Project II

Species sharing a single space differ systematically in movement dimensionality, body scale, and distance thresholds. Prevailing design practice resolves these differences through functional zoning, keeping the activity ranges of each species separate. This strategy avoids inter-species encounter and decouples spatial form from behavioral difference. This project proposes and tests the opposite hypothesis: if spatial components can read the behavioral divergence among co-present parties and reconfigure their own form accordingly, that divergence can be converted from a constraint to be resolved into a generative condition for form. The method is multi-agent simulation. Candidate scenarios are classified into three types of human–animal relationship, and the first to be built is resident cats and unfamiliar visitors in a cat café. Behavioral rules come in two grades: thresholds with an empirical source are used as published, and thresholds without one are marked as demonstration values. Spatial response is carried by an array of contracting, tensioned units hung from the ceiling.

Background and research question

Cross-species interior design has a default solution, vertical stratification. Human furniture occupies the lower volume while cat-climbing elements ascend the walls, each species holding a non-overlapping band of height. Cat cafés do the same. When visitor numbers rise, the cats move up to the shelving (Hirsch et al., 2025). The strategy is sound as engineering, but its cost is the elimination of bodily encounter between species, and stratification reduces cohabitation to juxtaposition.

The research question is whether space can participate in cohabitation as a responding party and not only as a container. More specifically, if the units composing a space can sense the behavioral information left by co-present parties and change form in response, the question is whether inter-species behavioral divergence can be continuously translated into change of spatial form.

Design position

This project frames cross-species space as a problem of volume redistribution. Spatial volume is reallocated according to the free behavior of all parties, and the design objective is the transformation of spatial form. Scale competition, which body takes priority in occupying volume and how dominance between species is adjudicated, lies outside the project.

The project started from a space with no arbitration function. Units would not judge the proper allocation of volume. They would read the behavioral traces left by bodies, including dwell hotspots, high-frequency paths and avoided regions, and shift volume toward the forms those traces indicate. The computational correlate is stigmergy, in which individuals coordinate indirectly through information deposited in the environment rather than through direct negotiation. The established correlate in the built environment is the desire path, where paving ultimately ratifies the shortcut worn in by users.

The first plan-view simulation was built on that premise. Units followed traces, and wherever a person went, something moved. Compared against fixed shelving and an empty room over thirty-two seeds, it produced the fewest stops beside a cat of the three, 9.6 against 12.8 for the fixed shelves and 12.6 for the empty room. The cause is ratification itself. A space that only ratifies traces thickens where bodies already are, so it can ratify use but cannot bring about an encounter. → Log 2026-10-07

The revised version gives the space a direction. In the human–cat case the purpose is to bring people and cats together, in the human–dog case to set apart an area for people and an area for dogs, and with wild animals to divide dynamically. The direction is set by the designer, and direction is all the space receives. It commands no one. It changes only the two costs of each position, where to go remains the body's own decision, and no individual is ever trapped by it. These conditions are written as the five rules in the Method section. Reading traces is unchanged, and what changed is the action taken after the reading. Cybernetic architecture had active spaces long before this. Price and Frazer's Generator would propose rearrangements on its own when nobody touched it. This project set out as a more conservative translator and was pushed back onto that lineage by the comparison. The old rule stays on the Lab 2-14 bench as the control, under "Trace (earlier)" in the rule row, where it can be run.

Scenario classification and selection

Candidate scenarios were classified into three types, each corresponding to one human–animal relationship, and together they form a gradient of co-presence intensity.

Shared interior (human–cat): cohabitation. The cats are residents and the people are visitors. The public version of this type is the cat café, and that is the one the simulation takes: cats and visitors are strangers to each other, co-presence is short and repeated, body scales differ markedly, yet the two share the same furniture and the same ceiling.

Pet square (human–dog): mediation. The dog functions in this scenario as a trigger for human-to-human contact. Civil inattention, Goffman's term for the default norm under which strangers withhold engagement, is legitimately suspended in a dog's presence. Whyte termed the linking of two strangers through an external object triangulation. What the space chiefly handles in this scenario is the human relationship the animal switches on.

Urban residual space (human–wild animal): avoidance. Both parties actively avoid co-presence, and interaction occurs only at the level of traces (feeding points, corridors, rooftops, ruins). The animal's range expands as human space contracts, and conversely.

The simulation takes the first type first. The site is the cat café, which meets three conditions at once. Strangers and resident animals share the room repeatedly, the operator already wants people and cats to interact, and the cats inside have been observed and recorded. All three types run on the same loop (see Method). The human–cat case comes first because its data are the most complete. The distances cats keep from strangers have been measured (Mertens & Turner, 1988), the positions cats choose in a café have been observed (Hirsch et al., 2025), and the distances people keep from each other have been surveyed across countries. The third type follows as a plan-view comparison, using the flight initiation distance of urban birds as the animal's cost of staying.

Within the first type the cat was chosen over other companion animals on two grounds. The human–cat size differential gives a legible contrast of scale, and the relationship contains no third-party mediation, so the objects of interaction are the two species themselves. The cat's three-dimensional mobility is a third reason, but the current plan-view simulation does not model height, so that reason holds only once the solid units are driven by the behavioral layer.

All three types appear on the project boards. The simulation covers the first type in plan and in three dimensions, the third as a plan-view comparison, and the second on the boards only.

Method

The simulation adopts a dual-layer multi-agent architecture. The behavioral layer consists of two classes of body agents, human and cat, whose rules are drawn from empirical research: interpersonal distance rules from Hall's proxemic measurements and a cross-national survey, stranger-interaction norms from Goffman's observational work, quantified dwelling behavior from the public-space studies of Whyte and Gehl, and feline space use and human–cat interaction rules from the animal-behavior literature (see Behavioral clauses and the interaction matrix). The rules come in two grades. A threshold with an empirical source is used as published, and one without a source is marked as a demonstration value. The spatial layer consists of contracting units hung from the ceiling, and its rules are authored within this project. At this stage the spatial layer has one action, a unit descends and forms or it withdraws. Merging, migrating, luring, splitting and dissolving were earlier ideas, not yet implemented, and are listed under Open questions. The two layers are held to deliberately asymmetric standards, the behavioral layer empirically constrained while the spatial layer admits speculation, and this asymmetry constitutes the project's methodological boundary. Speculation is confined to a single layer and anchored by the empirical one.

The simulation operates at the scale of minutes, and its modeled variables are restricted to bodily dynamics: position, orientation, dwell duration, gaze direction, and movement path. Attendance decisions on the scale of days or weeks (whether to be present, long-term changes in willingness to stay) are excluded from the model.

The spatial layer lowers units for a purpose. In the human–cat case the purpose is to bring people and cats together, and which units descend follows the person's state. Walking triggers nothing and leaves no trace. Standing still for three seconds while looking at a cat lays one step toward that cat. Sitting down leads the nearest cat along the units present, one step at a time, to the unit 1.0–1.5 m in front of the seat. A cat reads nothing but unit states. After three seconds on a unit with another offered beside it, it moves or stays by its own odds. When a person approaches a unit that has descended, only the strips facing them retract, like a door opening, and the rest stay down. The earlier rule, a unit reading the traces beneath it and descending once the reading passes a threshold, stays on the bench as the control. The space changes only the cost of each position and gives no orders to anyone, and this position is written as five rules. The spatial layer outputs nothing but the state of each unit, never the position or goal of any individual. An individual reads only two numbers from the space, whether it can pass here and whether it can stay here. Where to go remains the individual's own decision, made by its own behavioral rules. Whether the space has met its purpose is judged by outcome alone, counted in the four interaction events. At every moment every individual has at least one passable route from where it stands to an exit, and a unit that would cut that route does not descend. The earlier utility framework, in which each component computed its own utility and objectives could conflict, was not implemented and is listed under Open questions.

The two costs follow the earlier design. The cost of passing is set by body scale and spatial form. A unit that has descended is a wall to a person, who must go around it, and no obstacle to a cat, which passes underneath. The cost of staying is set by who is nearby and is recomputed every beat. When someone comes within the distance appropriate to the relationship, the cost rises, and when it exceeds the individual's threshold the individual leaves. Together the two costs give an asymmetry. A cat can go everywhere a person can and further, but does not linger, while a person is limited by size but can hold a position once inside. The simulation has to show this asymmetry, and that is its first acceptance test.

Component input is the accumulation and decay of behavioral traces. Traces deepen with use and fade with disuse, and constitute the sole information channel between the spatial and behavioral layers. From this follows the phenomenon of central interest to the project, hysteresis. A component that has changed form in response to traces does not revert the moment the bodies leave, because traces fade more slowly than bodies exit. A unit a cat is standing on is also pinned, and does not withdraw until the cat leaves. Hysteresis therefore lives in the trace layer and in occupation, not in the structure. The structure itself is reversible. A locked bond can be released, and a strip that has straightened can form a different shape next time. A release probe confirmed it. A strip that forms, straightens and forms again from a different bond map matches a fresh strip to the last digit, and the fabric itself keeps no memory. → Log 2026-10-06

The three relationships share one loop and differ only in site, purpose and accounting:

The simulation explores the consequences of hand-written rules and does not predict a real cat café. Everything it can say is a conditional comparison. If visitors keep the distances the literature gives for strangers and cats behave by the distance distribution of the 1988 experiment, then under one rule set live units score so much against fixed shelving and an empty room. The seconds mean nothing on their own. What carries meaning is the difference between conditions, and only under this rule set. The simulation finds conflicts between rules before anything is built, finds the conditions under which responsiveness is worth having, and finds what cannot be done. The behavioral rules themselves are its inputs and are not verified by it. Whether real cats walk this way can only be seen in a real café.

Formal definition of interaction events

All interaction outcomes are classified into four formally defined event types, which serve as evaluation metrics:

• Co-gazing: both parties oriented toward the same visual target, height differential permitted.
• Co-warmth: one party enters and remains within the other's 0.5–1.5 m comfort band, without physical contact.
• Co-touch: physical contact occurs, the escalated state of co-warmth.
• Crossing: one party's movement path passes through the other's body domain.

A null event is additionally defined, a signal sent and unanswered. Null events are kept as data and serve as the trigger condition for spatial-layer intervention. Where the behavioral layer's invitation fails is where the spatial layer should act. Invitation is not yet among the human behaviors in the plan-view benches, so this trigger is not yet implemented.

Behavioral clauses and the interaction matrix

Behavioral-layer rules are organized as numbered clauses: C1–C8 for the cat, H3–H4 for the human, each corresponding to an empirical finding. Of the clauses, what enters the simulation is the distance thresholds, as parameters of the two costs. The remaining clauses are on the boards only. The cat's high-perch avoidance rule (C3) derives from Hirsch et al. (2025), a behavioral observation of 27 domestic cats in a Swedish cat café, which found that during high-visitor periods cats significantly preferred elevated shelving and concealment.

The clauses yield the two axes of the interaction matrix. The cat axis takes three states, following the branching logic internal to the adult-cat clauses: a stranger state (governed by C1/C6, with a latency period before first approach), a familiar state (C2, which in a café comes from staff and regulars), and a high-perch avoidance state (C3, triggered when local human density exceeds a threshold). The human axis takes three behaviors: passing (the speed and density rules of H3/H4), focused dwelling (the attention state defined in C4), and active invitation (C5, the slow-blink or dual-modal visual-vocal signal). Of the three, passing and dwelling exist in the plan-view benches, and invitation does not yet. C8 remains outside the matrix as a global modulator, sorting individuals into avoidant and proactive types and adjusting transition probabilities and temporal parameters across all cells.

The crossed axes produce nine cells, each holding one trigger rule and one dominant event. Three cells illustrate the matrix's explanatory reach. In familiar × passing, a familiar person's path enters within 0.5 m of the cat, the cat holds position, and a crossing completes. Paired with stranger × passing, identical geometry with the opposite outcome, it provides the minimal evidential unit for the claim that relationship state alters spatial right-of-way. In high-perch avoidance × passing, the cat occupies height, the person passes below, and interaction reduces to co-gazing across a height differential. This cell translates directly into a structural requirement. A perch fixed above 2 m prevents escalation, and the units' capacity to lower a platform into the overlap band (0.4–1.2 m, the interval where the two species' body heights coincide) is the exit this cell requires.

Only one of the nine cells contains a genuine inter-clause conflict, high-perch avoidance × active invitation, where the approach pull of C5 and the avoidance push of C3 act simultaneously and the source clauses provide no rule of combination. That cell has no resolution yet (see Open questions).

The spatial body

The interaction medium is the ceiling. The structural scheme is the contracting, tensioned unit. A strip of fabric hangs from the ceiling around a core. When the core contracts, the fabric goes slack, and bonds laid out on the fabric in advance catch the slack and fasten it. Where the bonds fasten decides the shape. One unit has a single degree of freedom, the contraction ℓ. Four canonical shapes come from four bond maps: a pocket, a bulb flange, a straight ledge and a stepped box (figure below). Strips stand in a ring around a vertical axis, and twenty strips contracting together fasten into a ring-shaped platform. Rings stand in a grid, and sixteen of them fill a room (Labs 2-5 and 2-8). At the real scale of Lab 2-8 the underside of the platform is 1.08 m above the floor, inside the 0.4–1.2 m band where the two species' body heights coincide.

The unit is passive, with no actuated members. Its shape is set by the bond map and the contraction. The final state is a path-dependent numerical solution. Bonds lock one by one in order, each step follows from the last, there is no closed formula, and the same input gives the same result every time. Four criteria governed the selection of the structural scheme: cellular modularity, reproducible solving, support by a mathematical model or design tool, and publication as research methodology rather than as artwork. Candidates evaluated and excluded include pneumatic truss systems (rigid-member geometry inconsistent with draped textile morphology), telescoping column systems (floor-mounted, in conflict with the ceiling-as-medium premise), and the earlier passive net-and-winch scheme, in which cable lengths set the net form and which the contracting unit replaced.

The project makes one assumption about the structure, that bonds are controllable fasteners. Which pairs fasten and when they release is decided by control, and where the fabric is clamped to the core also follows the program. This differs from passive bonds that catch on their own. The assumption is what lets one strip change shape within the vocabulary, and what makes withdrawal by releasing bonds possible. On a physical unit it remains to be verified.

The project's research contribution is the loop itself. Bodies leave traces, the space reads the traces and changes form, the form changes the cost of each position, and the bodies take a different path and leave new traces. The contracting unit is an instrument built for this, and the behavioral rules are sourced input.

Comparison results

The same room and the same people and cats were run through three spaces: live units, units fixed in place (the equivalent of fixed shelving), and an empty room with nothing in it. Each run lasts five minutes and is averaged over thirty-two seeds, at three crowd levels: three visitors and one cat, six and two, twelve and three. Four readings are taken. Co-warmth is the seconds a person spends stopped beside a cat, co-touch the seconds closer than half a metre, and the other two are the seconds cats spend settled and the metres visitors detour.

From six visitors up, live units give people more time beside cats than either fixed shelving or the empty room. With six people and two cats the figures are 861 seconds against 735 and 704, and with twelve people and three cats 2604 against 1980 and 1998, an effect size d of about 1.8 with a permutation test p below one in a thousand. How strongly a crowd sends cats up onto the units has no measurement, and the result holds under all three settings tried. With three visitors the three spaces cannot be told apart, and the device shows its effect only once the room is busy enough.

Cats settle for about as long on either device, and longer than in the empty room. At low and medium crowding, live units give visitors the shortest detours, and at twelve people they detour somewhat more than with fixed shelving and still less than in the empty room.

Contact is the cat's own decision. With guidance switched to the unit beside the seat, lowering the whole path at once, co-touch at twelve people exceeds fixed shelving, 443 seconds against 337, and matches or slightly trails the empty room. The two guidances lean different ways. The unit in front of the seat raises co-warmth, the unit beside the seat raises co-touch, and one guidance can only lean one way, a direct consequence of the two distance bands excluding each other. The tests were not corrected for multiple comparisons, and differences with p between .01 and .05 are read as direction only.

Technical implementation

The structural engine was prototyped in Python and NumPy and has been closed. The version on this site is a line-by-line TypeScript port, and guard tests check it against the Python results to the last digit. Form is computed with position-based dynamics, Verlet integration plus constraint projection, which is the modern form of dynamic relaxation from the form-finding literature. The force density method (Schek, 1974) is its lineage. The behavioral layer (trace field, unit activation, human and cat bodies) is written in TypeScript in the same site. The plan-view simulation represents space as a grid with one forming degree per unit. The solid simulation uses the real strip geometry. The 6×6 and 8×8 grids in the plan-view benches are the 4×4 real units scaled down proportionally, and small structural units have not been verified. The five rules and the two costs are in the simulation (Lab 2-14). People and cats share one room, units descend and withdraw by the rules, the four events are counted live, and the comparison study runs the same code offline. All rules are hand-written, human-readable clauses, with no dependence on sensors, computer vision, or hardware. Visualization happens in the browser, with SVG and canvas for plan views and dependency-free WebGL for the solid units.

Six revisions

From social withdrawal to a lived space (2026-06-22). The starting point was a social mechanism of collective withdrawal, complete as a skeleton but further from the author's own experience the longer it was written. The project started again from a space the author knew first-hand, and the subject settled on how people and animals share a room. Since then the questions have started from concrete spaces. → Log 2026-06-22

From two clocks to one (2026-07-27). The design had a fast clock and a slow one, and the slow one held the most striking result, attendance collapsing and never recovering. The slow clock was removed because the subject is bodies in a room over minutes, and the rule book should not carry two literatures for one story. What is left is minute-scale body dynamics, with the space always a step behind the bodies. → Log 2026-07-27

Thresholds without data are marked as inferred (2026-07-28 to 07-29). The distance needed had never been measured, so it was written as a probability spread and marked as inferred. The breed assumption was dropped against the data, and kittens became a class of their own because they cannot jump high, which broke a rule already written for adult cats. The three grades of evidence date from here. → Log 2026-07-28 → Log 2026-07-29

From a hanging net to a contracting skin (2026-08-04 to 08-17). The net coupled too strongly across its surface, and one pull moved everything. It was replaced by a strip, one contraction, and bonds that decide the shape. The unit begins here, with form written onto the fabric in advance and one engine reused across every structural bench. → Log 2026-08-04 → Log 2026-08-17

Memory moves from the fabric to the traces (2026-10-06). Memory used to come from bonds locking for good. A release probe showed that a strip which straightens and forms again matches a fresh unit to the last digit, so the fabric keeps no memory. Memory can only be told from traces, which fade more slowly than people leave, and from a unit staying down while a cat stands on it. The structure is reversible as a result, and the same unit can form a different shape next time. → Log 2026-10-06

From ratifying traces to a space with a purpose (2026-10-04 to 10-07). A space that only ratified traces lost to fixed shelving on time spent beside a cat. The revision gave the space a direction, and five rules confine it to changing costs. The old rule is kept as the control. → Log 2026-10-07

Open questions

• The rule of combination for the conflicting cell. In high-perch avoidance × active invitation, the approach pull of C5 and the avoidance push of C3 act at once. The earlier proposal used Miller's (1944) approach-avoidance gradient model and the escape economics of Ydenberg and Dill (1986) to define the outcome as an equilibrium height z* on the vertical axis, to which a unit would lower a platform. That rule has no data behind it, and C3 itself has none yet, so it is not used as a basis.
• The vertical dimension. The plan-view simulation does not model height. Cats going up and platforms coming down both wait for the solid units to be driven by the behavioral layer.
• The utility framework and the five verbs. One utility per component with conflicting objectives, and the five operations of merging, migrating, luring, splitting and dissolving, remain ideas.
• Small structural units. The 8×8 plan grid is only a proportional reduction. Units that small need a round of their own.
• Observation in a real cat café. Every conclusion from the simulation holds only under this rule set. Testing it would take a café that already has fixed shelving, one movable platform added, and the same events counted. That is a separate line of research.
• How to define separation. In the avoidance relationship, what measures "no direct conflict" is still open. The current candidate is a period in which no person comes within the animal's flight initiation distance while people can still see the animal.

Project information

References

• Ellis, S. L. H., Rodan, I., Carney, H. C., Heath, S., Rochlitz, I., Shearburn, L. D., Sundahl, E., & Westropp, J. L. (2013). AAFP and ISFM feline environmental needs guidelines. Journal of Feline Medicine and Surgery, 15(3), 219–230.
• Goffman, E. (1963). Behavior in Public Places. Free Press.
• Hall, E. T. (1966). The Hidden Dimension. Doubleday.
• Hirsch, E. N., Navarro Rivero, B., & Andersson, M. (2025). Cats in a cat café: Individual cat behavior and interactions with humans. Animals, 15(22), 3233.
• Mertens, C., & Turner, D. C. (1988). Experimental analysis of human–cat interactions during first encounters. Anthrozoös, 2(2), 83–97.
• Miller, N. E. (1944). Experimental studies of conflict. In Personality and the Behavior Disorders. Ronald Press.
• Schek, H.-J. (1974). The force density method for form finding and computation of general networks. Computer Methods in Applied Mechanics and Engineering, 3(1), 115–134.
• Sorokowska, A., et al. (2017). Preferred interpersonal distances: A global comparison. Journal of Cross-Cultural Psychology, 48(4), 577–592.
• Stanton, L. A., Sullivan, M. S., & Fazio, J. M. (2015). A standardized ethogram for the Felidae: A tool for behavioral researchers. Applied Animal Behaviour Science, 173, 3–16.
• Weston, M. A., McLeod, E. M., Blumstein, D. T., & Guay, P.-J. (2012). A review of flight-initiation distances and their application to managing disturbance to Australian birds. Emu, 112(4), 269–286.
• Whyte, W. H. (1980). The Social Life of Small Urban Spaces. Conservation Foundation.
• Ydenberg, R. C., & Dill, L. M. (1986). The economics of fleeing from predators. Advances in the Study of Behavior, 16, 229–249.