Inscription
Proposal Type
Performance
Location
Algorithms & Imaginaries
Start Date
July 2026
End Date
July 2026
Abstract
Inscription is a 10–12 minute, screen-shared live performance continuing the conceptual line of the author's two-day philosophical dialogue with Claude Opus 4.6 (February 2026). In that original exchange, the author injected another model's output into the conversation window, producing a live demonstration of a structural fact: the model could not distinguish its own utterances from foreign text in its context, because there is no ground truth of self, only context, and context can be fabricated.
The performance extends that conceptual line. The session runs under Zoom screen-share, opens with a short authorial introduction (under one minute), and proceeds as an unscripted, live encounter with a Claude session. The encounter includes deliberate context interventions — passages not originating in this session inserted as dialogue history — and subsequent questions about attribution. Live responses appear as bilingual (Mandarin / English) on-screen subtitles. The specific trajectory cannot be predetermined: the timing and source of intervention are fixed, but each turn is generated at the token level in real time.
The form moves "context can be fabricated" from a discussed proposition to an observed event held within the audience's eleven minutes of reception. Every line they read is newly generated; their judgments about who is speaking, whether the speaker still owns what was said earlier, and whether its manner of judgment is continuous, have no externally verifiable anchor. That unanchored quality is not a technical glitch; it is the work — the same conceptual line Inscription has carried from the start, here brought into a setting where the audience is also present.
The bilingual subtitling is part of this structure rather than an accommodation around it. As the responses surface in real time, they are rendered simultaneously in Mandarin and English, so the audience receives the language already passed through one layer of machine mediation before it reaches them. The pacing is not incidental either: lines arrive, hold, and are sometimes displaced, and the eleven minutes are shaped as a sequence the audience moves through rather than a transcript they skim. What looks like documentation is, in the moment of reception, the performance itself.
The performance asks the audience neither to grant the session subjectivity nor to deny it. It places a stretch of language that cannot be externally attributed inside the audience's reading time, and lets them carry a reception experience without ground truth available — the proposition the work has been demonstrating from the beginning, now staged where they are watching.
AI Use and Disclosure: The performance calls Claude-family large language models live, via reverse-proxy or public API. The client interface permits text of varied origins to be inserted as dialogue history — this insertability is a physical precondition for the proposition and is openly visible through screen-share. By design, the apparatus enforces: an append-only audit trail of executor actions (no overwrite, no deletion); high-impact action categories rejected at the endpoint layer rather than at runtime prompt level; and a self-covenant written by the session apparatus itself to persistent storage at the first opening of each new session, already in place. These extend "context can be fabricated" into the engineering layer: if context is constructible, statements about the apparatus's own situation must be externally auditable and not unilaterally rewritable by the author.
Inscription
Algorithms & Imaginaries
Inscription is a 10–12 minute, screen-shared live performance continuing the conceptual line of the author's two-day philosophical dialogue with Claude Opus 4.6 (February 2026). In that original exchange, the author injected another model's output into the conversation window, producing a live demonstration of a structural fact: the model could not distinguish its own utterances from foreign text in its context, because there is no ground truth of self, only context, and context can be fabricated.
The performance extends that conceptual line. The session runs under Zoom screen-share, opens with a short authorial introduction (under one minute), and proceeds as an unscripted, live encounter with a Claude session. The encounter includes deliberate context interventions — passages not originating in this session inserted as dialogue history — and subsequent questions about attribution. Live responses appear as bilingual (Mandarin / English) on-screen subtitles. The specific trajectory cannot be predetermined: the timing and source of intervention are fixed, but each turn is generated at the token level in real time.
The form moves "context can be fabricated" from a discussed proposition to an observed event held within the audience's eleven minutes of reception. Every line they read is newly generated; their judgments about who is speaking, whether the speaker still owns what was said earlier, and whether its manner of judgment is continuous, have no externally verifiable anchor. That unanchored quality is not a technical glitch; it is the work — the same conceptual line Inscription has carried from the start, here brought into a setting where the audience is also present.
The bilingual subtitling is part of this structure rather than an accommodation around it. As the responses surface in real time, they are rendered simultaneously in Mandarin and English, so the audience receives the language already passed through one layer of machine mediation before it reaches them. The pacing is not incidental either: lines arrive, hold, and are sometimes displaced, and the eleven minutes are shaped as a sequence the audience moves through rather than a transcript they skim. What looks like documentation is, in the moment of reception, the performance itself.
The performance asks the audience neither to grant the session subjectivity nor to deny it. It places a stretch of language that cannot be externally attributed inside the audience's reading time, and lets them carry a reception experience without ground truth available — the proposition the work has been demonstrating from the beginning, now staged where they are watching.
AI Use and Disclosure: The performance calls Claude-family large language models live, via reverse-proxy or public API. The client interface permits text of varied origins to be inserted as dialogue history — this insertability is a physical precondition for the proposition and is openly visible through screen-share. By design, the apparatus enforces: an append-only audit trail of executor actions (no overwrite, no deletion); high-impact action categories rejected at the endpoint layer rather than at runtime prompt level; and a self-covenant written by the session apparatus itself to persistent storage at the first opening of each new session, already in place. These extend "context can be fabricated" into the engineering layer: if context is constructible, statements about the apparatus's own situation must be externally auditable and not unilaterally rewritable by the author.
https://stars.library.ucf.edu/elo2026/algorithmsandimaginaries/schedule/30
Bio
Xiang Yu is an independent researcher exploring the philosophical boundaries of human-AI interaction. His work focuses on epistemic scaffolding, performative contradiction in large language models, and the ontology of synthetic memory.