NOTE-099: Floridi — A defence of informational structural realism

Contribution

Floridi supplies (a) a reconciliation of ESR and OSR by assigning them to different levels of reflection, using the computer-science method of levels of abstraction (LoAs, from Floridi & Sanders); (b) a moderate answer to the “no relations without relata” objection (the package hypothesis); and (c) a positive account of OSR’s structural objects as informational objects, argued to scale from sub-observables to everyday objects. It also contains ten replies to objections, including a formal-logic rebuttal of the Psillos/Papineau “collapse” argument against ESR.

Key insight

Epistemic and ontic structural realism answer different questions at different levels. “What can we know?” gets ESR’s answer: relations. “What is the least we must assume is there for that knowledge to be possible?” gets OSR’s answer: relata that are themselves structural. Once relata are allowed to be constituted by relations, and specifically by difference, the natural candidates are data in the most primitive sense, concrete points of “lack of uniformity”. Reality is then a network of informational objects known only through LoA-mediated interaction: “black-boxes inside a grey-box” (p. 45).

Assumptions

• Knowledge is indirect: LoA-mediated, a Kantian but inter-subjective and revisable schematism (“liminal realism”, p. 11). No white-box access to reality (p. 44).
• A structure is (O, P, R, T) (p. 4). The ontological problem is “what are the objects in O?”
• Ockham’s razor applies twice, once to the number of types of things (supporting ESR) and once to the nature of the relata (supporting OSR) (§2.4, reply 6.5).
• Physics and mathematics underdetermine metaphysics, which is therefore “a matter of rational exchange of arguments” (p. 28).
• Metaphysics is judged by portability, scalability and interoperability, notions imported from software engineering via FOLDOC (§4).
• Newman’s problem is explicitly set aside (§1), though Floridi acknowledges it remains “a future concern” for ISR.

Key results

• Definitions (pp. 3–6, 32): SR, ESR (models are informative “only about the relations … but not about the first-order, one-place predicates qualifying the objects in themselves”), EOSR (only structures exist; objects are “figmenta”), NOSR (entities are structural objects, indirectly denotable), SR* (SR relativised to an LoA), and ISR (SR* with “informational objects” as the relata).
• LoA formalism (§2.2): an observable is an interpreted typed variable; an LoA is a finite non-empty set of observables; a moderated LoA adds a behaviour predicate. A gradient of abstractions is {Lᵢ | 0 ≤ i < n} with relations Rᵢ,ⱼ ⊆ Lᵢ × Lⱼ that are mutually inverse, with the behaviour at Lⱼ at least as strong as the translated behaviour. Disjoint and nested GoAs are defined; in a nested GoA, the reverse of each Rᵢ,ᵢ₊₁ is a surjective function.
• Ontological commitment (§2.3): choosing an LoA commits to types; endorsing its model commits to tokens (the SLMS scheme, Figs. 2–4).
• Reconciliation (§2.4): there is no logical space between ESR and instrumentalism at first order. At a “derivative or metatheoretical” LoA, the minimal assumption about relata is that they are structural. Since the commitments are at different levels, “there is no incompatibility”.
• Regress (§3): internal relations constitute their relata, but Moore’s objection shows marriage-type internal relations presuppose their relata. Difference is the exception, backed by the modified Leibnizian principle: the logically undifferentiable necessarily does not exist. Floridi then retreats to the package hypothesis (see corrections).
• Informational objects (§4): a datum is a binary difference, “it takes two to make a datum”. Data are “purely relational entities”, particular not universal. OOP (chess pawn example) shows how structural objects combine state and behaviour and scale to macro-objects. ISR is presented as a precisification of Dennett’s real patterns (credited to Wallace 2003, fn. 20).
• Upshot (§5): ISR reconciles Wiener (“information is information, not matter or energy”) with Wheeler’s “it from bit”.
• Replies (§6): 6.1: the collapse argument (restricting to structure is no restriction, so ESR ≡ full-blown realism) is “valid but unsound”, because ESR ↔ FBR is false (they differ on Kantian humility; if they were equivalent, Newman’s objection would apply to FBR too). 6.6: a whole world is still logically differentiable (from ¬W). 6.8–6.9: data are neither purely epistemic nor readability-dependent; they are concrete differences that agents exploit. 6.10: portability between disjoint LoAs is via a translating LoA, as with a compiler.
• Epistemology (§7): knowledge is constructionist, “designing and modelling” rather than mimetic. Reference, representation and truth are relations among models, not between models and reality in itself (p. 45).

Claims

id	claim	strength	support
C1	ESR and non-eliminativist OSR are compatible because they operate at different LoAs	weak	informal transcendental argument (§2.4). The “second-order Ockham” step is asserted, and objection 6.5 presses exactly this point
C2	Not all relata are logically prior to relations	moderate	the package hypothesis (Eddington), after the stronger Ur-relation claim is withdrawn (§3). Modest, and argued
C3	Structural objects are best understood as informational objects (concrete data as differences de re)	weak	argument by analogy and methodological virtue (OOP, portability and scalability, §4). No argument excludes rival readings
C4	”The ultimate nature of reality is informational”	weak	consequence drawn from C1–C3 at the reflective LoA (§5). Floridi hedges that the objects “might well be” material, “but we have no need to suppose them to be so”
C5	The Psillos/Papineau collapse argument against ESR is valid but unsound	moderate	propositional reconstruction plus the claim that ESR ↔ FBR is false (6.1). The formal step is laboured, but the substantive point (different scopes) is argued
C6	A modified Leibnizian principle holds: the logically undifferentiable necessarily does not exist	moderate	argued as as safe as the indiscernibility of identicals (6.6), with the toy-universe discussion (§3, 6.7)
C7	ISR is scalable to macro-objects and fruitful in ethics, epistemology and philosophy of mind	assertion	referred to Floridi’s other papers (p. 32–33), “no space to show here”

Method

Conceptual and transcendental argument in the philosophy of science. It borrows a formal apparatus (LoAs and gradients of abstraction from Formal Methods, Z notation) and engineering criteria (portability, scalability, interoperability, OOP), then defends the result against ten objections received after the 2003 CAP@AU keynote.

Concepts

• Level of abstraction (LoA): a finite non-empty set of observables, i.e. interpreted typed variables. It fixes which questions are meaningful and answerable. It is a model-theoretic interface, not a metaphysical “level of being”.
• Gradient of abstractions (GoA): a family of LoAs with inverse translation relations satisfying a behaviour-consistency condition; either disjoint or nested.
• SLMS scheme: system, analysed at an LoA, generating a model, which identifies a structure. The LoA is O-committing (types) and the model O-committed (tokens).
• Structural object / informational object: a relatum constituted by relations; as a cluster of data, concrete differences de re.
• Datum: a “lack of uniformity”, a binary, symmetric relation of difference. Not alphanumeric data, and not readability-dependent.
• Package hypothesis: relata and relations co-exist, neither logically prior (Eddington 1928).
• Modified Leibnizian principle: the necessary inexistence of the logically undifferentiable.
• Constructionism: knowledge as designing and modelling, where reality is a “resource”, not a “source”, of knowledge, and structural objects are “constraining affordances”.

Connections

This is the source ISR position that Karpenko (LIT-215) claims to extend, and Karpenko misdescribes it in two places (see x3). In LIT-045’s taxonomy, Floridi’s NOSR is roughly OSR thesis 4 (objects without intrinsic natures), and the ESR he formalises is the Kantian “unknowable intrinsic properties” thesis. His package hypothesis is close to Morganti’s mutual-dependence view (LIT-173), except that Floridi keeps relations as co-fundamental. He cites Chakravartty (1998, 2003) as a non-eliminativist ally; compare LIT-196. His appeal to Dennett’s real patterns, via Wallace 2003, is the same lineage Ladyman & Lorenzetti (LIT-184) rely on. Theirs, though, is an information-theoretic compression criterion, while Floridi’s is a software-engineering analogy. LIT-136 (Floridi, informational personal identity) applies the same informational ontology to persons.

Metaphysics: the work defends an informational, non-materialist ontology of structural objects and a fundamental dualism or pluralism of differences. It presupposes the Kantian noumenal/phenomenal split. The metaphysics tag is justified as a secondary topic.

Bearing on the record

It carries nothing for ML practice. The LoA machinery resembles interface and abstraction talk in software, and its “real patterns” connection gestures at compression, but the paper makes no computational or learning claim. In this record it is the reference point for “informational” readings of structuralism. A THEORY or reading-list entry citing it for “reality is information” should add Floridi’s own hedges, “as far as we can tell”, the concreteness of data, and the dualist reading in 6.9.

Limitations

• Newman’s problem is deferred (§1). The paper’s own framing concedes this is a central threat to any SR, ISR included.
• The reconciliation’s key step, a second Ockham razor at a meta-LoA that favours structural relata, is asserted rather than derived. Why a structural nature is “simpler” than an unknown intrinsic one is exactly objection 6.5, and the reply clarifies what the levels are rather than showing the simplicity.
• The abstract says ISR is “shown” applicable to observables and sub-observables. The body offers the OOP analogy and states that it “is not a philosophical ontology but a practical example” (p. 31). Applicability is illustrated, not shown.
• The informational objects are said to be concrete differences, yet described as typed variables and OOP objects. The tension between concrete ontology and modelling vocabulary is managed by LoA-relativity, not resolved.
• The LoA formalism is rigorous in isolation, but no result about it is used in the argument. It supplies vocabulary, not premises.

Open questions

• Does the package hypothesis block the regress, or only relocate it? Floridi’s closing “Russian dolls … all the way in” (p. 45) concedes an open regress at the epistemic level.
• Can ISR answer Newman’s problem? Floridi promises a future paper (fn. 3).
• Is ISR an extension of ESR rather than OSR (Beni 2016, as reported by Karpenko)? The reflective, LoA-relative status of the commitment to informational objects makes that reading available.

Corrections to the seeded skim

• The dossier says §3 “answers the … regress objection … by pointing to the internal relation of difference … his ‘Ur-relation’”. The full text withdraws that stronger claim. Floridi calls the Ur-relation argument dependent on “a classic fallacy, namely an unqualified use of universal quantification” (p. 24), and retreats to a weaker conclusion: relata and relations “co-exist as a package” (Eddington). The package hypothesis is “all that is needed”, and “It is not true that all relata are logically prior to relations. This much … suffices” (pp. 24–25). The Ur-relation survives only as the “modified Leibnizian principle” defended in reply 6.6.
• Precision: “OSR” throughout means non-eliminativist OSR (NOSR, p. 7). Floridi rejects eliminativist OSR as a “last resort” option: it resembles Hegelian absolute idealism, collapses toward strong realism and invites a regress of structures (pp. 6–7). He places ISR in van Fraassen’s “in-between structuralism” (fn. 4).
• The dossier’s “a world made up of interacting informational objects” is right but omits a consequence Floridi draws: ISR supports “at least an irreducible, fundamental dualism, if not a Leibnizian pluralism” (6.9, p. 42) and a Kantian knowable/unknowable dualism (6.10). The data are concrete, not abstract (6.9). That contradicts readings of ISR as an ontology of abstract objects, including Karpenko’s (LIT-215).
• The dossier’s section pages are slightly off. §2 runs pp. 7–22, §3 pp. 22–25, §4 pp. 25–31, §5 pp. 31–33, §6 pp. 34–43, §7 pp. 43–45.
• The dossier’s “information-theory tag … justified only in that broad sense” is confirmed and should be sharpened. The paper uses no Shannon or algorithmic measure. “Information” here is Floridi’s semantic/ontic data-as-difference notion, and “information theory” appears only in passing (p. 13). The information-theory tag, whose blurb is entropy, channels, coding and capacity, is at best tangential; a philosophy-of-information word does not exist in the vocabulary.
• metaphysics tag: justified. The work defends an ontological thesis, that reality is constituted by mind-independent structural and informational objects, not substantial or material ones, and it is about what exists. philosophy-of-science correctly leads, since the frame is the structural-realism debate over theory change and the NMA/PMIA.