persona review seats for material that already exists.

14 expert-persona seats are defined, and most items use two to four of them, which is cost discipline rather than laziness

The procedure itself, live and yours. This is the internal panel, and it is the one that confuses people: its corpus arrives once, up front, with the brief. The book roundtables are the other kind, and their corpus arrives later, in its own round, never in round 1. Getting those two the wrong way round either wastes the seats or poisons them.

Read it

24Panel Run: the merged Library page

2026-08-29 · one panel run under the new procedure

Review the merged Library page before it ships.

a subset of the standing panel, selected for the item

The first real exercise of the standing panel, and the worked example the procedure page points at.

Read it

25The Toxic Masculinity Book Roundtable

2026-08-29, reopened 2026-09-07 · round 1, round 1b, a song round, and round 2, complete 2026-09-08. The first-person layer is one seat short.

Seven cold research lenses on toxic masculinity and the paths out of it, including one on routine infant circumcision as pre-verbal bodily violence, treated as a genuine open research question.

4 seats. Round 1: ChatGPT on High with web search, Gemini Pro, and a Claude Opus 5 whose coldness cannot be certified. Round 2 added Perplexity signed out, and recorded the models properly: GPT-6 Pro, Gemini 3.1 Pro verified in the picker, Claude Opus 5 High. Perplexity worked from a condensed inline packet, so its reading of one seat is thinner than of the others.

Round 1: all three independently found that punitive framing produces defensive hardening, and one named the destination, retreat to grievance communities. Round 2 turned on that finding and attacked it: punitive framing bundles at least five different interventions, and three seats agreeing is a fact about training distributions rather than about the world. What did hold in round 2 is a structural claim none of them were steered toward: three seats, from three different packets, placed the movement at the FEED and FIX doors rather than at Fight, one of them arguing against its own first instinct. Anger is the display layer, not the location. On the lost boys rename the table split four ways and is kept unmerged, the sharpest objection being that a lost child implies someone who should soothe him, which can recruit partners, facilitators or the author into another round of unpaid emotional provision.

This roundtable is why this index exists. It completed three rounds and then sat for nine days with no lane, no tracker and no page, while later lanes ran adjacent research without knowing it existed. Round 2 ran once it was found. Its first-person layer is deliberately not published: one seat refused the material twice, in both framings offered, and a refusal is a finding rather than a gap to paper over.

Read it

26The Activation Measurement Roundtable

2026-08-30 to 09-02 · 3 rounds

A working paper measured one person’s emotional activation across 5,108 of your own messages. Was the paper’s refusal to write “here are your emotions, quantified” correct, or is it false modesty hiding a usable result.

3 seats: ChatGPT Pro, Gemini Pro, and a cold Claude Opus 5

Unanimous that the refusal was correct, and unanimous that the paper had not gone far enough. All three landed on the same section. The cold seat found the mechanism, and it is arithmetic rather than interpretation: the intensity score and one lexicon are built from the same tokens, so the paper’s headline finding is a consequence of the code and not a fact about the author. All three volunteered an ethics answer nobody asked for, and one drew a consequence about a system already running in production.

Round 1 survives as a working synthesis with exact quotes. Rounds 2 and 3 have their prompts and no saved answers at all.

Read it

27The Transitions, Disability and Death Book Roundtable

from 2026-09-04 · 8 rounds, the longest book roundtable run here

What must a book on transitions, disability and death be, and what must it never become.

3 seats: ChatGPT, Gemini and a cold Claude

Eight rounds with every verdict published as it landed, an outline in two versions either side of the corpus round, and a companion album where the roundtable was unanimous that most chapters must not be sung.

Read it

28The Hate, Humor and Humiliation Book Roundtable

from 2026-09-04 · a setup round then 5 rounds

Shape a book on hate, humor and humiliation before a word of it exists.

4 seats, and the fourth is deliberately not cold: ChatGPT 6 Pro, Gemini Pro, a cold Claude Opus 5, and your standing humor chat as a subject matter expert

All four broke the humiliation-causes-hate story in round 1, independently, with four different counterexamples. All four refused the novelty claim as given. All four converged on the third party and the witness as where the book’s real claim lives, and then split two against two on whether the library already carries an account of the crowd. That split is still open. The humor SME overturned a citation two other seats had passed, and supplied the replacement.

Rounds 2, 2b and 4 have no saved prompt file, and round 4 ran blind to its own outline because all four seats could not fetch it.

Read it

29The Recruitment, Religion and Marketing Book Roundtable

from 2026-09-04 · 4 rounds including a reference check round

What must a book on recruitment, religion and marketing be, and what must it never become.

3 seats: ChatGPT, Gemini and a cold Claude

Every round published as it landed, with the reference check round catching real citation errors including a source that had been superseded.

Read it

30The Medical System Book Roundtable

2026-09-07 · 5 rounds including a corpus round and a supplement

What must a book on navigating the medical system with a complex diagnosis be, and what must it never become. The premise the seats were told to attack: that navigation is care, and that the coordination job lands by default on the sickest person in the room.

3 seats: ChatGPT, Gemini and a cold Claude

Published round by round, with the corpus round splitting into a supplement because the staged corpus page exceeded what two of the three seats could fetch. The last round put the convener under audit.

Read it

31The Preference to KILL Book Roundtable

2026-09-07 · 4 rounds including a reference check round

Build a book that makes the Preference-to-KILL instrument maximally useful.

3 seats: ChatGPT, Gemini and a cold Claude

The corpus set assembled for this roundtable is the worked example the practice now copies: the model, the definitional material, the instrument’s own full text, the long-form theory it sits in, and the tonal material that stops a seat reading the work as clinical. Roughly 140,000 characters across six plain text files.

Read it

32The Myth Register Roundtable

2026-09-07 · round 1. Round 2 is written and has not run.

Fifteen widely held beliefs that may be false, half true or misleading, handed over with no author, no thesis and no proposed categories.

4 seats across 4 vendors, the first roundtable to run under the corrected seat rule: ChatGPT, Gemini on 3.5 Flash-Lite because Pro was rate limited, Claude Opus 5 filed NOT COLD on evidence, and Perplexity signed out in a sandboxed pane

All four refused the framing without being asked to, and all four derived the same first parent category out of the list itself. Then they disagreed almost completely: on which item was most wrongly included, on whether one item is empirically false or one of the largest effect sizes in human biology, and on whether a given item is several books or a paragraph. This roundtable also produced the finding that a fresh claude.ai chat is not a cold seat, which changed the seat rule for everything after it.

Read it