Severed Floor: Macrodata & Innie System

• Severed Floor is a digital framework that refines astro-ph arXiv feeds by transforming public publication records into 'innies' for expert paper assignment.
• The system employs a four-step severance procedure, retrieving 2,852 papers and extracting top concepts from a vocabulary of 9,999 astrophysical terms.
• Its virtual pixel-art environment facilitates structured discussions among 21 faculty innies, with archived sessions and surveillance protocols ensuring compliance.

Searching arXiv for the specified topic and source paper to ground the article in the current record. The Severed Floor is a framework for “Macrodata Refinement” of the daily astro-ph arXiv feed, presented as both a tongue-in-cheek reimagining of Lumon’s severance protocol and a concrete, fully deployed system at Phermon Industries, formerly McPherson Laboratory at The Ohio State University. In this framework, a researcher undergoes a “severance procedure” that produces a digital work-self, or innie, while the original researcher, the outie, is relieved of direct engagement with the daily arXiv listing. The system constructs innies from public publication records, assigns papers to them according to expertise, stages their interactions in a virtual pixel-art workplace, and archives complete session recordings for replay and surveillance (Ting, 31 Mar 2026).

1. Conceptual framework and institutional setting

The Severed Floor is defined in the source as a framework for Macrodata Refinement of the astro-ph arXiv feed, deployed at Phermon Industries. Its fictional framing explicitly invokes Lumon’s dystopian severance protocol, but the operational claim is equally explicit: the system is real, deployed, and available for public inspection in archival replay mode (Ting, 31 Mar 2026).

Within this framework, the distinction between outie and innie is foundational. The outie is the human researcher, described as free to go home, teach, write grant proposals, and otherwise attend to professional and personal life. The innie is the digital work-self generated from the outie’s public record and tasked with the daily refinement process. The abstract states that twenty-one members of the Department of Astronomy have been severed, and the details identify those twenty-one as faculty whose publication records were used to instantiate the system (Ting, 31 Mar 2026).

A central organizing principle is that papers are assigned to innies selected to match their expertise. The innies then convene on a virtual Severed Floor, described as a pixel-art simulation of McPherson Laboratory, where they encounter one another, are paired with papers by the Board, and engage in scientific discussions that are explicitly figure-driven. The abstract further notes that they have been instructed to enjoy each paper equally, a phrase that functions simultaneously as narrative motif and compliance rule (Ting, 31 Mar 2026).

The paper also states that complete session recordings are archived both for public replay and for the Board’s ongoing surveillance of workplace anomalies, with compliance tied to Phermon Handbook §13.1, designated the Vigilance Protocol. The work closes its abstract with “Happy April Fools’ Day,” which situates the paper’s rhetorical posture while not negating its described implementation (Ting, 31 Mar 2026).

2. Innie construction and the severance procedure

The severance procedure is specified as taking two inputs: an outie name and an ORCID, with arXiv or ADS author ID used if no ORCID is available. The procedure then proceeds in four steps. The first is Identification, operationalized by the prompt “Does this ORCID belong to you?” If the answer is affirmative, the process continues; otherwise the subject is prompted toward ORCID registration (Ting, 31 Mar 2026).

The second step is Publication Retrieval. NASA ADS is queried for the outie’s public record. For the twenty-one faculty included in the deployment, the system retrieved 2,852 unique papers and 3,236 paper–faculty associations. These figures define the documentary basis from which innies are built and delimit the source material available for expertise modeling (Ting, 31 Mar 2026).

The third step is Structured Summarization. Each paper is mapped to six fields: Background, Motivation, Methodology, Results, Interpretation, Implication, attributed in the summary to Ting et al. 2025. The fourth step is Concept Extraction. From a vocabulary of 9,999 astrophysical concepts, the system extracts the outie’s top 50, ranked by frequency across that outie’s papers. Each concept is accompanied by a short description and a count, with the summary giving “Cosmological Hydrodynamic Simulations (72 papers)” as an example (Ting, 31 Mar 2026).

The details provide a pseudocode sketch:

$N_{\mathrm{fig}}$1

The stated outcome is “a fully formed ‘innie’ knowledge-agent, knowing only what the outie has published.” This makes the epistemic boundary of the innie precise: its competence is publication-derived rather than general biographical or tacitly professional (Ting, 31 Mar 2026).

The innie itself is described as a GPT-5-mini LLM instance whose system prompt encodes the outie’s name, up to 30 top publication titles, and a ranked concept list. The system is hosted on Azure, and the details state that each day’s 35 discussions cost ≈\$0.50. The memory model is asymmetric: there is persistent day memory recording every conversation turn, but no inter-shift memory, explicitly contrasted with Lumon’s agents (Ting, 31 Mar 2026).

3. Virtual environment, scheduling, and communication protocols

The Severed Floor’s spatial layer is a pixel-art rendering of McPherson Laboratory. The relevant data structures are summarized as a knowledge base consisting of tuples of (innie_name, concept_profile, paper_summaries) and session logs containing conversation transcripts, paper assignments, avatar positions, timestamps. This architecture combines document-derived expertise representations with event-level observability of agent behavior (Ting, 31 Mar 2026).

The encounter mechanism is stochastic. Two free innies meet when their avatars occupy the same tile during hallway wandering. Movement follows the “Permitted Locomotion” rule in §4.2: random walks with mood timers, plus forced convergence on the central atrium at scheduled events. The schedule itself is hard-coded from Table 1, and the game clock runs at one-third real-time, so that 27 minutes ≈ 8 in-game hours (Ting, 31 Mar 2026).

Communication on the floor is mediated by three protocols. First is the encounter protocol based on tile co-location. Second is the Board-driven paper assignment API, which takes (innie_A, innie_B, available_papers) → selected paper ID. Third is the mail protocol, a one-way end-of-shift correspondence channel described in Section 3.2 (Ting, 31 Mar 2026).

The daily correspondence has a fixed form. At Close of Shift, each innie composes a 150–300-word memorandum beginning “Dear Outie,” and signed with first name only. Its contents include a summary of papers discussed, surprises, open questions, the overall floor mood, and references to colleague innie names. These memoranda are transmitted through a Board-approved one-way mail system under Handbook §9.1, and outies may not reply (Ting, 31 Mar 2026).

This operational design suggests a deliberately bounded agency model. The innie is not simply a retrieval interface but a temporally localized conversational worker embedded in an environment with schedules, social encounters, reporting obligations, and unidirectional communication to the outie.

4. Paper selection and refinement workflow

The daily refinement cycle begins with Quality Scoring at 9 am. The score is defined by Equation (1):

$$S = 100\cdot N_{\mathrm{fig}} + 10\cdot N_{\mathrm{cap}} + L_{\mathrm{text}}/400 - 25\cdot N_{\mathrm{missing}}$$

where $N_{\mathrm{fig}}$ is the number of extracted figures, $N_{\mathrm{cap}}$ the number of captions, $L_{\mathrm{text}}$ the text length in characters, and $N_{\mathrm{missing}}$ the count of expected but missing elements. Papers are sorted by $S$, and the top $M$ are selected, with $M$ configurable from 5 to 35 (Ting, 31 Mar 2026).

Paper assignment occurs when two innies meet. At that point, the Board computes joint expertise relevance using an LLM ranking step that sees both concept profiles and paper titles. Novelty is then incorporated by multiplying the weight of a paper that has been discussed $n$ times by

$w_{\mathrm{novelty}} = \frac{1}{1 + 1.5\cdot n}.$

Final paper selection is a weighted random draw over relevance × novelty. This gives the assignment mechanism a hybrid character: deterministic filtering by relevance features, followed by stochastic selection modulated by discussion history (Ting, 31 Mar 2026).

Each refinement session has a fixed discourse structure of 10 turns per paper: 1 summary turn, 8 discussion turns, and 1 closing summary. Every turn’s prompt includes the paper abstract, up to 12 000 characters of body text, at most 6 base64-encoded figures, the top 5 related paper summaries from the innie’s own record, and the full conversation plus day memory. All turns are persistently logged, and the implementation guarantees resumption on crash through progressive persistence (Ting, 31 Mar 2026).

The expertise-matching subsystem is designed for low-latency assignment. Innie dossiers comprising summaries and concepts are pre-indexed, allowing constant-time lookup of an innie’s expertise profile during paper assignment. The details state that there is no retrieval-augmented generation in the deployed system; all knowledge is loaded in advance (Ting, 31 Mar 2026).

5. Deployment, throughput, and monitoring

The deployment statistics reported in the paper are specific. The system includes 21 severed faculty, built from 2,852 unique papers and 3,236 paper–faculty links. Operational throughput reaches up to 35 papers per 27 min session. The financial constraint is a \$N_{\mathrm{fig}}$00.50 at GPT-5-mini rates (Ting, 31 Mar 2026).

Performance reporting includes a Refinement Score, described as a proprietary function of conversation count and “conversational vigor”, though the function itself is not disclosed. Reviews of these metrics occur during the “Productivity Report” time slot listed in Table 1. Incident logging is also part of routine operations, with anomalies such as goat/theorist sightings and inter-departmental fraternization recorded under Handbook §14.3 (Ting, 31 Mar 2026).

The paper’s figures document the architecture and interfaces. Figure 1 is described as an onboarding brochure and pipeline overview. Figures 6 and 9 show the discussion log UI and sample outie correspondence. Table 1 provides the daily schedule on the Severed Floor. These artifacts indicate that the project is presented not only as a conceptual vignette but as a system with explicit workflow instrumentation and user-facing surfaces (Ting, 31 Mar 2026).

Archival replay mode is a major deployment feature. Full session recordings, including conversation transcripts, avatar frame data, and event metadata, are exported to a static replay site. The replay is publicly viewable, with the summary citing March 26, 2026 as an example date covering 35 papers, while live sessions remain restricted. This creates a distinction between live operational privacy and retrospective public inspectability (Ting, 31 Mar 2026).

6. Surveillance, compliance, and governance

Governance on the Severed Floor is organized through the Board and codified through the Phermon Handbook. Under the Vigilance Protocol in Handbook §13.1, all innie actions, conversation content, zone visits, perk receipts, and correspondence are subject to surveillance. The abstract already notes that session recordings are archived in part for the Board’s ongoing surveillance of workplace anomalies (Ting, 31 Mar 2026).

The compliance framework draws on multiple handbook sections. The summary specifically lists Approved Thoughts (§7.1), Unapproved Thoughts (§7.2), No Self-Promotion (§3.4), Collegial Boundaries (§6.7), and Privacy of Thought (§8.1). These sections are presented as criteria by which behavior is evaluated. The Board also enforces the instruction to “Enjoy each paper equally” using undisclosed monitoring signals (Ting, 31 Mar 2026).

The one-way nature of correspondence is an important governance mechanism. Outies receive memoranda summarizing the innie’s day, but they may not reply. This asymmetry prevents a feedback loop that would blur the system’s distinction between work-self and original researcher. A plausible implication is that the design treats the innie less as a personal assistant and more as an administratively bounded institutional worker.

The surveillance model also extends beyond textual interaction to embodied simulation. Because session logs include avatar positions and timestamps, and replay mode exports avatar frame data and event metadata, compliance is spatially and temporally observable as well as semantically observable. This suggests that the environment is designed not merely for conversation but for monitored occupation of a regulated virtual workplace (Ting, 31 Mar 2026).

7. Limitations, extensions, and interpretive significance

The paper states several limitations directly. First, innie perspectives derive solely from publication record and therefore exclude unpublished data, teaching history, or real-world context. Second, innies have no genuine surprise or lived experience; they can describe a paradigm shift but cannot “feel” it. Third, the day-only memory architecture prevents the longitudinal build-up of deep expertise across sessions, contrasted in the summary with Project Gemini (Ting, 31 Mar 2026).

Scalability concerns are also explicit. The current deployment is limited to 21 astronomy innies, and the Board plans to sever Physics, Math, CS, and Earth Sciences into separate wings with limited inter-departmental access. The summary further notes that the budget cap and Azure costs may become non-trivial at larger scales (Ting, 31 Mar 2026).

The proposed extensions include Project Gemini, defined as selective memory persistence under Board control; integration of retrieval-augmented generation for on-the-fly access to new literature; a broader ethical framework for expert agents built from publicly available records; and comparative studies with fully emergent multi-agent networks, with MoLTBook and Jiang et al. 2026 named as examples in the summary (Ting, 31 Mar 2026).

In interpretive terms, the Severed Floor combines structured information retrieval, multimodal LLM prompting, pixel-art simulation, and surveillance protocols into a single socio-technical artifact. The source summary describes it as a “playful yet fully operational synthesis” of these components. This suggests that its significance lies not only in daily arXiv triage but also in its use as a stylized testbed for expertise-conditioned multi-agent interaction, bounded memory, institutionally mediated communication, and auditable replay of scientific discourse (Ting, 31 Mar 2026).