---

NON-BIOLOGICAL SEMANTICS (NBS)

A proposed field-level research programme for the study of semantic organisation in artificial systems

Author: Tristan Radford — Cross-Domain Cognitive Systems Architect

Publisher: Riff Systems Ltd · Company No. 17419314

Version: 1.0  ·  Current revision: 7 September 2026

Status: Proposed field-level research programme

Canonical publication: riffsystems.org/research/non-biological-semantics/

Abstract

Artificial neural systems form high-dimensional internal representations that can exhibit stable geometric, statistical and task-relevant structure. Contemporary work in representation learning, computational linguistics and interpretability has shown that linguistic and semantic distinctions can sometimes be decoded from these representations, that representational geometry can vary systematically across layers and tasks, and that internal structure does not automatically imply causal use by the model. The scientific problem is therefore not simply whether artificial systems contain “meaning,” but how to distinguish measurable representational organisation, functional contribution to behaviour, dynamic change during processing, and the human interpretations placed on those measurements.^{[1][2][3][4][5][6]}

This paper proposes Non-Biological Semantics (NBS) as a field-level research programme for that problem. NBS is not presented as an already-established field, nor as a replacement for mechanistic interpretability, computational semantics, representation learning, information theory or cognitive science. Its present status is a testable and falsifiable research programme with discipline-level scope and field-level ambition: if the programme succeeds empirically, NBS aims to establish a future scientific discipline for the systematic study of artificial representational geometry, dynamics and human interpretability. It proposes a common architecture for relating three levels of analysis: a Representational Geometry Layer (RGL) for structural organisation, a Representational Dynamics Model (ReDyM) for changes in representation over time or computation, and an NBS Interpretive Interface (NBS-II) for translating validated measurements into human-interpretable descriptions.

NBS uses semantics operationally. It does not assume consciousness, subjective meaning, intention or human-like understanding. A semantic claim is warranted only to the extent that a representational pattern can be related to defined distinctions in inputs, tasks, outputs or behaviour and survives appropriate controls. Stronger claims require stronger evidence: decodability is not equivalent to causal use, geometric separability is not equivalent to conceptual understanding, and an interpretable visualization is not itself an explanation of mechanism.

The immediate scientific task of NBS is methodological: to develop testable ways of mapping artificial representational structure, measuring its dynamics, evaluating whether identified structure is functionally relevant, and constructing interfaces that expose those findings without overstating what they establish. Its broader ambition is field-building: if this programme produces a coherent body of reproducible methods, findings and unresolved questions that is not adequately organised by existing disciplinary boundaries, NBS aims to provide the basis for a future scientific field devoted to those problems.

1. Introduction: the problem of artificial semantic organisation

Modern artificial systems transform inputs through distributed, high-dimensional internal states. In language models, these states can contain information about lexical, syntactic, semantic and task-related properties, but the presence of decodable information does not by itself show how the model uses that information, whether the representation is stable across contexts, or whether a human description of it corresponds to the model’s operative mechanism.^{[5][6][7][8][9]}

The central problem addressed by NBS is therefore:

How can the structure, transformation and functional relevance of artificial representations be measured and translated into human-interpretable form without confusing representation with consciousness, decodability with causality, or visualization with explanation?

This question is narrower and more defensible than the claim that artificial systems possess an inaccessible form of “meaning” in the human or phenomenological sense. NBS treats artificial semantics as an empirical target: structured relations among internal representations, task-relevant distinctions and observable behaviour.

1.1 Working definition

Non-Biological Semantics is a proposed field-level research programme for studying how artificial systems organise, transform and use representational structure that is systematically related to distinctions in inputs, tasks, outputs or behaviour.

The term non-biological specifies the substrate under investigation. It does not imply that biological and artificial representation are fundamentally incomparable, nor that artificial systems possess the same kinds of cognition or meaning as humans.

1.2 Four evidential levels

NBS distinguishes four claims that are often blurred together:

1. Sensitivity or association — a model state varies systematically with a defined feature.
2. Decodability or structure — information about the feature can be recovered from the representation using a specified analysis.
3. Functional relevance — intervention on the identified representation changes model behaviour in a manner predicted by the hypothesis.
4. Human interpretation — a description, projection or visualization makes the finding understandable to a person.

These levels are related but not interchangeable. A successful probe may reveal decodable information without showing that the model uses it. A causal intervention may establish functional relevance without showing that the human label attached to the intervention exhaustively describes the underlying computation. An interface may be useful while remaining only a partial model of the system.

2. Position relative to existing fields

NBS does not begin from the premise that existing fields cannot study artificial meaning. Several established areas already address essential parts of the problem.^[5][7][10]

Representation learning studies how useful structure is encoded in learned spaces.
Computational linguistics and computational semantics analyse linguistic and semantic structure in artificial models.
Interpretability and mechanistic interpretability investigate internal features, components, circuits and causal mechanisms.
Information-theoretic probing and diagnostic methods test what information can be recovered from representations.
Cognitive and systems sciences provide mature concepts for representation, dynamics, control, stability and interpretation.

The proposed contribution of NBS is integrative rather than exclusionary. It asks whether structural, dynamic and interpretive analyses can be linked within a common protocol while keeping their evidential status explicit. Its field-level claim is prospective rather than declarative: NBS proposes that these connected problems may justify a dedicated scientific discipline if the programme demonstrates sufficient empirical coherence, methodological distinctiveness and cumulative value.

This positioning creates a clear novelty criterion: NBS should be judged not by whether geometry, dynamics, probing or interpretability are individually new, but by whether the combined architecture improves measurement, cross-method consistency, causal clarity and human interpretation relative to existing approaches. Discipline-level status would therefore have to be earned through a sustained body of work, not inferred from the breadth of the proposal alone.

3. The three-layer NBS architecture

NBS organises the problem into three analytically distinct layers.

The three layers are unified because none is sufficient on its own. Structure without dynamics is incomplete: a static geometry cannot show how representations transform, stabilise, diverge or switch during computation. Dynamics without structure is uninterpretable: trajectories have little explanatory value unless the representational space and the relations being traversed are defined. Interpretation without evidential discipline is misleading: a human-readable interface can create apparent understanding even when the underlying structural or causal claim is weak. NBS therefore treats geometry, dynamics and interface as mutually constraining levels of one research programme rather than as independent topics.

3.1 Representational Geometry Layer (RGL) — structure

The RGL describes the measurable organisation of model states in representational space.

For a representation vector $h$ in a layer or component of a model, relevant questions include:

• Which distinctions are linearly or non-linearly separable?
• Which distances or similarity relations remain stable across paraphrase, context or model scale?
• Are task-relevant features concentrated in identifiable subspaces or distributed across many dimensions?
• Does geometry change across layers, time steps, modalities or tasks?
• Are apparent clusters, directions or manifolds robust to the analysis method and control conditions?

Possible measurements include representational similarity, dimensionality, local and global distance structure, subspace alignment, probe performance, information-theoretic quantities, cluster stability and geometry under controlled perturbation.^{[1][2][3][4][5][7][8]}

The RGL does not assume that a geometric feature is a human concept merely because a human label predicts it. Geometry establishes structured organisation; semantic interpretation requires additional evidence.

3.2 Representational Dynamics Model (ReDyM) — change

The ReDyM concerns how representations change during computation, across context, under perturbation or over learning.

A generic form is:

$$h_{t+1} = T(h_t, c_t, \theta)$$

where $h_t$ is a current representational state, $c_t$ represents relevant context or input conditions, $\theta$ denotes the model parameters or other fixed system properties, and $T$ is the transformation that produces the subsequent state.

The empirical target is not a metaphorical “force” acting on meaning. It is a measurable trajectory or transformation. Candidate phenomena include:

• representational drift — systematic change in encoding across time, context or learning;
• representational conflict — incompatible or competing task-relevant structures identified under controlled conditions;
• convergence or stabilization — reduction in representational variability toward a reproducible state or subspace;
• trajectory switching — transitions between distinguishable representational regimes;
• coupling — dependencies between representations across layers, components or modalities.

Terms such as attractor, equilibrium or collapse should be used only when the relevant mathematical or dynamical criteria are actually demonstrated. NBS therefore replaces the earlier language of “semantic force,” “drift pressure,” and “paradox collapse” with operational descriptions unless stronger evidence warrants the dynamical terminology.

3.3 NBS Interpretive Interface (NBS-II) — human interpretation

The NBS-II is the human-facing layer. It translates validated structural or dynamic measurements into forms that can be inspected, compared and tested by people.

An NBS-II might contain:

• low-dimensional projections with explicit distortion warnings;
• labelled directions or subspaces whose validation procedure is shown;
• trajectory views across prompts, tasks or model layers;
• uncertainty and confidence indicators;
• causal intervention results alongside correlational measurements;
• comparative views across models or checkpoints;
• provenance showing which analysis generated each claim.

The NBS-II is not assumed to reveal a model’s “true thoughts.” Its purpose is to provide a disciplined interface to measurements. A valid NBS-II should preserve the distinction between what was observed, what was inferred, what was causally tested and what remains interpretive.

4. From “meaning” to operational semantic claims

The strongest scientific risk in NBS is semantic inflation: treating any organised internal state as evidence of meaning in a rich philosophical sense.

NBS therefore uses a graded operational vocabulary.

4.1 Representational structure

A pattern in model states that is statistically stable or recoverable under a declared analysis.

4.2 Semantic association

A representational structure that covaries with a defined semantic distinction, such as topic, entailment relation, semantic role, referential relation or task label.

4.3 Functional semantic relevance

A representational structure whose controlled alteration produces a predicted change in behaviour while appropriate controls rule out simpler explanations.

4.4 Human-interpretable semantic model

A human-readable description of representational structure whose fidelity can be evaluated against held-out observations and, where possible, causal interventions.

This hierarchy prevents the word meaning from doing evidential work that has not been earned.

5. Core research questions

The NBS programme can be organised around six testable questions.

RQ1 — Structural organisation

Do artificial systems contain reproducible representational geometries associated with defined semantic distinctions?

RQ2 — Generalisation

Do those structures persist across paraphrases, datasets, model scales, architectures, checkpoints and task contexts, or are they local artefacts of a particular probe or dataset?

RQ3 — Functional relevance

Are the identified structures causally involved in model behaviour, or are they merely decodable correlates?

RQ4 — Dynamics

Do controlled semantic or task manipulations produce reproducible representational trajectories, transitions or convergence patterns?

RQ5 — Cross-level correspondence

Can geometric and dynamic measurements predict observable output behaviour better than simpler behavioural baselines?

RQ6 — Interpretive interface quality

Can human-facing interfaces communicate validated internal structure in ways that improve prediction, diagnosis or intervention without increasing false confidence about model internals?

6. A minimum empirical protocol for NBS claims

A scientific NBS study should specify the claim before choosing the visualization that best supports it.

6.1 Declare the target

Define the semantic distinction, behaviour or computational property under investigation.

6.2 Freeze the representation source

Specify model, checkpoint, layer or component, prompt protocol, token-selection rule and any preprocessing before decisive analysis.

6.3 Establish a structural baseline

Measure whether the target is recoverable or geometrically distinguishable and compare against appropriate random, lexical, frequency, positional or task-specific controls.

6.4 Test robustness

Repeat under paraphrase, held-out data, alternative seeds, alternative probes and relevant distribution shifts.

6.5 Separate decodability from use

Where a functional claim is made, perform an intervention or other causal test. Probe accuracy alone is insufficient.^{[5][6][7][8][9]}

6.6 Analyse dynamics only when temporal or computational order matters

If trajectories are claimed, define the state sequence and distance or transition metric before interpreting it.

6.7 Validate the interface independently

If an NBS-II visualization or description is presented, test whether it improves human prediction, comparison or diagnosis on held-out cases. Visual plausibility is not sufficient.

6.8 Preserve negative results

Failure to recover stable geometry, causal effects or cross-model generalisation is an informative result and should constrain the NBS claim rather than trigger post-hoc redefinition.

7. Candidate empirical programme

7.1 Geometry study

Select a small number of semantic relations with externally defined labels. Compare representational geometry across multiple open-weight language models and layers. Test whether identified directions or subspaces replicate across paraphrase and held-out domains.

Pass condition: a structure replicates under predeclared similarity or decoding criteria and survives control features.
Failure condition: the effect disappears under paraphrase, control variables or probe changes.

7.2 Functional intervention study

For a replicated structure, intervene on the associated activations while using matched random-direction and magnitude controls.

Pass condition: intervention changes the predicted semantic or behavioural outcome more than matched controls and without indiscriminate degradation.
Failure condition: intervention has no specific effect or produces broad nonspecific disruption.

7.3 Representational dynamics study

Construct controlled ambiguity, contradiction or competing-constraint tasks. Measure trajectories across layers or generation steps rather than inferring “tension” from output language.

Pass condition: predefined trajectory features distinguish conditions on held-out examples and replicate across model instances.
Failure condition: apparent dynamics are not distinguishable from ordinary activation variability or prompt-specific artefacts.

7.4 Interpretive interface study

Build an NBS-II prototype displaying only measurements that have passed structural and, where relevant, causal validation. Compare expert performance with and without the interface.

Pass condition: the interface improves calibrated prediction or diagnosis while preserving uncertainty.
Failure condition: it increases confidence without improving accuracy, or encourages unsupported mental-state attribution.

8. Falsifiability and failure conditions

NBS should be weakened, revised or rejected where its distinctive claims fail empirical tests.

The programme would be substantially undermined if:

• purported semantic geometries do not replicate beyond individual datasets or probes;
• identified structures are dominated by superficial lexical or positional confounds;
• decodable features repeatedly fail causal relevance tests;
• proposed representational dynamics cannot be distinguished from generic activation change;
• cross-model comparisons require incompatible ad hoc definitions;
• human-facing interfaces systematically overstate mechanism or reduce calibration;
• the three-layer architecture provides no explanatory or methodological advantage over existing interpretability workflows.

These are not peripheral limitations. They are explicit conditions under which NBS would fail to justify itself as a distinct field-level research programme or to support its longer-term claim to disciplinary status.

9. Applications if the programme succeeds

If the proposed architecture proves empirically useful, NBS could support several areas.

9.1 Interpretability

Relating structural measurements to causal tests and clearly separating them in human-facing tools.

9.2 Model comparison

Comparing representational organisation across architectures, checkpoints, languages or modalities using common measurement protocols.

9.3 Semantic diagnostics

Detecting instability, shortcut representations, brittle separability or changes in task-relevant structure under distribution shift.

9.4 Human–AI interaction

Building interfaces that help users understand what internal measurements do and do not establish, reducing anthropomorphic overinterpretation.

9.5 Safety and alignment research

Testing whether safety-relevant distinctions are merely decodable, reliably functionally represented, or manipulable without broad model degradation.

9.6 Cognitive comparison

Providing carefully bounded points of comparison between biological and artificial representational systems without presuming equivalence of mechanism or experience.

10. Research agenda

The immediate NBS research agenda is field-building and methodological rather than declarative. The programme is intended to test whether a sufficiently coherent empirical and conceptual territory exists to justify a future discipline, rather than assuming that status in advance.

Phase 1 — Measurement discipline

• define operational semantic targets;
• establish geometry and probing baselines;
• specify control conditions;
• distinguish association, decodability and causal use;
• establish cross-model replication procedures.

Phase 2 — Dynamics

• define trajectory metrics;
• test controlled conflict and ambiguity conditions;
• distinguish drift, switching and convergence from generic activation variability;
• assess temporal and layer-wise reproducibility.

Phase 3 — Interface

• build NBS-II prototypes;
• display provenance and uncertainty;
• test human calibration and prediction;
• measure whether interface explanations correspond to causal evidence.

Phase 4 — Cross-domain extension

• test whether the architecture transfers beyond text-only language models;
• compare multimodal, agentic and other artificial systems;
• identify which constructs are architecture-specific and which generalise.

11. Non-claims

NBS does not presently claim that:

• artificial systems are conscious or sentient;
• internal vectors possess subjective meaning;
• geometrically separable states are equivalent to human concepts;
• a probe reveals the mechanism used by a model;
• a low-dimensional projection faithfully preserves all relevant high-dimensional structure;
• attractors, equilibria or dynamical “forces” exist unless demonstrated under appropriate criteria;
• the proposed NBS architecture is already empirically validated;
• NBS already constitutes an established scientific field or discipline;
• existing fields are unable to study these problems.

These restrictions are part of the framework, not disclaimers added after the fact.

12. Conclusion: a field-level proposal, not a declaration by fiat

Artificial systems contain high-dimensional internal representations that can be studied structurally, dynamically and causally. Existing research already demonstrates that useful linguistic and semantic information can sometimes be recovered from representational geometry, while also showing that decodability, causal use and human interpretability must be carefully distinguished.

Non-Biological Semantics is proposed as a field-level, testable research programme for connecting those levels of analysis. Its scope is intentionally discipline-sized, but its present status is not that of an established discipline. The programme is designed to determine whether a coherent, cumulative and empirically distinctive body of work can be built around these problems. Its central architecture is deliberately simple:

Representational Geometry Layer (RGL) — structure
Representational Dynamics Model (ReDyM) — change
NBS Interpretive Interface (NBS-II) — human interpretation

A compact heuristic is:

Structure as geometry; change as dynamics; interpretation as interface.

The value of NBS will not be established by naming a field into existence. It will depend on whether this architecture supports reproducible measurements, stronger causal tests, clearer distinctions between evidence and interpretation, and interfaces that improve understanding without manufacturing certainty. The field-level ambition is therefore conditional on the performance of the research programme.

If those criteria are met across a sustained body of research, NBS aims to support the establishment and eventual recognition of a distinct scientific discipline for studying artificial representational geometry, dynamics and human interpretability. Until then, it should be understood as having discipline-level scope without discipline-level status. If the programme does not establish sufficient empirical or methodological distinctiveness, it should contract, merge with existing approaches, or be abandoned. That falsifiability is a requirement of the field proposal rather than a threat to it.

References