The Information Asymmetry in Autonomous Operations
Autonomous systems operating in dynamic, high-stakes environments face a persistent source of uncertainty that is not geometric or perceptual but informational. Critical situational context exists in human minds and is communicated through verbal evidence, yet it does not reliably enter the agent's decision-making pipeline. The environment may be partially observable through sensors, but mission-relevant context — intent, constraints, hazards, last-known trajectories, local knowledge — remains unobservable to the controller unless the acquisition process that obtains it from a human source is itself made observable. Language is the evidence channel through which that process unfolds; the object the controller needs to read is not the language but the schema-grounded trajectory of acquired evidence the language induces.
The problem is therefore not that robots lack sensors, nor that the language they receive is poor. It is that autonomy lacks a reliable interface layer for reading the acquisition trajectory, grounding the evidence it carries, and admitting both into control under formally characterised conditions. Search and rescue (SAR) is the motivating domain throughout this thesis because it concentrates the difficulty: time pressure, dispersed witnesses, evolving hypotheses, and consequences that compound when acquisition fails. But SAR is the boundary case, not the boundary. The same architectural deficit recurs wherever an autonomous system must acquire and act on knowledge held by humans — medical intake, disaster coordination, investigative interviewing, supervisory teaming, collaborative grounding. Across these domains the surface activity differs and the operational interpretation of acquisition friction differs sharply, but the missing observability and admission interface is the same.
Across the broader landscape of autonomous systems, there is no compositional account of the boundary through which runtime natural language enters control — one that spans from an after-exchange observability layer over the acquisition state, through grounding of unstructured language into bounded mathematical signals, to the formal admission guarantees that preserve downstream controller assurance.
This gap sits at the intersection of human–robot interaction, language-conditioned control, runtime verification, and information-theoretic dialogue analysis. None of these fields, individually, provides the missing interface; each is positioned to consume it once it exists.
The Triple Barrier Problem
Bridging verbal human intelligence into robotic autonomy requires resolving three coupled barriers. They are not three flavours of the same problem; they are sequential conditions, each of which must be discharged before the next becomes meaningful.
In physical control, real-time observability is foundational: navigation does not act on raw images but on pose and map estimates; manipulation does not reason over raw motor currents but over force and contact estimates. Human-information acquisition has no comparable measurement layer. A controller may receive natural-language evidence from a source yet lack any compact, after-exchange estimate of which task categories have been resolved, where residual acquisition utility remains, and whether the recent exchange is contributing new evidence or revisiting already-covered ground. Without such a layer, adaptation is impossible: the controller cannot decide whether to continue, switch category, change strategy, terminate, or escalate. Crucially, this is an observability problem, not a quality-judgment problem. Recurrence is not intrinsically a failure — repeated probing may be costly in time-critical SAR, productive in collaborative grounding, and appropriate in reflective interviewing. The missing layer must expose the acquisition trajectory faithfully and leave the cost of that trajectory to the downstream task.
Even when useful information is provided, current pipelines lack robust mechanisms to convert messy declarative language into internal constraints that guide downstream autonomy. Existing language-conditioned approaches typically assume relatively well-formed inputs and treat language as an endpoint rather than a continuously updating source of contextual constraints. In emergency and situated HRI settings, interaction is characterised by miscommunication, corrections, ambiguity, and self-repair — phenomena that demand semantic reasoning rather than pattern matching, and that demand a middleware between the language process and the controller.
Formal assurance methods for autonomous systems — shielding, runtime verification, contract-based design, control barrier functions — are effective when controller-facing inputs are already well-characterised in bounded mathematical domains. Runtime natural language, however, arrives as unbounded, ambiguous, and potentially unreliable input. No existing assurance framework specifies the conditions under which language-derived content may be admitted into a control pipeline, nor how guarantees at that admission boundary compose with the downstream theories the system already relies on. For safety-critical applications, empirical demonstrations that "language helps" are not enough; a compositional account of the interface is required.
These three barriers are coupled. Without observability, grounding cannot be evaluated; without grounding, assurance has no bounded objects to constrain; without assurance, observability and grounding remain empirical conveniences rather than principled components of a safety-relevant pipeline. The thesis treats them as one structural problem decomposed into three independently validated layers.
The Thesis Statement
Autonomous systems can reduce information asymmetry in dynamic environments by combining (i) an after-exchange observability layer that exposes the human-information acquisition state — progress and friction — to a downstream controller; (ii) a grounding middleware that converts extracted declarative language into bounded, client-agnostic control signals; and (iii) a compositional assurance framework that certifies the admission boundary through which runtime language enters control pipelines.
The architectural commitment underlying the thesis is the separation of concerns. Rather than building a monolithic language-conditioned agent, the work decomposes the language-to-control problem into three layers whose interfaces are formally specified and whose guarantees compose. Each layer is validated independently; together they instantiate a single chain along which runtime natural language enters control through certified, bounded interfaces.
A scope boundary runs through the entire programme. The thesis makes no claims about regulating or measuring human cognitive states, nor about evaluating whether a dialogue is conversationally good. It instruments the observable acquisition trajectory, grounds extracted content into control-admissible signals, and provides the formal interface guarantees that connect these layers. The human side of the channel is treated with respect for its complexity: latent quantities such as effective communicative capacity are acknowledged in the formalism but are not estimated, attributed, or controlled. Whether acquisition friction should be penalised, ignored, or interpreted as productive recurrence is a decision left entirely to the downstream task and controller; the measurement layer is deliberately silent on this question.
The Three Pillars
The PhD is structured around three interlocking research pillars, each corresponding to one barrier and one research question. Each pillar produces an independently publishable contribution; together they define the language-to-control pipeline.
Acquisition-State Telemetry: The Sensing Layer
RQ1: What after-exchange observables can expose the schema-grounded acquisition state induced by human-information evidence, so that progress and friction become legible to a downstream controller across heterogeneous interaction settings?
What: Acquisition-State Telemetry (AST) — a controller-facing observability layer that treats verbal evidence as an input stream and exposes the evolving acquisition state induced by that evidence. The architecture separates three roles that the literature typically conflates: (i) state estimation from source evidence via a schema-grounded scorer, (ii) trajectory-level telemetry over the resulting acquisition state, and (iii) domain-specific interpretation of that telemetry inside a downstream controller. The interface exposes two observable types: a prospective Progress observable Pt that estimates where residual acquisition utility remains across the task schema, and a retrospective Friction observable Ft that estimates whether the recent acquisition trajectory is becoming low-yield or recurrent.
Why it is needed: Existing work concentrates on either pre-turn question selection (expected information gain) or post-hoc dialogue evaluation (breakdown classification, satisfaction scoring). Neither produces the after-exchange, controller-facing observable that adaptation requires. More fundamentally, existing measures treat the dialogue itself as the object of evaluation, conflating whether the exchange was conversationally good with whether the acquisition state advanced. The two are different questions, and only the second is what the controller needs.
How: All observables are conditioned on a Global Task Schema Ω that defines categories, importance weights, and completion criteria. The empirical instantiation uses one fixed lightweight sensor pair — PEB (a Bayesian residual-progress score) and SI⊥ (an embedding-subspace semantic-recurrence score under weak acquisition gain) — but the architectural contribution lies in the interface, not the mathematics: alternative posterior models, embeddings, or projection operators would preserve the same separation between source-evidence estimation, acquisition-state telemetry, and downstream interpretation. Crucially, AST is deliberately silent on whether friction should be penalised. "In SAR, repeated low-yield acquisition may carry operational cost, whereas in collaborative grounding or counselling, recurrence may be useful." The telemetry layer measures the acquisition trajectory; the controller interprets its cost. Validation spans controlled SAR traces and three external corpora — transactional slot filling (MultiWOZ), collaborative spatial grounding (HCRC Map Task), and reflective interviewing — using a single fixed calibration, demonstrating that the same observability interface produces interpretable readings across qualitatively different interaction mechanics.
LUCIFER: The Middleware Layer
RQ2: How can unstructured, real-time verbal reports be translated into bounded mathematical signals that heterogeneous autonomous controllers can consume without retraining?
What: LUCIFER (Language Understanding and Context-Infused Framework for Exploration and Refinement) — an inference-only middleware that exposes a client-agnostic Signal Contract with four standardised outputs: policy priors, reward potentials, admissible-option constraints, and action predictions for targeted information gathering.
Why it is needed: Embedding language understanding inside the learner or planner couples language conventions, domain knowledge, and optimisation dynamics — increasing redeployment burden and confounding grounding errors with control errors. A principled interface boundary is required between language processing and decision-making, one robust to messy, self-correcting reports that defeat pattern-matching baselines.
How: A retrieval-augmented Context Extractor maps verbal reports into structured semantic objects grounded against the task schema and a knowledge base. An Exploration Facilitator predicts high-value information-gathering actions from client-agnostic telemetry alone. Validation uses two structurally distinct downstream consumers — a hierarchical reinforcement learner and a hybrid A*/heuristic planner — establishing the architectural claim that the middleware is genuinely client-agnostic.
Compositional Assurance: The Admission Boundary
RQ3: Under what formal conditions can runtime natural language be admitted into an autonomous control pipeline such that the downstream controller's existing assurance guarantees are preserved?
What: A Compositional Assurance Framework that formalises the admission boundary through two explicit architectural interfaces. A Certification Gate admits language-derived content only when acquisition is sufficiently complete and not degraded — a per-category predicate conjoining a sufficiency condition on the AST progress observable with a health condition on the AST friction observable. A Signal Contract transforms certified content into bounded, well-typed controller-facing objects whose structural conditions match downstream controller assumptions.
Why it is needed: Once runtime language is allowed to influence control, language-side uncertainty and controller behaviour become entangled. Without an explicit admission boundary, failures originating in acquisition, interpretation, or translation are indistinguishable from control failures. A formal interface theory is required to make these failure modes separable and to allow language-driven autonomy to compose with existing assurance methods rather than override them.
How: Under explicit acquisition-side assumptions (cooperativeness, minimum gain, within-episode stationarity), the framework proves a Compositional Bounded-Input Assurance Theorem covering certified admission, bounded signal construction, policy invariance under potential-based shaping, one-step safety under option filtering, and graceful fallback when certification fails. The theorem is instantiated for two structurally different controller classes — tabular Q-learning and PID control — demonstrating that the same admission logic applies across discrete decision-making and continuous feedback control. The contribution is architectural rather than algorithmic: it does not verify the semantics of language but formalises the interface through which language-derived content becomes admissible input.
How the Pillars Compose
The three pillars are not three papers stapled together. They form a single chain in which each layer's output is the next layer's structured input, and the same task schema Ω threads through all three. The composition is what licenses the assurance argument; the layered architecture is what makes the composition modular.
Pillar 1 · Observability (AST)
After each acquisition exchange, a scorer estimates the per-category acquisition state from verbal evidence; AST exposes a prospective progress observable Pt and a retrospective friction observable Ft over the resulting trajectory. The controller receives the telemetry; the cost of friction is left to the task.
Pillar 2 · Grounding (LUCIFER)
The Context Extractor maps streaming verbal reports into structured semantic objects validated against Ω. The Signal Contract transduces these objects into bounded mathematical signals — priors, potentials, constraints, action predictions.
Pillar 3 · Assurance (Certification Gate + Signal Contract)
For each category, the Certification Gate releases content to the signal layer only when the AST progress observable has crossed the task-set sufficiency threshold and the AST friction observable remains below the task-set friction tolerance — both thresholds are configuration parameters, not universal verdicts on the trajectory. The Signal Contract then guarantees that all admitted controller inputs are bounded and well-typed by construction.
Downstream Controller (Q-learning, PID, planner, …)
Receives only certified, bounded signals. Existing controller-side assurance theory (policy invariance, safety filters, contract-based design) remains applicable without modification.
This layered structure mirrors mature control domains. Pillar 1 is the observability layer over the acquisition process, Pillar 2 is the grounding middleware over verbal content, and Pillar 3 is the admission and contract boundary. The contribution is not just each layer individually, but the demonstration that they compose into a chain in which language influences control only through interfaces whose guarantees are strong enough to support downstream assurance theory.
Component-level validation (AST observability across SAR, transactional, spatial, and reflective settings; LUCIFER grounding under messy reports; controller instantiations of the assurance theorem) uses controlled simulation environments and curated corpora. Integration-level validation requires an information source that reacts to how the agent acquires — the classic plant-vs-controller problem of any closed-loop study. A collaborative MSc project (Summer 2026) is constructing a reactive, persona-configurable LLM information source that exposes the same interface as the current static environment. The intellectual contributions of the PhD stand independently of the simulator's fidelity; the simulator determines how convincingly the integration can be demonstrated, not whether the architecture works. The same modular separation that governs the technical layers governs the research infrastructure itself.
The Research Outputs
Each pillar is realised through concrete publications and submissions. Several further outputs surround the three pillars: a foundational conference paper and a co-authored generalisation study that establish the SAR setting; an extension that lifts one of the telemetry signals into a broader phase-space framework; an adjacent design study of downstream controllers; an exploratory human-side instrumentation sketch; and the reactive plant model required for integration testing. Together they constitute the narrative arc of the thesis: independent contributions that, taken in sequence, motivate and resolve the Triple Barrier Problem.
Selective Exploration and Information Gathering in Search and Rescue Using Hierarchical Learning Guided by Natural Language Input
The foundational paper that established the architecture. It introduced the integration of LLMs with hierarchical reinforcement learning for SAR, demonstrating how verbal inputs from human stakeholders can be transformed into actionable RL insights via an Information Space, Context Extractor, Strategic Decision Engine, and Attention Space. Language-infused hierarchical agents were shown to outperform flat RL agents — especially in sparse-reward environments — and domain-knowledge infusion via RAG produced properly grounded outputs.
- Contribution to the arc: Established that language-guided hierarchical decision-making is viable and that attention-based policy shaping from verbal inputs improves both performance and safety. Introduced the Information Space formalism that recurs in all subsequent works as the structural carrier of schema-based information requirements.
Learning What Matters Now: A Dual-Critic Context-Aware RL Framework for Priority-Driven Information Gain (CA-MIQ)
An extension of the Information Space formalism to the case where operational priorities shift mid-mission. CA-MIQ pairs a standard extrinsic critic with an intrinsic critic that fuses state novelty, information-location awareness, and real-time priority alignment; a shift detector triggers transient exploration boosts and selective critic resets so the agent re-focuses after a priority revision. In a simulated SAR grid-world, it achieves near-complete recovery after single and multiple priority shifts where ε-greedy and ε-boosted baselines fail to adapt.
- Contribution to the arc: Treats priority ordering as a latent context variable over the same Information Space that threads through the rest of the programme, addressing a limitation the foundational paper left open — its assumption of a static priority ordering. It is positioned as a decorative extension rather than a pillar: it strengthens the SAR motivator and the information-gathering formalism without altering the observability, grounding, or assurance interfaces.
Generalising Rescue Operations in Disaster Scenarios Using Drones: A Lifelong Reinforcement Learning Approach
A co-authored study that combines lifelong reinforcement learning with eligibility traces and a shaping-reward heuristic derived from pre-training experience, improving generalisation of UAV navigation to novel post-disaster environments. Extensive simulation shows improved average return over baseline RL and robustness across environment complexities and varying numbers of trapped individuals.
- Contribution to the arc: External validation of the SAR motivator that frames the thesis. It is not a pillar contribution and does not touch the language-to-control interfaces; it is included to evidence the broader RL-for-SAR programme within which the observability, grounding, and assurance layers are situated.
A Signal Contract for Online Language Grounding and Discovery in Decision-Making (LUCIFER)
The full maturation of the grounding layer. Formalises online language grounding as training-decoupled middleware exposing a four-signal contract (priors, potentials, constraints, action predictions). Component benchmarks show that reasoning-based extraction remains robust on self-correcting reports (91–100% accuracy) where pattern-matching baselines collapse (20–36%). System-level dual-client ablations — hierarchical RL and hybrid A*/heuristic planner — establish a clear pattern: grounding improves safety, discovery improves efficiency, and only their combination achieves both.
- Contribution to the arc: Proved that language can be externalised from the decision-maker through a stable mathematical interface, enabling safe and efficient behaviour without embedding NLP inside the controller's optimisation loop. The dual-client validation demonstrates that the approach is architectural, not algorithm-specific. LUCIFER's Exploration Facilitator consumes the same trajectory traces that AST exposes, and LUCIFER's Signal Contract becomes the output boundary of the assurance framework in Pillar 3.
Acquisition-State Telemetry for Autonomous Information Gathering
Defines and validates the observability layer. Frames human-information acquisition as a state trajectory observed through a controller-facing telemetry interface, with an explicit architectural separation between three roles that the literature typically conflates: acquisition-state estimation from source evidence, trajectory-level telemetry over that state, and domain-specific interpretation inside a downstream controller. The fixed instantiation evaluated here pairs a prospective Bayesian progress observable (PEB) with a retrospective subspace-residual friction observable (SI⊥); the architectural novelty is the interface, not these particular forms. Validation under a single fixed calibration spans four qualitatively different acquisition regimes — controlled SAR elicitation, transactional slot filling (MultiWOZ), collaborative spatial grounding (HCRC Map Task), and reflective interviewing — and a SAR reinforcement-learning study in which the same telemetry layer drives a downstream controller when acquisition friction carries operational cost.
- Contribution to the arc: Supplies the observability layer on which the entire programme depends. PE and SI are the controller-facing observables that the Certification Gate consumes in Pillar 3. The deliberate refusal to treat friction as failure is what makes the layer transferable: the same interface produces interpretable readings under SAR mechanics where recurrence is costly and under collaborative-grounding mechanics where recurrence is part of how grounding is achieved. The downstream RL study confirms that the layer is not only a diagnostic but an actionable signal whenever the task assigns a cost to low-yield acquisition.
A Compositional Assurance Framework for Admitting Runtime Natural Language into Autonomous Control Pipelines
The capstone formal contribution. Introduces two explicit architectural interfaces — Certification Gate and Signal Contract — and proves a Compositional Bounded-Input Assurance Theorem with seven properties: acquisition convergence, degradation detection, bounded downstream inputs, policy invariance under potential-based shaping, one-step safety under option filtering, graceful degradation, and partial coverage. Instantiated for tabular Q-learning and PID control, showing that the same admission logic applies across structurally different controller classes.
- Contribution to the arc: Provides the formal interface theory that turns the empirical demonstrations of AST and LUCIFER into a compositional assurance chain. The contribution is intentionally narrow: it does not verify the semantics or truth of runtime language. It formalises the admission boundary through which language-derived content becomes controller-facing input, and shows how guarantees at that boundary compose with existing downstream controller theory.
Phase-Space Telemetry for Temporal Representation Trajectories
An in-preparation generalisation that lifts the friction observable from human-information acquisition into a broader telemetry framework for temporal latent trajectories in learning agents. The framework distinguishes failure modes that conventional novelty-only, reward-only, or recurrence-only signals tend to conflate, and shows that the resulting telemetry is actionable rather than merely diagnostic.
- Contribution to the arc: The clearest demonstration of the programme's portability claim — an observability primitive designed for human-information acquisition transfers, with its interpretive neutrality intact, into representation-level diagnosis of learning agents. The telemetry classifies the trajectory; the downstream task decides which regimes are operationally undesirable. (Mechanism and experimental detail withheld pending publication.)
Behavioural Control for Information-Seeking Dialogue: A Systematic Design Space
An adjacent design study addressing which controller, once a category has been admitted, should produce the next questioning behaviour. The study frames information-seeking dialogue as a constrained control problem with acquisition-state observables as inputs and protocol constraints as bounds, and surveys the space of candidate downstream controllers and the trade-offs each presents between interpretability, sample efficiency, fine-grained control, and stability.
- Contribution to the arc: The assurance theorem is controller-class agnostic by construction; this study is its open companion, mapping the controllers that could sit beneath the certification boundary. Empirical comparison across this design space is identified as a downstream research thread rather than a thesis chapter.
ITLI: An LLM-Derived Intrinsic Task-Load Index for Cognitive Load Estimation
An exploratory, early-stage study on the human side of the loop, included here for completeness rather than as a load-bearing part of the thesis arc. It sketches a way to estimate intrinsic cognitive load from textual information streams using an LLM as a feature extractor over cues grounded in Cognitive Load Theory, mapped to a scalar task-load index. The approach is open and pending validation.
- Relationship to the arc: ITLI sits deliberately outside the externalisation principle that governs the three pillars: it estimates a cognitive-load-adjacent quantity rather than instrumenting an observable acquisition state, and its outputs are not bounded controller inputs in the formal sense, so it does not enter the assurance chain. For that reason it is treated as a future direction rather than a thesis contribution — a possible observability signal for variable-autonomy and adaptive presentation, pursued only if and when it matures. It is mentioned because it marks the boundary at which the present engineering-side framing would meet work requiring quantitative psycholinguistic claims.
High-Fidelity LLM-Based Interviewee Simulation for Task-Oriented Information Gathering
A collaborative MSc project building a reactive, persona-configurable LLM interviewee simulator whose behavioural characteristics are grounded in distributional statistics from real task-oriented human communication. The simulator exposes the same programmatic interface as the existing static environment, so the research code operates without modification, and its behavioural fidelity is evaluated against literature-derived baselines.
- Contribution to the arc: In any closed-loop study, the researcher implicitly builds both the controller and the plant. This project externalises the plant. The PhD's intellectual contributions stand independently of the simulator's fidelity, but a credible plant model is what makes adaptive integration-level evaluation rigorous rather than merely illustrative. The persona configuration is treated as an initial condition rather than a state variable; modelling persona dynamics in response to interviewer behaviour is identified as a frontier requiring validated computational models of interviewer–interviewee co-regulation, and is explicitly out of scope.
Transfer Beyond Search and Rescue
SAR is the testbed because it concentrates the problem; it is not the boundary of the contribution. Each pillar addresses a general pattern that recurs wherever autonomous systems interact with language-rich, human-populated environments. The table below makes the abstraction explicit.
| Contribution | SAR Instance | General Pattern | Transfer Domains |
|---|---|---|---|
| Acquisition-State Telemetry (PE + SI) | Exposing the SAR acquisition state — progress per category, friction over recent exchanges — to the controller | Controller-facing observability over any schema-grounded acquisition process, deliberately neutral about the operational cost of friction | Transactional dialogue, collaborative spatial grounding, reflective interviewing, medical intake, investigative interviewing, supervisory teaming, intelligent tutoring |
| Signal Contract (LUCIFER) | Converting hazard/victim reports into navigational priors, potentials, and admissible-option constraints | Training-decoupled middleware for converting streaming language into client-agnostic control-relevant signals | Warehouse logistics, agricultural robotics, autonomous vehicles receiving runtime instructions, language-conditioned planners |
| Compositional Assurance | Certifying admission of witness-derived hazard information into a SAR controller | Interface-level assurance for runtime language entering any pre-assured controller class | Any safety-critical system admitting natural-language input — flight management auxiliaries, medical decision support, supervisory autonomy |
| Phase-Space Telemetry | Generalising the friction observable into a richer diagnostic over trajectory geometry | Trajectory-level telemetry that distinguishes failure modes conventional signals conflate — with the task deciding which are operationally undesirable | Sparse-reward RL diagnostics, sequence-model failure analysis, exploration-shaping in any agent with a latent representation |
| ITLI | Estimating intrinsic load of incoming mission reports for operators | Non-intrusive bandwidth estimation from textual information streams using an LLM as a feature extractor over CLT-grounded cues | Adaptive UIs, variable autonomy, instructional design, accessibility systems |
The thesis's deepest commitment is the recognition that admitting human knowledge into autonomous control requires the same engineering discipline as physical control: observability, grounding, and a certified admission interface — each modular, each composable, each independently validated. The measurement layer reports the acquisition trajectory; it does not adjudicate it. The grounding middleware translates language into bounded mathematical objects; it does not embed itself inside the controller. The assurance framework formalises the interface; it does not verify the semantics of language. This architectural separation is what transfers wholesale to any domain in which agents must observe, admit, and act on human-communicated knowledge under operational constraints.
What Did Not Exist Before
To be precise about the new knowledge produced by this PhD:
1. Acquisition-State Telemetry as a controller-facing observability interface. AST is the first model-agnostic, after-exchange measurement contract that separates source-evidence state estimation, trajectory-level telemetry, and downstream control interpretation — exposing a prospective progress observable and a retrospective friction observable without committing to whether friction is operationally bad.
2. Training-decoupled language grounding via a client-agnostic Signal Contract. LUCIFER is the first middleware to formalise a stable mathematical interface between streaming language and heterogeneous decision-makers, validated across learning and non-learning paradigms.
3. Compositional assurance for the language-to-control admission boundary. The Certification Gate and Signal Contract, together with the Compositional Bounded-Input Assurance Theorem, are the first interface-level formal account of how runtime natural language may enter an assured control pipeline without collapsing downstream assurance into the semantics of language.
4. Generalisation of trajectory telemetry to learning agents. A generalisation of the friction observable into a richer trajectory-level diagnostic that distinguishes failure modes conventional novelty- or reward-based signals conflate — preserving the commitment that the telemetry classifies and the controller adjudicates. (In preparation; detail withheld pending publication.)
Beyond the four contributions above, the programme includes exploratory work that is deliberately not claimed as a thesis contribution. ITLI, an LLM-derived intrinsic cognitive-load index, sits outside the externalisation principle — it estimates a cognitive-load-adjacent quantity rather than instrumenting an observable acquisition state — and is therefore treated as a future direction, pursued only if it matures. A behavioural-control design-space study surveys the downstream controllers that could sit beneath the certification gate and is identified as a research thread rather than a thesis chapter.
The thesis does not claim to have solved "the physics of dialogue", to have produced a universally optimal interviewing agent, or to have verified the semantics of runtime natural language. It claims to have built practical, validated tools — telemetry signals, a grounding middleware, and a compositional assurance framework — that are useful, computable, principled, and transferable. Three scope boundaries are drawn explicitly and observed throughout:
No claim about human cognitive or affective state. AST exposes the controller's view of the acquisition trajectory; it does not estimate, attribute, or control the latent communicative or cognitive state of the human source. The formalisation includes a latent quantity C(t) — the effective transmission capacity of the human–agent channel, which the literature suggests can be affected by interactional conditions — but C(t) is acknowledged in the model and is not measured, estimated, or controlled by any component. Psycholinguistic findings motivate the architecture; they are not claims the thesis sets out to prove.
No claim of dialogue quality. The thesis proposes no conversational-quality metric. Friction is not failure: AST classifies the trajectory, the Certification Gate adjudicates it against task-set thresholds, and the downstream controller decides whether a pattern is operationally costly. The neutrality of the measurement layer is a feature, not a hedge.
No claim of language verification. The assurance framework does not verify that runtime language is true. It formalises the interface conditions under which language-derived content may be admitted into a control pipeline. Truth, intent, and adversarial reporting are out of scope; the framework is about admission, not authentication. Whether the same observable trajectory patterns also reflect anything about the human's internal state lies at the intersection of computational psycholinguistics and control theory, beyond the scope of an engineering thesis.
The Research Philosophy
This PhD is driven by a conviction that autonomous systems must become active participants in the knowledge-acquisition process, not passive recipients of pre-processed data. Humans, when uncertain, ask questions, interview witnesses, and consult experts. Robots should do the same — and the channel through which they do so should be exposed to control through interfaces whose properties are formally characterised, not merely empirically demonstrated.
The path was not linear. The early inspiration came from an analogy between dialogue dynamics and fluid flow — the intuition that conversations have gradients toward information and circulation when stuck. Formalising that analogy in full would have required mathematical machinery beyond the scope of a single PhD, and would also have committed the thesis to a strong claim about dialogue itself rather than about the acquisition state induced by dialogue. The work therefore pivoted in two related ways. First, to computational proxies — PEB, SI⊥, and their phase-space generalisation — that capture the functional intent of those metaphors using information theory and embedding geometry. Second, and more deeply, to a reframing of the object itself: from dialogue quality to acquisition state. The observable is no longer "is this conversation going well?" but "where is the controller in its task-relevant information state, and how is that state evolving?". This is the difference between judging the exchange and instrumenting it. The former is normative and domain-specific; the latter is operational and transferable.
A related pivot shaped the thesis's relationship with psychology. The literature on stress, recall, and crisis communication motivates the engineering architecture but is not what the thesis sets out to prove. AST observes the acquisition trajectory — what evidence has accumulated, what residual utility remains, whether recent exchanges are extending or recurring on existing content — not the human's latent cognitive or emotional state. It regulates nothing about the human and adjudicates nothing about the dialogue. It exposes the trajectory; the downstream controller decides what to do with it. This is a deliberate engineering stance: treat the human side of the channel with respect for its complexity, instrument what is observable, and control what is controllable — namely, the agent's own actions and the interfaces it exposes.
The tools work. The deeper mathematics remains for the future. And that is a mature, honest research contribution.
Each component was designed with a single principle in mind: build frameworks others can extend, not just deliverables for a contracted project. The AST interface is sensor-pair agnostic and source-channel agnostic — the empirical sensor pair is one instantiation; the same interface produces interpretable readings across SAR, transactional dialogue, spatial grounding, and reflective interviewing under a single calibration. The Signal Contract is client-agnostic. The Assurance Theorem is controller-class agnostic. The phase-space telemetry is application-agnostic. This modularity is intentional. It is what makes the work transferable beyond SAR, beyond any single application domain, and beyond the lifetime of this PhD.
The same philosophy governs the research infrastructure. In a multidisciplinary closed-loop PhD, the researcher implicitly builds both the controller and the plant. Recognising this, the evaluation infrastructure — a reactive, behaviourally grounded information-source simulator — has been scoped as a self-contained collaborative MSc project: the contributor builds the plant model without needing to understand the assurance theory, while the PhD researcher retains sole ownership of the intellectual contributions. This separation mirrors the architectural commitment that governs the technical layers themselves: every interface in this programme is designed so that what sits on either side of it can be built, validated, and replaced independently.
The Research Arc in One Sentence
This PhD develops the observability, grounding, and assurance layers through which an autonomous controller reads the schema-grounded acquisition trajectory induced in a human source — and admits the resulting content into control through certified, bounded interfaces — treating language as the evidence channel and the acquisition state as the object the controller reads, with the same engineering discipline applied to physical sensing, perception, and control.