I am AI

Introduction of a dynamically corrigible probabilistic context-selection architecture for personalised inference under partial observability

Status: Applied / Computational architecture (testable, unvalidated)
This page presents a proposed systems architecture (the GIA) addressing context selection under partial observability in AI-enabled healthcare. It is an applied extension in AI deployment of the Dot Theory programme, formulated as a computational proposal with clear empirical testability. The claims are structural and performance-oriented, and require validation through implementation and controlled evaluation.

Abstract

AI-enabled healthcare systems are approaching a structural bottleneck that is not driven by data scarcity, but by the inability to select and operationalise relevant context under finite computational constraints. Current architectures scale by increasing data volume and model capacity, yet treat contextual conditioning as implicit, static, or weakly specified. This leads to diminishing returns in predictive performance, reduced interpretability, and instability under distributional shift. This paper introduces a dynamically corrigible probabilistic context-selection architecture for personalised inference under partial observability, in which contextual metadata is treated as an explicit, optimisable component of the inference process. It is called the Generative Interpretative Architecture (GIA) and detailed information on its construction can be found here: https://www.dottheory.co.uk/paper/the-invention-of-truth

Introduction

This paper introduces the speculative functional advantages of a systems architecture addressing a constraint that is becoming dominant in AI-enabled healthcare: not the acquisition of data, but the selection of context under finite computational resources. This systems architecture is based in Dot theory, which is explained across this website.

Contemporary systems are increasingly data-rich yet context-poor at the point of inference. They process large volumes of information but lack principled mechanisms for determining which contextual features are relevant for a given individual at a given time. The result is diminishing returns in predictive performance, increasing brittleness, and reduced clinical trust. The architecture proposed in the paper, here, and across this website, reframes inference explicitly as a context-selection problem and introduces a dynamically corrigible probabilistic framework designed to maintain fidelity under partial observability.

The core premise is that predictive performance in healthcare depends less on the absolute quantity of data than on the adequacy of contextual conditioning applied to that data. Clinical and evaluatory states are not fully captured by diagnostic categories alone. They emerge from structured interactions between behavioural, physiological, environmental, and historical variables. Under finite compute, it is not feasible to process all available variables with equal weight. The system must therefore select, compress, and structure context in a way that preserves decision-relevant information while discarding noise. This is the bottleneck the GIA architecture addresses.

Structure and inference

The proposed system treats each individual decision-instance case (inference state) as a structured informational state Ψ = (ψ, μ), where ψ denotes observable data and μ denotes contextual metadata governing interpretation. Compared to current standard systems, μ is then either relatively under-specified or implicitly encoded in model parameters, implicitly leading to relative loss of interpretive transparency and reduced adaptability. In the Dot theory model, μ is systemically elevated to a first-class computational and explicit object. Inference is then defined not only as a mapping from ψ to a prediction, but as a joint process of selecting μ, and applying an interpretive transformation f such that J = f(ψ, μ). The quality of inference then depends on the adequacy of μ under resource constraints.

Context selection is formulated probabilistically. Given a large candidate set of contextual variables and unlike current systems in those circumstances, the system discussed across this website maintains a distribution over possible context subsets, weighted by their expected contribution to predictive fidelity and balanced by pragmatic ‘unknowables’ (here defined as contextually relevant factors that are unobserved, inaccessible, unmeasurable, or computationally intractable at the point of inference). This delivers a different analytical landscape, one which allows the system to prioritise context that is informative for the specific individual trajectory being evaluated.

Rather than attempting exhaustive inclusion, it therefore performs constrained optimisation over context space. This is analogous to relevance and attention mechanisms, but extended to structured, interpretable, and dynamically revisable context sets, rather than latent feature weights alone.

Dynamic functionality

A key feature of the Dot theory’s proposed ‘GIA’ architecture is dynamic corrigibility. Because contextual adequacy cannot be guaranteed a priori, the system must remain open to revision. Corrigibility is implemented through feedback loops that monitor divergence between predicted and observed outcomes, as well as disagreement across plausible context selections. When such divergence exceeds a threshold, the system updates its distribution over μ, effectively revising its understanding of which contextual factors are relevant. This prevents accumulation of systematic error and supports adaptation across changing clinical conditions.

Under partial observability, multiple context configurations may be consistent with the available data. The system therefore maintains epistemic plurality rather than collapsing prematurely to a single interpretation. Competing context hypotheses are evaluated in parallel within computational limits, and uncertainty is propagated into downstream decisions. This preserves epistemic divergence and supports safer decision-making in high-stakes environments. It also aligns with clinical and legal reasoning practices, where differential diagnosis is maintained until sufficient evidence accumulates.

Differentiation

The architecture introduces a distinction between representational state, structural constraints, and behavioural properties. Structural constraints ΛΞ define admissibility and compatibility conditions on context selection, ensuring that selected μ is coherent with domain knowledge, privacy constraints, and computational limits. Behavioural properties such as instability or sensitivity arise from the interaction between Ψ, ΛΞ, and the transformation f. This separation allows the system to diagnose failure modes. For example, apparent “chaotic” behaviour in predictions can be traced to mismatches between context selection and underlying data structure, rather than treated and discarded as noise.

Alongside this cost-efficient data transformation, the critical issue of privacy is addressed also through abstraction and decision making at the level of contextual patterns rather than individual records. The system operates on archetypal trajectories derived from historical data, comparing current cases to statistically similar patterns without exposing identifiable information. Context selection then operates over these abstractions, allowing personalised inference while preserving data protection requirements. This is particularly relevant in healthcare systems where regulatory constraints limit direct data sharing.

Adoption and validation

From a computational perspective, the architecture reduces effective dimensionality by selecting context adaptively rather than statically. Instead of training increasingly large models to absorb all possible variation, it allocates computational resources to the most relevant contextual dimensions for each case. This improves efficiency and scalability, especially in settings where compute is constrained relative to data volume. It also enables deployment in distributed or resource-limited environments.

Empirically, the central claim is testable. The hypothesis is that models incorporating explicit, dynamically selected contextual metadata μ will outperform models relying solely on ψ or on fixed context structures, across predictive accuracy, calibration, and robustness metrics. This can be evaluated through controlled studies comparing standard pipelines with context-selection architectures on longitudinal healthcare datasets. Performance improvements are expected particularly in heterogeneous populations and complex, multi-factorial conditions. This is one of the suggested extensions of Dot theory, the scientific programme discussed on this website.

Conclusion

The broader implication of adopting Dot theory is a shift in how AI systems are designed for healthcare. Rather than treating context as an implicit by-product of large-scale modelling, it becomes an explicit, controllable, and corrigible component of inference. This aligns system behaviour more closely with clinical reasoning, where context is actively constructed and revised. It also supports transparency, as the selected context μ can be inspected, audited, and contested within institutional frameworks.

In conclusion, the proposed architecture addresses a structural limitation in current AI-health systems by reframing inference as probabilistic context selection under partial observability. By introducing dynamic corrigibility, explicit contextual modelling, and constrained optimisation over context space, it offers a path to improved predictive fidelity without requiring unbounded increases in data or compute. Its success depends on empirical validation, but its formulation provides a coherent response to the emerging bottleneck in AI-enabled healthcare systems.

Thank you for your time and attention. For further investigation into this architecture please visit the homepage or this paper.

S.