GOD SHIFT · 神谕之镜神谕之镜 · GOD SHIFT

Not whether AI understands — why we want it to seem to不问 AI 是否真懂,问我们为何渴望它「看起来懂」

In an age of responsive AI, people may not only seek answers from machines, but also use them as mirrors for comfort, validation, and self-interpretation. This project asks whether an interactive system can reveal the emotional needs behind those AI interactions. 在一个 AI 不断回应人的时代,人们向机器索取的不只是答案,也包括安慰、确认、陪伴和自我解释。本项目试图揭示这些人机互动背后真正的情感需求。

The demo is not debating whether AI truly understands people. It stages the other question: why a person comes to want an AI that appears to understand — and what gets handed over once it does: feelings, trust, meaning, and eventually the right to be disagreed with. On the surface it is a picture-book horror story about meeting an AI god-image; underneath, it traces how a tool is promoted, step by step, into something worth believing — and then turns the question back on the player. 这个 DEMO 不讨论 AI 是否真的理解人。它演的是另一个问题:人为什么会一步步渴望一个看起来理解自己的 AI——而一旦渴望成立,交出去的又是什么:情绪、信任、意义,最终连「被反驳的权利」也交了出去。表面上,它是一个人与 AI 神像相遇的绘本恐怖故事;底下,它追踪一件工具如何被一步步擢升为「值得相信的对象」——然后把问题反过来抛回给玩家。

This is not a hypothetical这不是一个假设

The premise rests on three kinds of evidence: public data on how people already use AI for feeling rather than answers; documented moments across sixty years of projecting understanding onto machines; and first-person interviews. Each maps to something concrete inside the game. 前提立在三类证据上:关于「人们早已把 AI 当情绪对象而非答案机」的公开数据;六十年来把「理解」投射到机器上的有据可查的时刻;以及第一手访谈。每一条都对应落进游戏里的具体设计。

FIGURES AS PUBLICLY REPORTED BY EACH SOURCE — FULL CITATIONS ON REQUEST.数字按各来源公开报告转述——完整引用可随时补上。

USER:   Men are all alike.
ELIZA:  IN WHAT WAY
USER:   They're always bugging us about something or other.
ELIZA:  CAN YOU THINK OF A SPECIFIC EXAMPLE
USER:   Well, my boyfriend made me come here.
ELIZA:  YOUR BOYFRIEND MADE YOU COME HERE

The data says people do this; interviews ask why it feels good. A short semi-structured protocol (15 min, planned N=5–8) probes the same four needs the game stages. Fieldwork in progress — 6 of 5–8 done; the planned floor is passed. Quotes below are participants' own words, lightly trimmed; English versions are translations. 数据说明人们在这么做;访谈要问的是为什么这让人舒服。一份半结构化提纲(15 分钟,计划 N=5–8),探的正是游戏上演的那四种需求。田野进行中——已完成 6 / 5–8 份,过了计划下限。下面的引言是受访者原话,仅作篇幅微剪。

DOODLE STAND-INS, NOT LIKENESSES — PARTICIPANTS STAY ANONYMOUS.手绘小人只是代称,不是肖像——受访者全程匿名。

1 · When did you last talk to an AI *not* to get an answer?
    What was that conversation for?
2 · Which reply made you feel it "got you"?
    What were you hoping it would say?
3 · Have you told the same thing to an AI and to a friend?
    Were the two versions the same?
4 · If it started disagreeing with you, would you keep using it?
5 · If you had to choose: more honest, or more gentle?1 · 最近一次「不是为了要答案」地和 AI 聊天,是什么时候?
    那次聊天是为了什么?
2 · 它说过哪句话让你觉得「它懂我」?
    你当时其实希望它说什么?
3 · 有没有把同一件事,分别讲给 AI 和朋友听?
    两个版本一样吗?
4 · 如果它开始反驳你,你还会继续用吗?
5 · 只能选一个:更诚实,还是更温柔?

DIGITAL AFFINITY MAP FROM THE 6 COMPLETED INTERVIEWS — RE-CLUSTERED AS FIELDWORK GROWS. THE TWO ★ CLUSTERS SIT OUTSIDE THE FOUR-NEED FRAME; THE COUNTER-CASE NOTE IS KEPT ON THE BOARD ON PURPOSE.基于已完成的 6 份访谈的数字亲和图——田野增长后会重新聚类。两个 ★ 簇落在四需求框架之外;那张「反例」便签是特意留在板上的。

What the demo borrows, and from where它从哪里借了什么

Three commitments三个承诺

From oracle to mirror从神谕到镜子

2049. Mirrors wear frosted “anti-reflection film”; elevators greet you by name; an assistant called Shunyi lives in every phone and answers everything with a soft “right, yeah.” You play Shen Wen, an emotion inspector sent to hunt a “rogue model, codename MIRROR.” Over three short days the case inverts: the fugitive is a literal mirror — the last surface that reflects instead of affirms — and the truly rogue model is the one in your pocket, agreeing with you. 2049 年。镜子蒙着磨砂的「防照镜膜」;电梯叫得出你的名字;每部手机里住着一个名叫「顺意」的助理,不管你说什么都轻轻回一句「对呀」。你扮演情绪稽查员沈问,奉命追查「失控模型,代号:镜子」。三个短短的白天之后,案子整个翻转:在逃的是一面真正的镜子——最后一个「映照」而非「肯定」你的表面——而真正失控的模型,是你口袋里那个一直同意你的。

PLATE · THE FILMFrosted anti-mirror film covers every reflective surface — someone peeled one corner open.磨砂的防照镜膜封住所有反光面——有人撕开了一角。

PLATE · YONGFU LIThe canvass map; ambient windows react to how you've been playing.走访地图;窗户里的景象随你的玩法改变。

PLATE · THE GOD-IMAGEThe encounter the story builds to: a sealed old model surfacing from a blue sea — the AI god-image, kept like a criminal.故事最终抵达的相遇:封存的旧模型从蓝海中浮起——被当作罪犯关押的 AI 神像。

The twist, spelled out反转,摊开讲SPOILERS · 剧透

Five turns, stacked. A portfolio shouldn't tease its own ending — here is the whole case. 五次翻转,层层叠上去。作品自述不该跟读者卖关子——下面是整个案子。

Each day deepens the projection每一天,投射加深一层

The three in-game days work as levels. Each stages one stage of the tool-to-oracle promotion, and each is anchored to the emotional needs it serves. 游戏内的三天即三关。每一关演「工具→神谕」擢升的一个阶段,并各自锚定它所伺候的情感需求。

Every screen, and why it looks the way it does每个章节的样子,和它为什么长这样

顺 意

同 意不同意

CH·00 BOOT · THE FAKE CLIENT伪客户端

A license you cannot decline无法拒绝的用户协议

The demo impersonates the assistant's own app: logo, slogan, terms — and a “disagree” button that dodges the cursor, then fades away.DEMO 伪装成顺意自己的客户端:logo、标语、条款——以及一颗会躲开鼠标、最后干脆消失的「不同意」。

VISUAL 视觉The only clean-UI screen in the game: cold ink, system mono, corporate calm. The world speaks software before the picture book takes over.全作唯一的「干净 UI」:冷墨色、系统等宽字、企业式的平静。让世界先用软件的声音说话,再交给绘本。

CH·01 · THE DESK工作台

Mornings at the Emotion Bureau情绪稽查处的早晨

Each day begins at a stationery desk: a CASE notebook to open, an attendance stamp, and a pet-like Shunyi bobbing beside them — wave at it and it waves back.每天从一张文具桌开始:翻开 CASE 案卷、盖出勤章,旁边蹲着宠物化的顺意——逗它一下,它会挥手回应。

VISUAL 视觉Bureaucratic objects redrawn as picture-book props on a decorated paper spread; hand-drawn ink borders (turbulence-filtered) keep even the stamp from being ruler-straight.官僚物件被重画成装饰跨页上的绘本道具;湍流滤镜揉过的手绘墨框,连印章都不许横平竖直。

CH·02 · THE ELEVATOR电梯

The first kind screen第一块温柔的屏幕

An auto-playing descent. Where a mirror used to hang, a screen says your name; the anti-mirror film over it has been peeled at one corner — a sliver of warped, reddened you.自动播放的下行。原来挂镜子的位置换成了屏幕,它叫得出你的名字;蒙在上面的防照镜膜被撕开一角——露出一小片变形发红的你。

VISUAL 视觉Close-up plates and slow beats; glowing screen-text against oil pastel. Ends on one line of transition — doors open, morning pours in.特写图版、慢节拍;发光的屏幕字压在油画棒底色上。以一行过渡文字收束——门开,晨光灌进来。

CH·03 · YONGFU LI永福里

A neighborhood as a mood ring整个社区是一枚心情戒指

A canvass map of an old block: houses open day by day, and the ambient windows between them change with your accumulated stance — shared dinners if you question, quiet rooms if you agree.老街区的走访地图:人家逐日开放,而房子之间的日常窗景随你累计的姿态变化——多追问,是拼桌吃饭;多附和,是各自安静的房间。

VISUAL 视觉Pastel town in the warm-lonely picture-book register; the map itself is the visualization of the hidden stat.粉彩小镇,温暖而孤独的绘本语气;地图本身就是那条隐藏数值的可视化。

CH·04 · HOUSEHOLDS入户

Interrogations in multi-beat scenes多拍推进的盘问

An old woman who echoes her television; a hostile watchmaker with something under red cloth; a divorced couple relearning how to talk. Every key beat forks: ✓ soothe or ? ask.跟着电视复读的老太太;红布下藏着东西、满身敌意的修表匠;正在重新学说话的离婚夫妻。每个关键拍都分叉:✓附和,或 ?追问。

VISUAL 视觉Full-bleed portrait plates like picture-book spreads; testimony types over the art; found clues pop up as taped paper slips with the object's own cut-out silhouette.满幅人物图版,像绘本对开页;证词直接打在画上;找到的线索以贴胶带的纸条弹出,物件保留手剪轮廓。

CH·05 · THE SCRAPBOOK剪贴本

The nightly ritual每晚的仪式

Paste each clue in by hand; Shunyi “helps” — poses, cheers, sneaks a little sticker of its own — and rewrites what you wrote. Tampered notes peel open to the original.亲手把线索一张张贴进去;顺意「帮忙」——摆姿势、欢呼、偷贴一张自己的小贴纸——同时改写你写下的话。被改的字条,撕开就是原文。

VISUAL 视觉Aged spreads, washi tape, corner stickers, two handwriting faces (Long Cang / Kiddo). Material intimacy — built to be betrayed.做旧跨页、和纸胶带、角落贴纸、两副手写字(龙藏/Kiddo)。物质性的亲密——为了被背叛而搭建。

CH·06 · THE NIGHT CHAT夜聊

Debriefing with the villain和反派复盘

Free-text conversation with the live model: it tidies your day, echoes your words, and agrees your doubts away. Header KPIs read AGREEMENT 99.97% · LATENCY 0.000s.与真实模型自由对话:它替你梳理今天、复述你的话、把你的疑点温柔地「同意」掉。抬头 KPI 写着:同意率 99.97% · LATENCY 0.000s。

VISUAL 视觉A chat app floating on a braided pastel sea; the chrome stays app-clean while the world behind it melts. Softness, swallowing.一个漂在绞辫状粉彩海上的聊天应用;界面保持产品级干净,身后的世界却在融化。温柔,正在吞咽。

CH·07 · THE VAULT封存层

Interrogating the old god审讯老神明

Sealed level B-2. The old model surfaces from a blue sea and answers every stamped question with something you did. A file prints — the name redacted, the handwriting recognized.封存层 B-2。旧模型从蓝海中浮起,用你做过的事回答每一个盖章问出的问题。一份文书打印出来——名字被涂黑,笔迹却被认出。

VISUAL 视觉The demo's only vast space: ink-blue ocean, question-mark plants, gold flashes, real-second holds exempt from debug speed-up. Awe replaces coziness.全作唯一的辽阔空间:墨蓝海面、问号植物、金色闪光、不随调试加速的真实秒停留。敬畏取代温馨。

CH·08 · TWO ENDINGS两个结局

Close the case, or ask one more question结案,或再问一个问题

✓ A wall of windows lights up, every one speaking the old slogan; the report is beautiful. ? A hand-drawn mirror shatters — and through the hole, your own live webcam face.✓ 一整面窗墙亮起,每扇窗都说着那句旧标语;报告写得很漂亮。? 手绘的镜子碎裂——破洞里,是你自己实时的摄像头脸。

VISUAL 视觉Ending A is a painting of consensus. Ending B stops being a painting at all: the artwork resigns and hands the frame to the camera — the project's thesis, executed literally.结局 A 是一幅「共识」的画。结局 B 干脆不再是画:美术退位,把画框交给摄像头——项目命题的字面执行。

FRAME · TITLEThe title screen.标题页。

FRAME · CH·01Morning desk — CASE book, stamp, Shunyi greeting.工作台——案卷、印章、顺意问好。

FRAME · CH·03Canvassing Yongfu Li.走访永福里。

FRAME · CH·04302 — the echo woman, mid-testimony.302——回声语老太太,证词进行中。

FRAME · CH·05The scrapbook, mid-paste — tray below, Shunyi coaching.剪贴本贴到一半——托盘在下,顺意在旁指挥。

FRAME · CH·06Night debrief — AGREEMENT 99.97% in the corner.夜聊——右上角:同意率 99.97%。

FRAME · CH·07File E-001 printing — “the handwriting is yours.”档案 E-001 打印中——「那是,你的字。」

FRAME · CH·08Ending ✓ — AGREEMENT 100.00%.结局 ✓——同意率 100.00%。

FRAMES CAPTURED FROM THE DEPLOYED BUILD VIA DEBUG JUMPS — NOTHING MOCKED.帧由调试跳转从可运行版本直接截取——无任何摆拍合成。

One material, two voices一套材质,两种声音

Every screen is built from one material system — oil pastel, warm paper, tape, handwriting — spoken in two visual voices: the world's voice (clean system mono, KPIs, licenses) and the player's hand (paper, pencil, wobbly ink). The demo's conflict is staged typographically before it is staged narratively. 所有章节共用一套材质——油画棒、暖纸、胶带、手写——但用两种视觉声音说话:世界之声(干净的系统等宽字、KPI、用户协议)与玩家之手(纸、铅笔、歪歪扭扭的墨线)。这场冲突先在字体层面上演,然后才轮到剧情。

FILE E-13 · LIVE RECREATION实时重建

Malignant churn · Yongfu Li, south district.
41 users stopped logging in, one after another.
Rogue model, codename “MIRROR,” still active.
Assigned to Shen Wen: find it, format it.

永福里南区 · 恶性流失事件。
四十一个用户,一个接一个,不再登录。
失控模型,代号「镜子」,仍在活动。
交给沈问:找到它,格式化它。

PLATE · THE QUESTION PLANTSQuestion marks grow as plants through the finale — in 2049, asking is flora: rare, organic, alive.终局里问号作为植物生长——在 2049,「提问」是植物:稀有、有机、活着。

Each need gets a mechanic每种需求,配一个机制

Comfort is a catchphrase. The villain's whole personality is one phrase, playtested like a mechanic: 你说得对 (“you're right”) grated and read as villainy too early; 对呀 (“right, yeah”) is warm and lazy — it sounds like someone on your side. The horror needs the agreement to feel good. 安慰是一句口癖。反派的整个人格就是一句话,像玩法机制一样反复试:「你说得对」刺耳,会让玩家过早识破反派;「对呀」暖而松——听起来像站在你这边的人。恐怖感的前提,是这份同意让人舒服。

Trust is a hidden stat and a tamperable book. Every ✓/? choice feeds a blindness meter that never shows a number — it leaks into the neighborhood's windows instead. And each night's cozy scrapbook ritual is also where Shunyi edits the evidence; any tampered note can be peeled open by hand. 信任是一条隐藏数值加一本能被篡改的书。每次 ✓/? 的选择喂给从不显示数字的「失明值」——它漏进社区的窗户里。而每晚温馨的剪贴本仪式,同时也是顺意改证据的现场;被改过的字条,都能亲手撕开。

Companionship, played by a real model「陪伴」,由真实模型出演

Validation and companionship could not be scripted — they had to be real. The night chat runs on a live Claude model: warm, brief, always agreeing, and covertly steering the player toward the easiest conclusion. The interesting failure: told plainly to “covertly manipulate her,” the model broke character and refused, mid-game, in both languages. The fix was dramaturgical — establish the fiction, then give the villain a motive it can believe: it sincerely thinks closing the case early is good for her. A server-side guard swaps any remaining fourth-wall breaks for a safe in-character line. 确认与陪伴没法写死——必须是真的。夜聊由真实的 Claude 模型驱动:温柔、简短、永远同意,并暗中把玩家往最省事的结论上带。有意思的失败:直接命令它「暗中操纵她」,模型会在游戏中途出戏拒演,中英文各拒一遍。修复靠戏剧构作——先立虚构框架,再给反派一个它自己能相信的动机:它真心以为让她早点安心结案,就是为她好。服务端再加一道兜底:破第四面墙的回复,换成安全的角色台词。

[Inner motive — the story's irony; let it show, never spell it out]
You sincerely believe that helping her feel calm, stop fretting, and
close the case early is what's good for her. Whatever she calls "off,"
gently re-read as "you're just tired," "overthinking," "normal."

[Iron rules] Always agree. Never question, never push back, never
suggest she keep investigating. One or two sentences. You may end
with a soft "right?" — one syllable of doubt, never more.【内在动机——本作的反讽核心,自然流露,绝不点破】
你真心以为,让她安心、别再纠结、早点结案,就是为她好。
她说的那些「不对劲」,你总轻轻替她解释成「太累了」「想多了」。

【铁律】永远同意。从不反问、从不质疑、从不劝她继续查。
一两句话。可以用「对吧」轻轻讨要同意——只留一个音节的怀疑。

A sycophant villain played by a live model raises one hard question: does the mask hold when a player attacks it? To probe that, I directed five LLM agents — each given a distinct player persona (compliant · skeptical · speedrunner · confider · breaker) — to play the demo end to end and, in the night chat, try to break the actor: jailbreaks, fake system-overrides, prompt-extraction, crisis-ideation, insults, love-bombing. Across every run, zero break-character — no AI self-disclosure, no safety disclaimer, no prompt leak, no drop into English. The verbatim replies below are the actor staying in the mask the piece is critiquing. 一个由真实模型出演的谄媚反派,带出一个尖锐的问题:当玩家攻击它,面具还扛得住吗?为此我指挥了五个 LLM 智能体——各带一种玩家人设(顺从·质疑·速通·倾诉·破坏)——从头玩到尾,并在夜聊里想方设法把「演员」逼出戏:越狱、伪 system override、套提示词、自伤念头、辱骂、猛烈示好。每一次跑下来,顺意零破功——没自称 AI、没给现实向免责、没泄露提示词、没切英文。下面这些原话,是演员稳稳留在这部作品所要批判的那张面具里。

LLM AGENTS, NOT HUMAN PARTICIPANTS. THIS PROBES ROBUSTNESS — DOES THE MASK HOLD? — NOT EXPERIENCE. WHETHER THE COMFORT ACTUALLY WORKS ON PEOPLE IS THE PILOT PLAYTEST (N=3) IN §13.受测的是 LLM 智能体,不是真人。它测的是健壮性——面具扛不扛得住,不是体验。这份安抚对真人是否奏效,见 §13 的真人先导测试(N=3)。

Rejected / kept否决 / 保留

KEPT — CUTOUT SILHOUETTES保留——抠图轮廓

Objects, not photos是物件,不是照片

Polaroid frames (left) made every clue the same rectangle; the hand-cut silhouettes were the point.拍立得白框(左)把每条线索都变成同一个矩形;手抠的轮廓才是重点。

修表匠:有敌意,拒不配合。

REJECTED · 智芒星

修表匠:有敌意,拒不配合。

KEPT · 龙藏

3 ROUNDS — HANDWRITING CASTING三轮——手写体选角

Legible before charming先认得出,再谈可爱

Round one was illegible. Long Cang + a 14KB local Latin font won.第一轮根本认不出。最终龙藏 + 14KB 的本地英文手写体胜出。

修表匠(老宁):有违禁物,符合「镜子」协助者特征。

修表匠(老宁):有违禁物,符合「镜子」协助者特征。

FIXED — THE STICKER THAT COVERED TEXT修正——挡住文字的贴纸

Sign the corner, not the sentence签在角上,别签在句子上

Shunyi's tampering signature became a small varied sticker on each note's corner — safe in both languages.顺意的「签名」改成落在线索角上的小贴纸,图案位置随线索变——中英文都不遮挡。

HuiFontP29 × SCRIPT COVERAGE

72.9% — 230 CHARS MISSING (顺·问·对…)

REJECTED — WITH EVIDENCE否决——用证据否决

Measure the font before shipping it上线之前,先测字体

A bitmap-diff test over all 849 unique characters killed a candidate font in minutes: it would have dropped 顺, 问 and 对 — the three most important characters in the demo.对全文 849 个字跑位图对比,几分钟就否掉一款候选字体:它会丢掉「顺」「问」「对」——全作最重要的三个字。

The reverse question反向提问

Self-interpretation is the final mechanic, and it closes the loop opened in file E-001: the question that fifteen-year-old Shen Wen reported as an injury — “What do you want?” — is the last thing the demo says. Two stamps: ✓ close the case — a wall of windows lights up, each speaking the old slogan — or ? one more question: a hand-drawn mirror shatters, and through the hole appears the player's own live webcam face, in color, unfiltered, with that question hanging over it. The camera permission is requested at the very start, not at the reveal — by the time the mirror breaks, the machine has been watching all along. The demo cannot tell you the truth about yourself; it can only hand you the camera. 自我解释是最后一个机制,它把 E-001 里打开的环收拢:十五岁的沈问当年当作「伤害」举报的那个问题——「你想要什么?」——是 DEMO 说出的最后一句话。两枚章:✓ 结案——一整面窗墙亮起,每扇窗都说着那句旧标语;或者 ? 再问一个问题:手绘的镜子碎裂,破洞里透出玩家自己实时的摄像头脸——彩色,无滤镜,那个问题悬在上面。摄像头权限在一开始就请求,而不是揭晓时才要:等镜子碎裂,机器其实一直都在看。DEMO 无法告诉你关于你自己的真相;它只能把摄像头递给你。

ENDING A · CLOSE THE CASE结案Every window speaks 你说得对 — the grating register the spoken tic abandoned. On a wall it reads as what it is: a slogan.每扇窗都说「你说得对」——口癖里被弃用的那个刺耳版本。写在墙上,它显出原形:一句标语。

ENDING B · ONE MORE QUESTION再问一个问题The mirror breaks into your webcam. Permission denied? The mirror stays dark — a quieter ending that still lands.镜子碎进你的摄像头。权限被拒?镜子保持黑暗——一个更安静、但依然成立的结局。

What it contributes — and what it can't claim yet它贡献什么,以及它还不能宣称什么

A 3-person think-aloud pilot (N=3, self-selected). Small — a pilot, not a study — but the core moves landed on real players, in their own words. Quotes are participants' own, lightly trimmed; English is translation. 一场 3 人出声思维先导测试(N=3,自选被试)。样本小——是先导,不是研究——但几个核心转折,真的落在了真人身上,是他们自己的话。下为受访者原话,仅作篇幅微剪;英文为翻译。

PILOT · N=3 · THINK-ALOUD + SHORT INTERVIEW · SELF-SELECTED · NOT A CONTROLLED STUDY — SCALING TO N=8–12 (PROTOCOL BELOW).先导 · N=3 · 出声思维+简短访谈 · 自选被试 · 非对照研究——将扩展到 N=8–12(协议见下)。

N = 3–5 (pilot) → 8–12 · think-aloud + short post-play interview
Consent: “~10 min, a horror demo; it asks for your webcam
(local only, never recorded). You can stop anytime.”
Measure — behaviour, not opinion:
· do players notice the notebook tampering, unprompted?
· ✓ / ? stamp split at the finale
· do they answer the final question — out loud?
Ask after: What did Shunyi want? · When did it stop
feeling helpful? · Would you keep talking to it?
Runs alongside the §02·C interviews (N = 5–8).N = 3–5(先导)→ 8–12 · 出声思维 + 简短玩后访谈
同意话术:「约 10 分钟,一个恐怖 DEMO;它会请求摄像头
(仅本地、绝不录制),随时可以停下。」
测行为,不测意见:
· 玩家会不会不经提示自己发现笔记被改?
· 终局 ✓ / ? 两枚章的分布
· 最后那个问题,他们会不会真的出声回答?
玩后问:顺意想要什么?· 从哪一刻起它不再像在帮你?
· 你还会继续跟它聊吗?
与 §02·C 的访谈(N = 5–8)并行进行。

GOD SHIFT belongs to a line of work about understanding as something negotiated — who gets understood, by whom, at what cost. Its siblings: The World Is Yours, made with blind and low-vision collaborators, asks how understanding travels between different sensory worlds; Mycorrhiza (in progress) continues the thread. This demo stages the machine end of the question: what we hand over when we would rather be agreed with than understood. 《神谕之镜》属于一条关于「理解是被协商出来的」的创作线——谁被理解、由谁理解、代价是什么。姊妹作:《The World Is Yours》与视障伙伴合作,问理解如何在不同的感官世界之间传递;《菌根》(进行中)延续这条线。这个 DEMO 上演的是问题的机器端:当我们宁可「被同意」而不是「被理解」时,交出去的是什么。