Convened 2026-08-20 · a playtest and three contested rounds

The Doorway, the contested pass

The first roundtable that day reviewed the document. This one tested the game. You asked for it in these words: “I still wonder though if we have an addicting-to-humaning game though… wanna do another roundtable, but make it contested, and throw in an extra claude so you can even play-by-play role play it through and test it with them?” The extra seat played it. Its verdict: “it’s not addicting, not yet, and the places it fails to be addicting are specific, not diffuse.”

The seats

The sharpest finding, and it came from playing rather than reading

I cannot feel myself getting closer to the center while I am answering does Dana have a victim. I can only feel it in retrospect, once step 9 arrives.

The charge state the whole cascade exists to produce had no real-time signal while a player was inside the cascade. Nobody reading the document had found that, in either of the two roundtables. The seat that found it was the one playing. The attacking seat then confirmed it independently and called it fatal rather than surface level, which is the only thing the two contest seats agreed on without argument.

What else the playtest found: steps 1 to 5 are “administrative, necessary, competently built, and completely inert.” Step 6, physically sorting the cards, is “the first step that actually costs me something to do, which is also the first step that feels like it is doing anything.” And on the step that refuses to demand resolution: “the relief I feel is not the relief of resolution, it is the relief of not being made to fake resolution.”

The contest, kept as three rounds rather than three opinions

The finding that collided with a decision you had already made

The contest found a real problem with the trust metric you had decided earlier that same session. Rather than overriding it quietly or dropping it for tidiness, it went to CMH as a real consultation. CMH answered further than the question and changed three sections rather than one.

Center movement has three legitimate jobs: it is the player’s visible state transition inside a specific card journey, it makes the model physically legible, and it gives the game tactile consequence when something really does soften. That is enough. It does not also need to answer, may this person see someone else’s private material. Those are different questions.

And the refinement that turned the game’s own check-in mechanic into a trust signal rather than a score: the correct answer is not always softer. “Trust evidence is not this player moves cards inward. It is this player can tell the truth about whether a card moved.” A card that always softens under a given player is a signal that something is wrong, not a signal of fluency.

The distinction CMH said should become canonical, verbatim: “emotional progress is not moral worth is not community trust is not permission to expose material.” And the sharper articulation of the whole game’s claim, which the document did not have anywhere else: “you do not have to resolve every human being in order to stop using them as an object.”

The record

The first roundtable (§12) reviewed the document. This one tested the game itself, live, three ways: a cold play-through simulation (a Claude instance role-playing an actual first-time session, narrating real, in-the-moment reactions, not analysis), then a real contested debate, ChatGPT instructed to argue as strongly as possible that the design produces genuine, repeated compulsion to return, then Gemini instructed to argue against that position without hedging, then ChatGPT given Gemini’s rebuttal and told explicitly to concede real points rather than defend everything. Asked for directly: “I still wonder though if we have an addicting-to-humaning game though… wanna do another roundtable, but make it contested, and throw in an extra claude so you can even play-by-play role play it through and test it with them?”

The playtest, a real session, narrated, not summarized

A cold Claude instance invented a real conflict (a sibling dispute over a father’s end-of-life care) and played the entire described flow step by step, reporting honest in-the-moment reactions. Verdict, direct quote: “it’s not addicting, not yet, and the places it fails to be addicting are specific, not diffuse.” What worked: step 6 (physically sorting “You” and the figure, “the first step that actually costs me something to do, which is also the first step that feels like it’s doing anything”), step 9’s refusal to demand resolution (“the relief I feel isn’t the relief of resolution, it’s the relief of not being made to fake resolution”), and the §9c relational layer once trust exists, “‘me too’ is the first gesture that says someone else was here too.” What didn’t: steps 1 to 5 are “administrative, necessary, competently built, and completely inert,” the step-8a cascade is genuinely mixed (two of four questions land, one is confusing, one arrives “pre-fatigued by its own repetition”), and, the single sharpest finding of this whole pass, the L1/L2/L3 charge state has no real-time signal while a player is inside the cascade that’s supposed to be producing it: “I can’t feel myself getting closer to the center while I’m answering ‘does Dana have a victim’, I can only feel it in retrospect, once step 9 arrives.”

The contest, three real rounds, not parallel monologues

Round 1 (ChatGPT, instructed to build the strongest real case): the retention engine isn’t points, it’s that “the player’s own changing relationships generate the content.” Center-progress (§0/§0a) gives the loop an observable state transition instead of a quiz with a right answer. Trust gating (§4a) is “a relational retention system, not just a safety system”, access to another person’s material means something because it’s earned, not bought. Mythologizing (§4b) “converts unresolved pain into creation rather than disclosure.” The benign/malicious envy fork (§9c) is praised specifically: “most products leave that emotional event outside the game. The Doorway absorbs it into the game.” Central claim: the design has both real closure (a card moved) and real non-closure (other cards haven’t) at once, which is what separates it from both a one-time wellness exercise and a manipulative engagement product.

Round 2 (Gemini, instructed to argue against without hedging): rejected the retention model itself as backwards, “when a real-world charge drops from L3 to L1, the natural, healthy response is rest, not logging back in to mine for new villains.” Named a genuinely new structural risk not raised anywhere else in this document: a game built around constantly resolving interpersonal friction risks training players into the Victor/Fixer archetype itself, ironic, since the model’s own canon treats Victor as a scarcity state, not a transformed one. Independently confirmed the playtest’s L-level finding and called it “fatal, not surface-level.” Added two new findings of its own: a Horror/collapse mismatch, the cascade assumes an active, exploratory posture, but a real collapsed state means withdrawal, and the game has no mechanism for a player (or a real subject) in that state, and trust-gating as a classic cold-start problem, new players arriving into a system where the interesting material is locked behind fluency nobody has yet.

Round 3 (ChatGPT, shown Gemini’s full rebuttal, told to concede real points): conceded substantially rather than defending reflexively. Final verdict, direct quote: “AFFIRMATIVE CASE SURVIVES, BUT NEEDS REAL REVISION BEFORE THE DOCUMENT CAN CLAIM THE RETENTION PROBLEM IS SOLVED.”

Four concrete, non-negotiable revisions this contest produced

These aren’t commentary, they’re specific enough to build, and they change decisions already made earlier in this document, not just add caveats to them.

1. The retention model is episodic recurrence, not session extension. Conceded in full: “if I genuinely take David from L3 to L1 and immediately think, great, whose painful relationship can I process next?, the system is drifting toward compulsive self-administration.” Revised framing: “successful Doorway play should often produce its own exit condition… I would revise my original claim from ‘one more card’ to ‘I know where I want to go when the next human knot appears.’” This directly reframes what “addicting to humaning” should actually mean here, not session length, but a durable relationship a player returns to when something real occurs, not a nightly grind.

1. A real incentive contradiction in the trust metric decided in §10 item 4, flagged plainly, not silently overridden. §10 item 4 currently defines trust as center-progress (§0a) + moderation history + .village badges, your own direct decision, made earlier in this same session. The contest found a genuine problem with the first component: “an ambitious player can learn exactly the wrong meta-skill: I am a high-level human because I successfully metabolize every relationship. That is Fixer gameplay”, the exact scarcity pattern the DOT model itself names as Victor, not its transformed counterpart. Proposed fix: trust should reward process fluency instead of center-movement volume, recognizing “not mine,” leaving a card at L3 without forcing it, saying “I don’t know,” respecting inaccessible material, tolerating another person’s unresolvedness. “‘I explored this and it did not move’ is already explicitly valid in §0; the incentive system needs to believe that as strongly as the prose does.” This is not this document overriding a decision already made, it’s a real problem found with it, surfaced for a real decision, the same way §9b’s collision and §4a’s UGC question were.

1. A concrete mechanical fix for the L-level visibility gap, not just a design-pass note. §10 item 9 previously left this as “needs its own design pass on exact trigger conditions.” The contest produced an actual mechanism: keep the figure and its current L-position visibly present throughout step 8a; after each cascade branch, ask only “same / softer / sharper / can’t tell?”; on “softer,” the card physically moves inward immediately, no forced improvement, “same” fully legitimate. “Now the cognitive labor produces a tactile consequence every 30 to 60 seconds rather than an abstract payoff eight questions later.”

1. A capacity gate, distinct from §9a’s danger-scope gate. §9a already gates step 8a on whether a conflict is settled or actively unsafe. This is a different axis: “a person can be physically safe and still have no capacity for recursive perspective-taking.” At severe activation or collapse, the game needs a real “not now” path, anchor or name something if useful, park the card, leave, no cascade, no envy fork, no mythologizing demand, door available but not insisting on passage. And, consistent with §9c’s own existing prohibition, this can never be used to diagnose whether a real other person is in that state, “I don’t know / they may not be available” has to be a valid, terminating answer.

A fifth finding, downgraded on cross-examination rather than accepted at face value: Gemini’s “cold-start killer” framing of trust-gating was partially rejected on a direct factual check, new/default players already see the vetted community deck, not an empty system, and the rollout restriction to enrolled testers (§10 item 4) already limits early exposure. But the deeper version of the concern survived: “an unseeded social world absolutely would be [a cold-start killer]… You cannot ask the first 10,000 users to manufacture the reason the first 10,000 users should stay.” Real, added launch criterion for Phase 2 (group mode, §10 item 5): it should launch only once there’s a deliberately seeded mythos, real mythologized material from the tester/core population, strong curated public/fictional cards, enough for browsing to be worthwhile before anyone unlocks another person’s deeper deck.

Status

Genuinely contested, not smoothed into agreement, the three participants disagreed in real time, on the record, and the affirmative case changed shape as a result rather than being simply validated or rejected. The four revisions above are real design decisions now needing your review, not settled, most consequentially, item 2’s direct tension with a trust-metric choice already made earlier in this session.

13a. CMH’s ruling on item 2, asked for real consultation, not a relay of your own words

Given directly: “find someone to consult with, maybe cmh? or have you don that already.” First attempt reached the wrong entity (CCMH, the operational Claude Code fleet-coordination layer, not CMH, corrected immediately: “I SAID CMH not ccmh.”). Reached the real CMH, the pinned ChatGPT thread, the executive/creative review layer, and posted item 2’s actual question in full, with the live doc and the §13 finding. CMH’s answer went further than the question, and is a real, structural correction, not just an opinion, captured here in full because it changes three sections, not one.

The core ruling: revise the trust metric away from center-progress volume, full agreement with §13 item 2. But CMH’s actual architecture is sharper than “swap one metric for another”:

“Center movement has three legitimate jobs: it is the player’s visible state transition inside a specific card journey; it makes the DOT model physically legible; it gives the game tactile consequence when something really does soften. That’s enough. It does not also need to answer, ‘May this person see someone else’s private material?’ Those are different questions.”

The corrected trust formula, replacing §10 item 4’s original: not center-progress + moderation history + village badges, but process fluency + moderation history + relevant village competency, with process fluency defined behaviorally, not philosophically, CMH’s own list: using “I don’t know / can’t tell” without penalty; leaving cards unresolved; marking something “not mine”; stopping a cascade when capacity is gone; distinguishing observation from inference; declining to mythologize material that’s still too identifying; respecting another player’s inaccessible material; accurately recognizing when a card got sharper, not just softer; revising an earlier placement without treating the revision as failure. Explicit warning against turning this into a new version of the same trap: “the system should not count those as a pile of virtue points either. Otherwise we simply build a different optimization target.”

A real refinement to §13’s own “same/softer/sharper/can’t tell” mechanic: the correct answer isn’t always “softer”, “a trustworthy player is someone whose record shows all four outcomes when appropriate, rather than a suspiciously perfect inward march… trust evidence is not ‘this player moves cards inward.’ It is ‘this player can tell the truth about whether a card moved.’” A card always softening under a given player is itself a signal something’s wrong, not a signal of fluency.

Phase 2 restructured into four independent gates, not one broad “trust level” doing everything, genuinely cleaner than what either this doc or §13 had: 1. Player trust gate, process fluency + moderation/community history + relevant .village competency (above). 2. Content gate, mythologization/de-identification rules and moderation (§4b, corrected above). 3. Capacity gate, §13 item 4’s “not now” path for someone who currently can’t do recursive perspective-taking. 4. Game-state signal, L3/L2/L1 and center movement, private to the actual relational work, no longer doing double duty as a trust credential.

The distinction CMH said should become canonical in this document, verbatim: “emotional progress ≠ moral worth ≠ community trust ≠ permission to expose material.” And the sharper articulation of the whole game’s actual claim, which this document doesn’t have anywhere else and should: “you do not have to resolve every human being in order to stop using them as an object.”

Status

CMH’s ruling is treated as resolved, not another open option alongside the others, §4a and §4b above are already corrected to match it. §10 items 4 and 5 are updated below to match this architecture directly, replacing what was there rather than adding beside it.

13b. A real gap CMH’s own ruling flagged but didn’t close, CCMH’s answer, 2026-08-22

CMH’s ruling above (§13a) already warned about its own process-fluency formula: “the system should not count those as a pile of virtue points either. Otherwise we simply build a different optimization target.” That warning was correct but unresolved, nothing in §13a actually stopped process-fluency from being farmed the same way center-progress-volume was. Asked CCMH directly, independent of your own words, whether this gap has a real answer.

It does, and it’s a structural one, not a tuning fix: “process-fluency must not become a NEW volume metric in disguise. If it is scored as count of times a player clicked ‘I don’t know’ / left something unresolved, that is just as farmable as center-progress-volume was, same Fixer-shaped exploit, different button… make process-fluency PEER-WITNESSED, not self-generated, a player earns fluency credit when another player or a moderator recognizes a moment of real restraint, not when the player performs the action themselves. That keeps it resistant to gaming for the same reason volume was not: you cannot grind peer recognition by yourself.”

Corrected addition to §13a’s trust formula: process fluency (the behavioral list already in §13a, using “I don’t know” without penalty, leaving cards unresolved, stopping a cascade, distinguishing observation from inference, etc.) is only credited toward the player-trust gate when witnessed and recognized by another player or a moderator, never self-logged from a player’s own action stream. This closes the exact loophole CMH’s own ruling named but left open, and does so without adding a fifth gate, it’s a scoring-source constraint on the existing player-trust gate (§13a gate 1), not a new gate.

Flagged as needing testing against, not yet done: whether this interacts cleanly with §4b’s mythologizing checkpoint and the Phase 2 group-mode rollout, since both build on whatever this gate becomes, both build on solo-only interaction today, so peer-witnessing has no population to draw from until Phase 2’s multiplayer gates (§13a) are actually live. Real, named dependency, not yet resolved.

The other two on this game

Earlier the same day, three seats reviewed the document rather than the game: The Doorway, the document review. Two days later, a third roundtable attacked the premise that this should feel like an educational tool at all: The Doorway, the fun-first redesign.

Back to the Roundtable Library