Abstract

Generative world models produce navigable structure without pre-authored structure. They therefore supply new test cases for the claim that hypertext is structural all the way down. We isolate one construct in this class of systems that extends hypertext theory rather than relabels it: the edit-distance link ( λ δ ). A λ δ link's representational weight lies on the transformation between its endpoints, carried by an explicit discrete channel, while invariant content persists through a continuous conditioning channel. We develop the construct through a close reading of Δ -IRIS, whose architecture—a discrete autoencoder that encodes stochastic inter-step deltas, paired with an autoregressive transformer that summarizes persistent state through continuous tokens—makes the two channels structurally visible. We distinguish λ δ from five neighbouring constructs: argumentation links, Storyspace guard fields, operational transforms and CRDTs, version-control diffs, and learned video codecs. We then show that the hypertext apparatus does real work: typing the transition as a link produces a reader-traversal criterion, a provenance regime, and an authorship locus that the ML term “context-aware tokenization” does not by itself entail. We read the construct as a specification of Halasz's computed-link programme rather than a displacement of it, and we locate the shared-content presumption λ δ revises in three post-Halasz works (Dexter, DeRose, RDF). We close with a prediction designed to discriminate the framing from a finite-token-budget autoregression null: at matched emitted-token budget, loop-closure inconsistency in conditional-tokenization world models should correlate with the variance of per-step delta magnitude—a property the null treats as irrelevant.

1 Introduction

Thirty-nine years after Conklin [8] codified the typed link as hypertext's atomic unit, a class of systems has appeared whose transitions invite re-reading against the received taxonomy. Generative world models—DreamerV3 [12], DIAMOND [1], Δ -IRIS [18], Genie 3 [11]—produce latent states and transitions between them. Most transitions, read hypertextually, fall into established categories: state graphs are graphs, Monte Carlo tree search is branching narrative, compounding error is lost-in-hyperspace. These are relabelings.

One transition type, in one architectural family, resists the relabeling and earns a name. In conditional-tokenization architectures such as Δ -IRIS, each predicted next-state is encoded conditionally on the previous state and action, and the tokenizer is explicitly trained to encode the inter-step delta. Representational capacity on the explicit discrete channel is therefore spent in proportion to what changes; invariant content is carried through a separate continuous conditioning pathway. We call the resulting typed transition an edit-distance link (λδ). This is the case where the hypertext apparatus earns its place. Once we type the transition as a link, three things follow that the ML term “context-aware tokenization” does not by itself entail. We get a reader-traversal criterion. We get a provenance regime. We get an authorship locus.

We are careful about scope. The construct appears cleanly in conditional-tokenization world models today, and we will argue (§6) that it is a partial reading at best of diffusion world models (DIAMOND) and of recurrent-latent world models (DreamerV3). But the stakes are broader than one architecture. Conditional encoding is a compute-efficiency frontier for generative rendering [18], and future systems that render navigable worlds on demand will face the same economic pressure to allocate representational weight on the differential channel. λδ is the first case in which a hypertext community long accustomed to content-anchored link semantics must reckon with links whose explicit representational weight lies on transformation.

The paper makes three contributions to the Principles, Directions, Reflections track. First, we define λδ operationally, distinguish it from five neighbouring constructs, and supply a criterion ML compression theory does not: a reader-traversal criterion (§2). Second, we read the construct as a specification of Halasz's computed-link programme [13]—not a displacement of it—and name the exact post-Halasz works (Dexter [14], DeRose [9], RDF [4]) in which the shared-content presumption λδ revises was installed (§3). Third, we close with a prediction designed to discriminate the λδ framing from a finite-token-budget autoregression null (§5).

2 The Edit-Distance Link

2.1 Δ -IRIS, accurately described

Our exemplar is Δ -IRIS [18]. We describe its architecture precisely, because an earlier sketch of this argument was read as claiming that Δ -IRIS discards invariant content. It does not, and the distinction matters for the construct.

IRIS [17] learns a discrete codebook of visual tokens and predicts future token sequences with a Transformer. Each frame is encoded independently. Δ -IRIS changes this in two coordinated ways. First, its discrete autoencoder encodes frame t + 1 conditionally on frame t and action at; the authors describe this as encoding the “stochastic delta” between successive time steps, and the training objective shapes the discrete codes to carry inter-step change rather than whole-frame content [18]. Second, its autoregressive transformer does not autoregress over the discrete delta tokens alone. It additionally summarizes the current state of the world through a stream of continuous tokens that condition future delta prediction. The architecture has two channels: a discrete explicit channel carrying deltas, and a continuous conditioning channel carrying persistent summary.

The hypertextual claim we want to make is narrower than the claim the earlier sketch could be read as making. We are not saying invariant scene content is “not linked,” nor that it is absent from the model's representation. We are saying that the weight of explicit discrete representation allocated to the link between consecutive states scales with the magnitude of inter-step change, while common content persists through the continuous conditioning channel. A classical typed link is anchored in shared endpoint content on a single channel. A λδ link is structurally two-channel: explicit-on-transformation, implicit-on-persistence.

This is not a generic property of transformer world models. IRIS (non-delta) does not have it. DIAMOND's diffusion process does not have it. DreamerV3’s recurrent latent has a related but different structure (§6). The property is architecture-specific, and naming it at the link level is what lets hypertext theory engage it.

2.2 Operational definition

Let n1, n2 be nodes in a hypertextual graph with a learned link ℓ(n1, n2). We call ℓ an edit-distance link, λδ(n1, n2), when all four clauses hold:

Learned operator. The link's semantics are specified by a transformation n2⊖n1 whose operator ⊖ is learned from a predictive objective, not authored.\

Channel asymmetry. The representational capacity of an explicit discrete channel scales with the magnitude of the transformation; persistent content shared between n1 and n2 is carried through a distinct implicit conditioning channel rather than the explicit channel.\

Arbitrary-overlap tolerance on the explicit channel. The explicit channel is meaningful and well-formed even when n1 and n2 share an arbitrarily small fraction of their explicit content; the implicit channel absorbs the asymmetry.\

Reader-traversal. The reader's movement between n1 and n2 enacts the transformation: traversal is application of ⊖ under reader-supplied conditioning (actions, inputs), not retrieval of a stored destination.\

Clauses (i)–(iii) are representational claims about the architecture. Clause (iv) is the one the hypertext apparatus supplies, and it is what separates λδ from learned video compression.

A learned video codec satisfies (i)–(iii): a learned inter-frame predictor is a learned operator, it allocates bits to motion rather than static background, and it handles arbitrarily small inter-frame overlap. But the viewer of an MPEG- or neurally-encoded video never traverses the operator. The delta-stream is an engineering artefact hidden behind a full-frame playback surface. The viewer engages frames, not operators. In a generative world model the reader's actions are precisely what instantiate each transition, so the reader engages the operator directly. This is the Bush/Engelbart commitment: hypertext is defined by traversal, not storage format [7, 10]. Clause (iv) writes that commitment into the link-type definition.

One clarification about reader phenomenology. A λδ reader does not perceive the explicit discrete channel. They perceive full rendered frames (in a pixel-space system such as Genie 3) or full decoded states (in a token-space system such as Δ -IRIS). The two-channel structure is a claim about the architecture and the representational economy of the link, not about what the user sees. The hypertextual consequences we draw—for interpretation, provenance, and authorship—derive from the architecture, not from a phenomenological claim that the reader “sees only transformations.”

2.3 Distinctions from neighbouring constructs

The hypertext and collaborative-systems literatures already contain transition-types that make difference partly constitutive. We distinguish λδ from five of the closest. The distinction in each case is load-bearing: the literatures are not wrong to have treated difference as primary in these ways, but none of them supplies what conditional-tokenization world models supply.

Argumentation links. gIBIS [8] and Trigg's TEXTNET taxonomy [9] include difference-typed relations: refutes, objects-to, questions, supports, alternate-version. In each case the relation is a predicate over a shared conceptual object—two nodes linked by refutes concern the same claim, two nodes linked by alternate-version concern the same underlying work. The difference is semantic, registered as a label. A λδ link has no labelled predicate. Its content is the operator itself. Argumentation links are typed by a finite vocabulary of relations; λδ links are typed by a learned continuous space of operators. The two do different work.

Storyspace guard fields. Storyspace [5] conditions transition-availability on reader state via authored guard expressions, which encodes a differential condition on traversal. But the guard is authored and symbolic; the transition it enables still connects lexia whose content is the basis of the connection. Guard fields violate clause (i) of §2.2: the operator is specified by an author, not learned from data. The hypertextual move Storyspace makes—letting transitions depend on what has happened—is conceptually adjacent and historically important, but it lives at a different level: the reader's trajectory through authored content, not a learned transformation between states.

Operational transforms and CRDTs. Operational transforms (OT) and conflict-free replicated data types (CRDTs) treat edits as first-class objects in collaborative editing. The analogy is the closest of the five: an OT operation is explicitly a transformation between document states, and CRDT theory is at root a theory of composable deltas. But OT and CRDT operators are designed to satisfy merge-laws—commutativity up to transformation, convergence—and their semantics are designer-specified for deterministic merge. λδ operators are learned, stochastic, and not required to satisfy merge-laws; their semantics come from a predictive training objective. OT/CRDT violates clause (i). What the collaborative-editing literature does offer λδ is a mature vocabulary for reasoning about edit-composition, and we flag this in §7 as a direction for the composed-operator link.

Version-control diffs. Diffs represent deltas between document states for presentation and storage. A diff is a description of a difference relative to fully-stored source and destination; the link structure is anchored in the persisted states, and the diff is a human-readable projection. Diffs violate clause (ii): they do not carry common content through an implicit conditioning channel—they carry it explicitly by storing both endpoints in full, with the diff sitting alongside as annotation. The channel asymmetry in λδ is a property of the link's internal representation, not a projection for display.

Learned video codecs. Neural video codecs with learned inter-frame prediction satisfy (i)–(iii). They violate (iv): no reader traverses them, and the operator is hidden behind a full-frame playback surface. This is the cleanest discrimination, because it shows that (iv) does genuine work. The architectural ingredients of λδ are present in neural compression; what promotes them to a link type is traversal.

None of these prior constructs satisfies all four clauses. Several come close on different axes. The hypertext and collaborative-systems literatures have long recognized that difference can be primary in linking; λδ names the specific case in which the operator is learned, the representation is two-channel, and traversal is what instantiates it.

3 A Specification of Halasz, and a Closer Reading of Bush, Nelson, Engelbart

3.1 Halasz, precisely

An earlier draft of this argument described λδ as “extending” or “revising” Halasz's programme [13]. That framing overreaches. Halasz's “Seven Issues” is explicitly a call for hypermedia systems to support computed links and virtual structures—link structures that do not pre-exist as authored artefacts but are generated from an underlying model or query. λδ links are a specific instance of what Halasz called for: computed (by a predictive objective), virtual (not pre-authored), and structurally uniform (every link is of the same type, generated the same way).

We reframe the paper's contribution accordingly. λδ is not a revision of Halasz but a specification of one branch of his programme under a constraint he did not anticipate: generative tokenization. What λδ does contest is a presumption that entered hypertext theory after Halasz, in the formalization of link anchoring. We locate this presumption in three specific places.

The Dexter Hypertext Reference Model [14]—co-authored by Halasz himself, and the most influential formalization of link structure in the field—defines a link as a relation between endpoints identified by anchors into content. The model is neutral on link typing, but its data structure installs the shared-content presumption at the level of representation: a link is a pair of content-pointers. DeRose's “Expanding the Notion of Links” [9] then types links by semantic relation—aggregation, annotation, alternate-version, rhetorical—each presupposing that endpoints share a conceptual object over which the relation holds. Semantic-web RDF [4] naturalises the same shape as subject–predicate–object triples, in which the predicate is a relation over entities that must, to be related, be co-denoted by something.

Read together, Dexter, DeRose, and RDF are where the shared-content presumption became the default grammar of typed linking. None of the three is wrong. Each is a correct formalization of its era's dominant link-type population. λδ specifies a link type these formalizations do not accommodate: one whose signature is a transformation rather than a predicate, and whose anchoring is to an operator rather than to content-pointers. This is the sense in which the construct extends Halasz's programme: it fills in a link-type slot his call anticipated but his successors, working with the dominant link populations of their time, had no occasion to model.

3.2 Bush, Nelson, Engelbart: specifically contested

A PDR contribution should do more with the canon than cite it. We read each of Bush, Nelson, and Engelbart as specifically contested by λδ, and we are careful to distinguish contestation from displacement.

Bush's trails [7]. Bush assumed each trail-step preserved a content pointer. The trail was a persistent record of reader associations, and its value lay in what it pointed back to. A λδ -structured trail preserves an operator, not a pointer. Its value is the composition of transformations it records. Bush's trail and a λδ trail are different kinds of memoranda: associative-retrieval in Bush, transformational-replay in λδ. The contestation is not that Bush was wrong but that his model of what a trail records does not exhaust what a trail can record.

Nelson's transclusion [19]. Nelson's theory rests on preserved identity: the docuverse links by reproducing, so that traversal from source to destination recovers something identifiable across both endpoints. λδ inverts the relation. What is identifiable across endpoints is the applied operator, not a preserved fragment; the implicit conditioning channel carries persistence but not in a form Xanadu-style provenance can address. Transclusion-style provenance is therefore not directly available for λδ links. This is a design constraint on any future Xanadu-like system that would incorporate generative link types, and we treat it in §3.3.

Engelbart's augmentation [10]. Engelbart framed human–computer systems as extending capacities the user already possesses. λδ links augment in a mode he did not theorize. The user is augmented into performing transformations whose operator is learned and not directly inspectable. Augmentation via inaccessible capability is a mode Engelbart's vocabulary has no term for. The contestation is the mildest of the three: Engelbart's programme is consistent with λδ, but needs a new term.

3.3 Interpretation and provenance

Two consequences follow for the reader.

Interpretation reconstructs transformations. Classical hypertext interpretation recovers a predicate. The reader asks what the author intended by a connection, and the answer has the form “these nodes share X.” λδ interpretation has no authored predicate to recover. The reader asks what operation takes n1 to n2, and the answer is the operator itself—learned, and not directly inspectable. The interpretive act is hypothesis-testing from examples, in the sense Iser described for Leerstellen-filling [15]: the reader infers the rules that make observed transitions coherent. Iser's reader fills gaps against an authorial intention; the λδ reader fills gaps against a learned distribution. Reader-response theory acquires a new case: interpretation under opacity, where the ground truth of the transformation is inaccessible in principle.

Provenance becomes operator-tracking. A provenance-preserving λδ hypertext cannot record transclusions. It must record operator applications. The required machinery has three parts. (a) A per-transition log capturing the operator's action as a human-interpretable summary—analogous to a commit message, but generated, not authored. (b) A translation layer that maps the operator's internal representation to natural-language or schema-level descriptions: “lane change,” “item collected,” “topic shifted.” Recent work on concept-level interpretability makes this tractable but not trivial. (c) A reverse-operator that permits audit: given n2 and the logged operator, the system must reconstruct a plausible n1. No current generative world model guarantees (c). It is the concrete open problem the construct puts on the HT community's agenda—and it is the place where OT/CRDT theory is most likely to contribute, because the reverse-operator question is structurally a merge-and-invert question the collaborative-editing literature has studied.

4 Community Positioning

The argument takes its shape from a turn in HT scholarship that treats AI-mediated content as a new object of study. Nuernberg and Atzenbeck [20] argue that structural hypertext is foundational “all the way down.” Generative world models produce navigable structure without pre-authored structure; λδ is the specific structural novelty this mode of production introduces in one of its architectural families.

Liu and Almeda [16] describe hypertextual friction as the cognitive cost algorithmic mediation imposes on reader agency. λδ links supply a specific mechanism for this cost. Under revisitation of a location in a conditional-tokenization world, the reader cannot rely on referential integrity: the operator chain that returns to a location is not guaranteed to reconstruct the same explicit content, because persistence lives on the implicit channel and is subject to its finite conditioning budget. The λδ reader is not denied a persistent view—Genie 3 and Δ -IRIS both render full frames or full states. What they are denied is any guarantee that a revisited location retains the content it previously held. This is a specifically λδ form of friction, distinct from the recommender-algorithm form Liu and Almeda analyse.

Bernstein [6] asks what it feels like to read alongside a machine. A λδ reader is augmented into performing operator-applications they cannot introspect. The phenomenology is not that they see only transformations—they see full rendered worlds—but that the ground they stand on is guaranteed only locally. Augmentation here comes with a specific loss: the user gains traversal capacity they cannot articulate, and loses referential anchoring they used to take for granted.

Antonini et al. [3] raise the authorship question. We take a position specific to λδ. Authorship under conditional-tokenization composition is not the generic claim that “the training corpus authors the text.” It is narrower: authorship resides in the distribution of deltas the training data contains—the set of transformations the operator has been shaped to reproduce. A Δ -IRIS trained on Crafter is authored by the delta-distribution Crafter states exhibit: short-horizon, sprite-level, action-conditioned. A driving world model is authored by the delta-distribution of driving scenes. Two models trained on different trajectory distributions are different authors, even with identical architectures. This is authorship-as-trajectory-curation, and we offer it as a candidate aesthetic category for λδ hypertexts rather than as a claim that displaces existing accounts of machine-mediated authorship.

Anderson [2] on spatial hypertext is the neighbouring case: Genie 3’s environments instantiate spatial hypertext whose space is synthesised rather than laid out, and λδ —to the extent it applies to Genie 3’s tokenizer, which is not publicly documented—is the link type that would make the synthesis legible.

5 A Discriminating Prediction

An earlier draft predicted that cyclic world-model trajectories should degrade faster than matched linear trajectories at equal length. That prediction is overdetermined. A finite-context autoregression null—FIFO eviction of KV-cache entries, positional-encoding extrapolation limits, training-distribution bias toward forward motion—predicts the same degradation with no λδ machinery required. We sharpen the prediction to one that actually discriminates.

The sharpening hinges on one move: measure the context in emitted tokens, not seconds or actions. In a conditional-tokenization system, high-delta content emits more discrete tokens per unit wall-clock than low-delta content. So matching wall-clock duration does not match context-window occupancy, and neither does matching action-count, since action tokens and emitted visual tokens live on different budgets. To discriminate the λδ framing from a finite-token-budget AR null, the test must match emitted-token count and look at a property of the trajectory the null is indifferent to.

Null hypothesis (H0, finite-token-budget AR). At matched total emitted-token count, loop-closure inconsistency in a conditional-tokenization world model depends only on (a) total token count and (b) the recency-ordering of first-visit tokens in the context window. The distribution of tokens across the trajectory—specifically, the variance of per-step delta magnitude—is irrelevant: the context window is a flat buffer.

λδ prediction (H1). At matched total emitted-token count, loop-closure inconsistency correlates with the variance of per-step delta magnitude along the trajectory. A trajectory of uniformly moderate deltas should reconcile loop closure more cleanly than a trajectory of spiky deltas—alternating low-motion and high-motion segments—with the same total token budget. The operator-path the model must reconcile carries higher entropy in the spiky case even at matched representational budget, because the channel-asymmetry structure of the link means that the load on the implicit conditioning channel is non-uniform across the trajectory.

The discrimination is precisely this: under H0, total emitted-token count is a sufficient statistic for degradation; under H1, it is not.

Protocol. (a) Trajectories. Matched cyclic paths of fixed loop geometry, generated in pairs—one uniform-delta, one spiky-delta—and truncated to matched emitted-token budgets. Delta magnitude is operationalized as inter-frame discrete-token turnover under Δ -IRIS's tokenizer applied offline to the rendered frames. Spikiness is operationalized as the variance of per-step turnover at matched total. (b) Outcome. Time-to-first-inconsistency at revisited landmarks, scored by two blinded coders against a pre-registered rubric (identity, position, colour, count), with inter-rater reliability reported as Cohen's κ. (c) Controls. Seed, build version, action cadence, frame rate, prompt, rendered scene content (same scene, different trajectory). (d) Minimum N. 20 trial-pairs; power analysis deferred to registered report.

A confirmation supports H1 and the λδ reading. A disconfirmation—delta variance irrelevant at matched token budget—is strong evidence for H0 and indicates that the λδ framing has been overextended. We flag a replication constraint: the protocol requires running Δ -IRIS's tokenizer on frames rendered by a second system (e.g., Genie 3 via Google AI Ultra), which is a cross-architecture operationalization of delta magnitude. A cleaner test would run entirely within a single conditional-tokenization system whose tokenizer is the one whose channel-asymmetry is under test. The single-system version of the protocol is available today for Δ -IRIS on Crafter; the cross-system version is what lets the construct be probed in production-grade systems like Genie 3.

6 Limits

The construct is architecture-specific in its cleanest form. Δ -IRIS satisfies all four clauses of §2.2. DIAMOND [1], which operates in pixel space via diffusion, jointly denoises full frames and does not exhibit clause (ii)’s channel asymmetry: there is no explicit discrete channel whose capacity scales with inter-step change. DreamerV3 [12] maintains a persistent recurrent latent ht alongside a stochastic component; the persistent latent is closer to classical stored state than to the implicit conditioning channel λδ describes, and we read the construct as a partial rather than dominant reading of DreamerV3 transitions. Genie 3’s tokenizer is not publicly documented in the detail needed to verify clause (ii), though its behaviour under cyclic trajectories is consistent with a conditional-tokenization architecture; we treat it as a plausible but unconfirmed case.

Neither DIAMOND nor DreamerV3 refutes the construct. Each clarifies its scope. Conditional tokenization is the cleanest architectural home for λδ today, and the compute-efficiency pressures on generative rendering make it a likely home for future systems. A separate question—whether analogous channel-asymmetry structures emerge in diffusion systems via latent-trajectory parameterization, or in recurrent systems via gated-update mechanisms—is open.

A second limit is operationalization. The variance-of-delta prediction in §5 requires a tokenizer external to the system under test when the system's tokenizer is not accessible. This is a methodological compromise, not a conceptual one: the prediction is about the architecture of the system under test, and the external tokenizer is a proxy for its delta measure. We flag it because a stronger version of the test would require tokenizer access that open-weights world models currently provide (Δ -IRIS) but production systems generally do not.

7 Conclusion

What does generative world modelling contribute to hypertext theory that is not already there? In one architectural family—conditional-tokenization world models, cleanly exemplified by Δ -IRIS—it contributes one construct: the edit-distance link. Its explicit representational weight accumulates on the differential between endpoints; its persistent content lives on a separate implicit conditioning channel; its operator is learned; its traversal is what the reader does. Its four-clause definition turns on a criterion only the hypertext apparatus supplies—reader-traversal. It specifies one branch of Halasz's computed-link programme, and contests a shared-content presumption installed not by Halasz but by Dexter, DeRose, and RDF. It recasts Bush's trails, Nelson's transclusion, and Engelbart's augmentation in ways those frameworks can accommodate but do not yet have vocabulary for. Its provenance regime tracks operator-applications rather than transclusions. Its authorship locus is trajectory-distribution curation. Its friction mechanism is referential-integrity failure under revisitation. Its discriminating prediction—variance of delta at matched token budget—is executable today on Δ -IRIS and extensible to production systems.

We offer one candidate for a second construct and do not develop it here: the composed-operator link. Its semantics are the composition of a sequence of λδ operators along a trajectory. Its interesting property is that composition of learned operators does not in general preserve the commutativity or invertibility assumptions classical trails take for granted. Here the collaborative-editing literature—OT and CRDT theory, with its mature account of operator composition and convergence—is the natural partner discipline. We expect the hypertextual consequences of non-commutative traversal to be substantial, and we note that the partnership would be the first case in which the HT community's theory of typed linking and the collaborative-systems community's theory of edit-composition meet on common ground.

Source

ACM full-text HTML

ACM Digital Library

Imported from ACM’s structured HTML source. ACM Reference Format: Raghav Sharma, Manan Mehta, Vineeth Kalluru, and Adit Kotwal. 2026. Edit-Distance Links: A Hypertextual Reading of Conditional-Tokenization World Models. In 37th ACM Conference on Hypertext (HT '26), September 14--18, 2026, London, United Kingdom. ACM, New York, NY, USA 6 Pages. https://doi.org/10.1145/3800935.3830861

References

[1] Eloi Alonso, Adam Jelley, Vincent Micheli, Anssi Kanervisto, Amos Storkey, Tim Pearce, and François Fleuret. 2024. Diffusion for World Modeling: Visual Details Matter in Atari. In Advances in Neural Information Processing Systems, Vol. 37.

[2] Mark Anderson. 2025. W(h)ither Spatial Hypertext?. In Proceedings of the 36th ACM Conference on Hypertext and Social Media. ACM, New York, NY, USA. https://doi.org/10.1145/3720553.3746683 — ACM HyperText copy

[3] Alessio Antonini, Lucia Lupi, Mariusz Pisarski, and Sam Brooker. 2025. Literary Hypertext, AI, and Google's New Web: An Aesthetics Discussion. In Proceedings of the 36th ACM Conference on Hypertext and Social Media. ACM, New York, NY, USA.

[4] Tim Berners-Lee, James Hendler, and Ora Lassila. 2001. The Semantic Web. Scientific American 284, 5 (2001), 34–43. https://doi.org/10.1038/scientificamerican0501-34

[5] Mark Bernstein. 2009. On Hypertext Narrative. In Proceedings of the 20th ACM Conference on Hypertext and Hypermedia. ACM, New York, NY, USA, 5–14. https://doi.org/10.1145/1557914.1557920

[6] Mark Bernstein. 2025. What is it Like to be Augmented?. In Adjunct Proceedings of the 36th ACM Conference on Hypertext and Social Media. ACM, New York, NY, USA, 15–20. — ACM HyperText copy

[7] Vannevar Bush. 1945. As We May Think. The Atlantic Monthly 176, 1 (1945), 101–108.

[8] Jeff Conklin. 1987. Hypertext: An Introduction and Survey. Computer 20, 9 (1987), 17–41. https://doi.org/10.1109/MC.1987.1663693

[9] Steven J. DeRose. 1989. Expanding the Notion of Links. In Proceedings of the 2nd ACM Conference on Hypertext. ACM, New York, NY, USA, 249–257. https://doi.org/10.1145/74224.74245

[10] Douglas Engelbart. 1962. Augmenting Human Intellect: A Conceptual Framework. Technical Report. Stanford Research Institute, Menlo Park, CA, USA.

[11] Google DeepMind. 2025. Genie 3: A New Frontier for World Models. DeepMind Blog. Accessed 21 July 2026. https://deepmind.google/discover/blog/genie-3-a-new-frontier-for-world-models/

[12] Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. 2024. Mastering Diverse Domains through World Models. Journal of Machine Learning Research 25, 334 (2024), 1–67.

[13] Frank Halasz. 1988. Reflections on NoteCards: Seven Issues for the Next Generation of Hypermedia Systems. Commun. ACM 31, 7 (1988), 836–852. https://doi.org/10.1145/48511.48514

[14] Frank Halasz and Mayer Schwartz. 1994. The Dexter Hypertext Reference Model. Commun. ACM 37, 2 (1994), 30–39. https://doi.org/10.1145/175235.175237

[15] Wolfgang Iser. 1978. The Act of Reading: A Theory of Aesthetic Response. Johns Hopkins University Press, Baltimore, MD, USA.

[16] Shuning Liu and Samantha G. Almeda. 2025. Agency among Agents: Designing with Hypertextual Friction in the Algorithmic Web. In Adjunct Proceedings of the 36th ACM Conference on Hypertext and Social Media. ACM, New York, NY, USA, 30–34. — ACM HyperText copy

[17] Vincent Micheli, Eloi Alonso, and François Fleuret. 2023. Transformers are Sample-Efficient World Models. In Proceedings of the 11th International Conference on Learning Representations (ICLR).

[18] Vincent Micheli, Eloi Alonso, and François Fleuret. 2024. Efficient World Models with Context-Aware Tokenization. In Proceedings of the 41st International Conference on Machine Learning(PMLR, Vol. 235). 35623–35638.

[19] Ted Nelson. 1981. Literary Machines. Mindful Press, Sausalito, CA, USA.

[20] Peter Nuernberg and Claus Atzenbeck. 2025. It Really Is Structure, All the Way Down. In Adjunct Proceedings of the 36th ACM Conference on Hypertext and Social Media. ACM, New York, NY, USA, 21–24. https://doi.org/10.1145/3720533.3750063 — ACM HyperText copy