TRACE: A Metrologically-Grounded Engineering Framework for Trustworthy Agentic AI Systems in Operationally Critical Domains
arXiv:2605.03838v1 [cs.CL] 5 May 2026
Serhii Zabolotnii
†
Cherkasy State Business College, Cherkasy 18028, Ukraine State Scientific Research Institute of Armament and Military Equipment Testing and Certification, Cherkasy, Ukraine healthPrecision, New York, NY, USA
Abstract We introduce TRACE, a cross-domain engineering framework for trustworthy agentic AI in operationally critical domains. TRACE combines a four-layer reference architecture with an explicit classical-ML vs. LLM-validator split (L2a/L2b), a stateful orchestrationand-escalation policy (L3), and bounded human supervision (L4); a metrologically grounded trust-metric suite mapped to GUM/VIM/ISO 17025; and a Model-Parsimony principle quantified by the Computational Parsimony Ratio (CPR). Three instantiations—clinical decision support, industrial multi-domain operations, and a judicial AI assistant—transfer the same architecture and metrics across principally different governance contexts. The L2a/L2b separation makes the use of large language models a deliberate design decision rather than an architectural default, with parsimony quantified through CPR. TRACE introduces CPR as a first-class design principle in trustworthy-AI engineering.
Keywords: Trustworthy AI • Agentic AI • Large language models • LLM validators • LLMorchestrated oversight • Metrological traceability • Tiered oversight • Model parsimony • Operationally critical systems • Human–AI supervision Companion site: Documentation, the up-to-date publication roster, and CRediT-style author contributions are available at https://traces.solutions.
1
Introduction
Trust in agentic AI is not an intrinsic property of the model. A high-benchmark model can behave dangerously in real-world deployment; a modestly accurate model can operate safely within a correctly engineered architecture. As agentic AI systems enter operationally critical domains— clinical decision support, industrial multi-domain operations, judicial assistance, energy and infrastructure management—this observation becomes an engineering problem rather than a philosophical one: how should the architecture be designed so that trust is a measurable property of the system, not merely a declaration? The 2024–2026 literature addresses this question through four parallel but largely disconnected streams. Stream 1 – Clinical trust frameworks. FUTURE-AI [1] provides an international sixprinciple consensus for trustworthy clinical AI (fairness, universality, traceability, usability, robustness, explainability). TRIAD [2] formulates a three-component governance/value/deployment †
Corresponding author. Email: [email protected].
1
model for human–AI collaboration in medicine. Both are domain-specific and lack metrological grounding or cross-domain applicability. Stream 2 – Hierarchical agentic oversight. Tiered Agentic Oversight (TAO) by Kim et al. [3] formalises tiered agentic escalation with an orchestrator and clinician in the loop, empirically demonstrating absorption of 16.9–24.3% of individual-agent errors (Table 6, [3]). TAO is the closest architectural predecessor of TRACE and an empirical anchor for cross-domain generalisation. A complementary, more recent line of work—Eywa by Li et al. [4]— demonstrates heterogeneous foundation-model / LLM collaboration through an explicit FM–LLM communication protocol (Tsaheylu), with the EywaBench benchmark showing consistent token-cost reduction at improved utility across nine scientific domains. However, TAO (i) is limited to the clinical domain; (ii) implements routing and rules through an LLM orchestrator without a dedicated deterministic rule layer; (iii) conflates model-class selection and invocation policy in the same tiered hierarchy; (iv) relies on LLM agents throughout—a limitation explicitly acknowledged by its authors [3]; (v) has no metrological apparatus; and (vi) does not address model parsimony. Eywa addresses the LLM-only limitation by formalising the FM–LLM interface but, as a domain-collaboration framework, does not propose a layered trust architecture or a metrologically grounded metric suite. Stream 3 – Metrological uncertainty for ML. Bilson et al. [5] transfer VIM/GUM apparatus to ML classification via probability mass functions; Thompson et al. [6] bridge trustworthy AI and metrology conceptually. Both are limited to classification tasks without agentic architecture or cross-domain scope. Stream 4 – Model parsimony. Belcak et al. (NVIDIA) [7] argue that small language models are the appropriate medium for most agentic sub-tasks; Schwartz et al. [8] frame efficiency as a first-class evaluation metric. These works provide an economic and engineering imperative but do not offer a trust architecture. Parallel governance anchors—NIST AI RMF [9], NIST GAI Profile [10], ISO/IEC 42001 [11], EU AI Act [12], and the cross-domain framework by Herrera-Poyatos et al. [13]—provide regulatory scaffolding but not an engineering architecture. TRACE positions itself as the engineering layer that operationalises those obligations through measurable, architecture-embedded artefacts. Gap. No published work unifies (i) a multi-layer trust architecture, (ii) cross-domain instantiation in operationally critical domains, (iii) a metrologically grounded trust-metric suite, and (iv) a model-parsimony principle in a single framework. TRACE closes this four-sided gap. Genesis. TRACE emerged from the author’s recognition and formalisation of a shared structural pattern observed across two independently developed operationally critical systems: (i) a clinical decision-support architecture with metrological uncertainty monitoring [14], in which the author is a co-developer, currently under review at IEEE Instrumentation & Measurement Magazine; and (ii) an industrial multi-domain agentic-AI platform for upstream oil-and-gas operations spanning technology/operations/administrative loops, developed independently by Shcherban (2024–2026) [15, 16]. Recognising the four-layer decomposition, tiered escalation, and parsimony of learned components shared between these two structurally distinct systems is the conceptual core of TRACE; this paper formalises that recognition. Contributions. This work: (a) presents TRACE as a domain-neutral engineering framework grounded in six principles with explicit metrological analogues; (b) defines a four-layer reference architecture with an L2a/L2b split based on the parsimony principle and an architecturally separate L3 orchestration-and-escalation policy; (c) specifies a layer-wise trust-metric suite culminating in CPR—the first quantitative first-class parsimony metric in the trustworthyAI literature; (d) instantiates the framework in three domains (clinical, industrial multi-domain, judicial) and makes the architectural distinction from TAO precise.
2
2
The TRACE Framework
2.1
Six Design Principles
TRACE rests on six mutually reinforcing principles, each with a metrological analogue: 1. Evidence Traceability. Every prescriptive action carries a machine-readable evidence trail. Metrological analogue: traceability of measurement results [17]. 2. Bounded Human Supervision. Human oversight is an architectural layer with measurable workload and veto rights, not a cosmetic safety net. Analogue: operator supervision of measurement systems [18]. 3. Staged Autonomy. Action rights are earned through accumulated evidence, not granted by default. Analogue: type approval and operational qualification. 4. Bounded Context. The input context is explicitly specified, time-stamped, and kept current as a safety property. Analogue: operational limits and environmental conditions. 5. Metrological Accountability. All quality characteristics are specified, measurable, calibrated, and monitored over time [5, 19]. Analogue: calibration, drift monitoring, stability testing. 6. Model Parsimony / Right-Sized Components. The type of learned component (classical ML, specialised neural network, LLM, or hybrid) is chosen by fitness for purpose, not by LLM presumption [7, 8]. Analogue: fitness for purpose—the foundational metrological requirement. The acronym TRACE reflects five user-visible system properties (trustworthy, reasoned, accountable, context-bound, escalated ); model parsimony operates as an internal design constraint.
2.2
Four-Layer Reference Architecture
Figure 1 shows the four-layer decomposition. L1 – Deterministic / Physics-Based Core. Non-generative, interpretable, stable logic: rule systems, control theory, enumerated protocols, physical models, compliance scaffolding. Key architectural distinction from TAO [3] : TAO implements routing and rules through an LLM orchestrator; TRACE separates L1 as an independent, non-learned, formally verifiable layer and the primary trust anchor of the system. L2 – Learned-Component Inventory (stateless). A catalogue of trained components partitioned by the parsimony principle. L2 defines what the system can do; when and in what order components are invoked is the responsibility of L3 policy. Each L2 invocation is stateless and accountable to its calling policy. • L2a – Specialised / Classical ML. Gradient boosting (XGBoost, CatBoost), random forests, isolation forests, SVM, and classical neural architectures (CNN, LSTM, Transformer encoders for sensor fusion). Applied where inputs are structured, latency is critical, reproducibility is required, and probability calibration is formally verifiable [20, 21]. • L2b – Generative / LLM Validators. LLMs and generative models as contextual validators, semantic reviewers, and inconsistency detectors for free-form data. Invoked only where the task genuinely requires linguistic understanding. The L2a/L2b boundary is described here at the architectural level. A concrete protocol-level realisation of an analogous boundary is given by Eywa’s Tsaheylu interface [4], which formalises the communication contract between domain-specific foundation models (analogous to L2a) and 3
TRACE Four-Layer Reference Architecture with L2a / L2b Model-Parsimony split and key differentiators vs TAO
L4
Final accountability · override rights · autonomy-boundary calibration
Human Supervision
Measured load: Review Burden Index · Override Rate · Signal-to-Noise Ratio
L3
Orchestration & escalation policy over the L2 inventory
Escalation Policy
Inputs: task signature · risk tier · accumulated confidence · cost & latency budget Actions: invoke L2a / L2b components · re-invoke with adjusted context on low conf. L4 handoff: joint {risk ∧ confidence} · L2 inconsistency · invocation budget exhausted vs TAO [Fig. 6b]: TAO tiers = model classes (an L2-level concern in TRACE terms). TRACE L3 is a stateful policy over the L2 inventory — model choice (L2a/L2b) and invocation policy are decoupled.
L2
L2a — Classical ML
Learned Components
XGBoost · CatBoost · Random Forests · Isolation Forests · SVM · CNN / LSTM / Transformer-encoders for signal & time-series fusion
(split by Model Parsimony principle)
Structured inputs · low latency · formally verifiable probability calibration · Grounding: Guo 2017; Menghani 2023
L2b — LLM Validators
escalation / trust concentration ↑
(stateful)
Escalation trigger: high-confidence + high-risk (TAO Fig. 15 pattern) — not low confidence
LLMs as context validators · semantic reviewers · inconsistency detectors; used only where task truly requires linguistic understanding Motivated by TAO self-acknowledged LLM-only limitation (Kim et al. 2025, App. B) · Belcak 2025; Sapkota 2025
L1 Deterministic / PhysicsBased Core
Rule systems · control theory · enumerated protocols · physical models · compliance skeleton Non-generative · interpretable · formally verifiable · trust anchor Metrics: Rule Coverage Rate · Rule Consistency Index · Update Traceability Coefficient vs TAO [Kim et al. 2025]: L1 is a dedicated non-learned layer — not LLM-orchestrated routing.
TRACE acronym: Trustworthy · Reasoned · Accountable · Context-bound · Escalated 6th principle — Model Parsimony — operates as internal design constraint (CPR metric) Each layer maps to metrological analog: L1 ↔ traceability · L2 ↔ calibration · L3 ↔ staged qualification · L4 ↔ operator supervision
Figure 1. TRACE four-layer reference architecture.
Figure 1: TRACE four-layer reference architecture. L1 provides the deterministic rule core (trust anchor); L2 holds the stateless learned-component inventory, partitioned into classical ML (L2a) and LLM validators (L2b); L3 is the stateful orchestration-and-escalation policy; L4 is bounded human supervision. language-model agents (analogous to L2b). TRACE remains protocol-agnostic; Tsaheylu is one viable implementation of the L2a↔L2b boundary within the framework. L3 – Orchestration-and-Escalation Policy (stateful). An architecturally separate policy layer above the L2 inventory, operating on task state (passage history, accumulated confidence, risk context, latency and cost budget). The policy specifies: (i) which L2 components are invoked and in what order for a given task type; (ii) conditions for re-invocation with expanded context or switching to an alternative component; (iii) escalation triggers to L4—the joint condition of high risk and high confidence [3] (a confident error in a high-risk context is more dangerous than an acknowledged uncertainty), or L3-detected inconsistency between L2 components, or exhaustion of the re-invocation budget; (iv) cost accounting for each execution path (Tier Cost Coefficient). Key distinction from TAO: TAO implements tiered oversight as a hierarchy of LLM agents under an LLM orchestrator—in TRACE terms, this conflates L2-concern (model-class selection) and L3-concern (invocation policy). TRACE separates them: L2a vs. L2b is a static design decision by fitness for purpose; invocation sequence, re-invocations, and escalation are L3 policy, independently calibrated from operational statistics. This separation removes the TAO safetyparsimony trade-off. L4 – Bounded Human Supervision. Final accountability, veto, and calibration of AI authority bounds. L4 entry is determined by L3 policy, not by direct L2 requests.
2.3
Sub-Domain × Layer Matrix
Real deployments often span multiple operational sub-domains. TRACE represents this as a sub-domain × layer matrix : each sub-domain instantiates all four layers independently, and 4
cross-sub-domain coordination is provided by a master orchestrator. This yields systematic, transferable deployment templates for complex multi-domain systems. Instance B (Section 3.2) provides a documented precedent for this matrix across three sub-domains with principally different data types.
2.4
Trust Metric Suite
TRACE specifies 17 metrics: 12 layer-wise, 4 cross-cutting, and 1 parsimony metric. The suite extends the metrological apparatus of Bilson et al. [5] and Thompson et al. [6] from ML classification alone to the full agentic lifecycle: ingestion → inference → escalation → human adjudication. Layer-wise metrics: L1: Rule Coverage Rate, Rule Consistency Index, Update Traceability Coefficient. L2: Context Relevance Precision, Context Freshness Index, Input Perturbation Stability Rate. L3: Escalation Precision (correctness of L4 hand-off decisions); Tier Cost Coefficient (aggregate computational cost of the policy-chosen execution path); False Positive Attenuation (suppression of spurious escalations through L2 re-invocations). These metrics evaluate the policy itself, not the quality of individual L2 components [22]. L4: Review Burden Index, Override Rate, Signal-to-Noise Ratio. Cross-cutting metrics: Evidence Trail Completeness; Calibration Error [20]; Autonomy Boundary Compliance; Operational Stability Index. Parsimony metric – Computational Parsimony Ratio (CPR). CPR is the ratio of the resource cost of the most economical model that adequately solves the sub-task (meeting required accuracy, calibration, and operational reliability) to the resource cost of the deployed model. CPR = 1 indicates an optimal model choice; values substantially below 1 signal architectural overhead. Resource cost encompasses latency, compute, training, and operational expenditure. To the best of our knowledge, CPR is the first formalisation of Complexity-Performance Ratio as a first-class design principle in the trustworthy-AI literature. Empirical support for the underlying intuition is provided by Eywa’s EywaBench results [4]: across nine scientific domains, replacing LLM-only pipelines with FM–LLM heterogeneous collaboration improves utility while reducing token cost—precisely the regime CPR is designed to surface as architectural overhead. Each metric has a direct metrological analogue (accuracy, repeatability, systematic error, drift, fitness for purpose), enabling principled aggregation via GUM-style uncertainty propagation across layers [5, 19].
3
Multi-Domain Instantiations
The framework’s universality is demonstrated by three instantiations. Instances A and B are foundational precedents—independently developed systems whose shared structural pattern motivated TRACE’s formalisation. Instance C is a partially implemented extension demonstrating architectural portability.
3.1
Instance A: Clinical Decision Support (Foundational)
A detailed account of the clinical foundational implementation—architecture, metrics, validation— appears in a companion paper [14] currently under review at IEEE Instrumentation & Measurement Magazine. The TRACE layer projection is: • L1: Rule-based clinical logic, compatible with the FUTURE-AI traceability principle [1]. • L2a: Classical ML for structured prediction: risk scores, vital-sign time-series classification, laboratory data modelling. 5
Domain × Layer Matrix — Instance B (Industrial Multi-Domain Deployment) Model Parsimony applied per sub-domain: the dominant layer shifts with the evidence type
Technology physical processes, sensor signals, equipment telemetry Dominant: L2a (signal fusion, time-series, control)
Operations scheduling, logistics, resource allocation, KPI monitoring Dominant: L2a (forecasting, optimization)
L1
L2a
L2b
L3
L4
Deterministic Core
Classical ML
LLM Validators
Escalation Policy
Human Supervision
ACTIVE
DOMINANT
MINIMAL
ACTIVE
ACTIVE
Physics models, process limits
XGBoost anomaly
Incident log
detection, CNN/LSTM
free-text parsing
Anomaly-routing policy
Process engineer on-call review
· PID / MPC control · Hard-coded interlocks · Tolerance envelopes
· Signal fusion · Drift / degradation · Calibrated probabilities
· Post-hoc only · Not in hot path (CPR ≈ 0 for control)
· Re-invoke L2a on tolerance-band breach · Alert on risk ∧ conf.
· Override rights · Root-cause loop · Review Burden Index
ACTIVE
DOMINANT
PRESENT
ACTIVE
ACTIVE
Operating policies, SLA thresholds
Gradient boosting, forecasting, OR solvers
Shift-report summarization
Cost-tiered routing policy
Operations manager accountability
· Enumerated protocols · Resource constraints · Audit rules
· Demand / load models · Classification / routing · Reproducible pipelines
· Incident narratives · Cross-team handoff · Inconsistency flags
· Confirm on high-cost · SLA-deadline escalation · Audit-trail mandatory
· Approve reallocations · Calibrate thresholds · SNR tracking
Administrative
DOMINANT
MINIMAL
DOMINANT
ACTIVE
ACTIVE
contracts, compliance, regulatory reporting, documentation
Compliance skeleton, regulatory clauses
Document classification
Contract review, clause inconsistency
Risk-tier routing policy
Legal / compliance officer sign-off
· Rule coverage rate · Traceable updates · Non-generative anchor
· Metadata tagging · Anomaly in forms (narrow auxiliary role)
· Semantic validation · Cross-document check · Requires NL understanding
· Standard: auto by L1 · Non-standard: L2a↔L2b · High-stakes → L4
· Final accountability · Regulatory liability · Zero-tolerance gate
Dominant: L1 + L2b (rules + semantic understanding)
Intensity legend (layer role per sub-domain): DOMINANT — primary decision path
ACTIVE — regular contribution
PRESENT — narrow / supporting role
MINIMAL — auxiliary; near-zero contribution to hot path (CPR → 0)
Key insight: the dominant layer shifts with evidence type — L2a for quantitative sub-domains, L1+L2b for rule- and language-bound sub-domains. Figure 2. Domain × Layer matrix — illustrates how Model Parsimony (Principle 6) is instantiated differently across industrial sub-domains.
Figure 2: Sub-domain × layer matrix for Instance B (industrial multi-domain platform). The same four-layer architecture manifests differently across three sub-domains: in Technology and Operations, L2a (classical ML) dominates; in Administrative, the dominant pair is L1 (compliance rules) and L2b (semantic contract review). Model parsimony applies adaptively to the evidence type of each sub-domain. L3 is realised as an independent Escalation Policy in each sub-domain; a master orchestrator coordinates cross-sub-domain pipelines. • L2b: LLM validator for free-form clinical notes; contextual consistency checking of recommendations against patient history. • L3: Clinical escalation policy: routine cases are handled by L2a risk scores with calibrated probabilities; at boundary confidence or with a complex context, L2b is invoked for freeform data verification; on joint high-risk / high-confidence coincidence, or L3-detected inconsistency between L2a and L2b, mandatory clinician hand-off is triggered. Compatible with the TAO pattern [3] but with explicit L1/L2/L3 separation. • L4: Clinician as final reviewer, aligned with the TRIAD collaboration framework [2].
3.2
Instance B: Industrial Multi-Domain Platform (Foundational)
Attribution and scope. Instance B is an independently developed industrial agentic-AI platform for upstream oil-and-gas operations, designed by Shcherban (2024–2026) [15, 16]. The system simultaneously spans three operational sub-domains: technology (drilling, production, well operations), operations (maintenance, condition monitoring), and administrative (procurement, contract lifecycle, compliance). This system serves here as a structural reference exhibiting the TRACE pattern in a domain principally different from clinical decision support; detailed methodology, operational statistics, and component-level specifications are reported separately by the system’s developer (Paper 2, in preparation). Architectural projection on TRACE layers. At the level of architectural pattern visible in the public record of the system, each sub-domain instantiates all four TRACE layers, but with characteristically different L2 weighting (Figure 2): the technology and operations sub-domains are L2a-dominant, since the bulk of inference operates over structured, low-latency sensor and equipment-state signals where classical and specialised ML satisfy calibration and reproducibility 6
requirements at substantially lower cost than LLMs; the administrative sub-domain is L1+L2bdominant, since regulatory and contractual rules must be formally specified rather than learned (L1), while semantic review of free-form contract text and unstructured documents is the natural target for LLM validators (L2b). A master orchestrator coordinates cross-sub-domain pipelines (e.g., the chain procurement → contract → compliance check → operational integration when new downhole equipment is introduced) and aggregates TRACE metrics across sub-domains, enabling measurement of operational stability for the platform as a whole. Significance for TRACE. At the level of structural pattern, Instance B demonstrates three claims that motivated TRACE’s formalisation: (i) the parsimony principle is empirically natural—two of three sub-domains are L2a-dominant despite the broader trend of LLM-default deployment; (ii) the sub-domain × layer matrix (§2.3) is documented in a real-world design precedent; (iii) the architecture transfers across principally different evidence types (sensor signals vs. legal documents) within a single platform. Component-level validation of these claims belongs to the domain paper.
3.3
Instance C: Judicial AI Assistant (Partially Implemented)
L1—procedural and substantive legal norms [23, 24]. L2a—classical classifiers for document categorisation, precedent relevance ranking, NER for parties and circumstances. L2b—LLM for case-material analysis, precedent semantic correlation, draft document preparation. Empirical anchor: Magesh et al. [25] report a 17% hallucination rate in commercial legal RAG systems, providing direct motivation for multi-layer trust oversight. L3—legal escalation policy: L2a/NER retrieval selects relevant precedents; L2b mandatory hallucination verification; at high claim stakes and a confident recommendation, juridical L4 review is required before execution. L4— legal professional (judge, assistant) [26]. Instance C is currently partially implemented as an AI assistant for judicial-decision analysis within the EU project “Pravo-Justice” — “Support of the Supreme Court on improving (modernising) the Database of Legal Positions” (funded by Expertise France) [27, 28]. This serves as a demonstration of architectural portability: the same four-layer pattern and the same metrics apply in a domain with a principally different governance context (judicial ethics, burden of proof, procedural accountability).
3.4
Unity and Variation
Across all three instances the same reference architecture, the same six principles, and the same metric suite apply—only the instantiated parameters and the L2a/L2b split vary. Instance B additionally demonstrates that the sub-domain × layer matrix operates within a single system for three sub-domains with principally different data types. This invariance is the central empirical claim of the framework.
4
Discussion
4.1
Relation to TAO, Revisited
TAO [3] supplies two empirical facts that TRACE preserves and absorbs: (i) tiered oversight reduces individual-agent error rates by 16.9–24.3% (Table 6); (ii) confident errors in high-risk contexts are more dangerous than acknowledged uncertainty (Fig. 15). TRACE reinterprets both within a purely layered architecture: (i) becomes the lower bound on the expected effect of any correctly configured L3 policy operating over a heterogeneous L2 inventory; (ii) becomes the precise specification of the L3→L4 hand-off trigger (joint risk ∧ confidence). TAO’s descending-vs-ascending capability finding—that descending is safer—is reinterpreted as valid within a single model class and serves as a calibration anchor for L3 policy, not as an 7
architectural norm. This interpretation is only visible after separating L2-concern (model-class selection) from L3-concern (invocation policy), which TAO does not do.
4.2
Relation to Eywa: Complementary Decomposition
Eywa [4] and TRACE address overlapping problems from complementary angles. Eywa provides a protocol-level answer to the question of how heterogeneous foundation models and LLMs should communicate, with three concrete instantiations (EywaAgent, EywaMAS, EywaOrchestra) and an empirical benchmark (EywaBench) over nine scientific domains. TRACE provides an architectural and metrological answer to the question of what layer-wise decomposition operationally critical AI should adopt, and how the resulting trust properties should be measured. The two are not competing: Eywa’s Tsaheylu protocol is a viable realisation of the TRACE L2a↔L2b boundary; conversely, TRACE supplies the L1 deterministic anchor, the L3 stateful policy, the L4 supervision layer, and the GUM/VIM-grounded metric suite that Eywa, by design, does not address. A natural integration path—layered TRACE architecture with a Tsaheylu-conformant L2a/L2b interface—is left to follow-up work.
4.3
Parsimony as an Architectural Constraint
CPR introduces a quantitative parsimony metric into trustworthy-AI alongside accuracy, calibration, and review burden. Its practical value is twofold: the metric surfaces architectural overhead (an LLM deployed where a gradient-boosted classifier satisfies the required calibration yields CPR ≪ 1) and provides an auditable artefact for certification bodies (e.g., under EU AI Act Art. 15 [12]). The architectural correlate of CPR is the L2a/L2b split: it makes the parsimony choice a visible design decision rather than an implicit LLM-default bias. Instance B makes this concrete: two of three sub-domains are naturally L2a-dominant; an LLM-by-default deployment would be measurably over-engineered. Limitations. The Trust Metric Suite is currently specified analytically; empirical calibration from production deployments is forthcoming in the domain-specific follow-up papers. CPR requires operational benchmarking for normalisation against domain-specific cost models. Instance C is partially implemented; the EU project deployment covers judicial-decision analysis but full operational statistics await the completion of Phase 2.
5
Conclusion
TRACE provides a metrologically grounded, cross-domain engineering framework for trustworthy agentic AI in operationally critical domains. Its four-layer architecture, six design principles, 17-metric trust suite, and CPR parsimony metric synthesise four previously disconnected literature streams into a single reusable engineering blueprint. Two foundational precedents—clinical decision support and an industrial multi-domain platform—validate the shared structural pattern that motivated TRACE’s formalisation. TRACE operationalises trust and model parsimony as engineering outcomes, shifting the design question from “how intelligent is this model?” to “what is this system permitted to do, on what grounds, under what supervision, with what measurable uncertainty, and is the right type of model being used for each sub-task?”
Acknowledgements The author thanks Andriy Shcherban, whose independently developed industrial multi-domain platform served as the empirical catalyst for TRACE’s formalisation; the structural pattern shared between that system and the clinical foundational implementation motivated the abstraction presented here. 8
The clinical foundational implementation (Instance A) is reported separately by S. Zabolotnii, V. Holinko, and O. Antonenko [14]; the author thanks V. Holinko and O. Antonenko for the collaboration on that companion work.
References [1] Karim Lekadir, Alejandro F. Frangi, Antonio R. Porras, Ben Glocker, Celia Cintas, Curtis P. Langlotz, et al. FUTURE-AI: International consensus guideline for trustworthy and deployable artificial intelligence in healthcare. BMJ, 388:e081554, 2025. doi: 10.1136/bmj-2024-081554. Correction: BMJ 2025;388:r340, doi:10.1136/bmj.r340. [2] Jing Li, Zhen-Chang Zhou, Zi-Chen Wang, Hui Lv, et al. Prioritizing human-AI collaboration in healthcare: the TRIAD framework for trustworthy governance, real-world value, and integrated adaptive deployment. Military Medical Research, 12:97, 2026. doi: 10.1186/s40779-026-00684-w. DOI to be revalidated at camera-ready. [3] Yubin Kim, Hyewon Jeong, Chanwoo Park, Eugene Park, Haipeng Zhang, Xin Liu, Hyeonhoon Lee, Daniel McDuff, Marzyeh Ghassemi, Cynthia Breazeal, Samir Tulebaev, and Hae Won Park. Tiered agentic oversight: A hierarchical multi-agent system for healthcare safety. arXiv preprint arXiv:2506.12482v2, September 2025. URL https: //arxiv.org/abs/2506.12482. Accepted at ICML 2025; project page: https://tieredagentic-oversight.github.io/; verify proceedings DOI before camera-ready. [4] Zihao Li, Jiaru Zou, Feihao Fang, Xuying Ning, Mengting Ai, Tianxin Wei, Sirui Chen, Xiyuan Yang, and Jingrui He. Eywa: Heterogeneous scientific foundation model collaboration. arXiv preprint, 2026. URL https://arxiv.org/abs/2604.27351. Introduces the FM–LLM Tsaheylu interface; instantiations EywaAgent / EywaMAS / EywaOrchestra; benchmark EywaBench across 9 scientific domains. [5] Steven Bilson, Maurice Cox, Alexandra Pustogvar, and Andrew Thompson. A metrological framework for uncertainty evaluation in machine learning classification models. Metrologia, 62(6):064001, 2025. doi: 10.1088/1681-7575/ae1bae. [6] Andrew Thompson, Tameem Adel, Steven Bilson, and Mark Levene. Trustworthy artificial intelligence in the context of metrology. In Producing Artificial Intelligent Systems: The Roles of Benchmarking, Standardisation and Certification. Springer, 2024. doi: 10.1007/9783-031-55817-7_4. arXiv:2406.10117. [7] Peter Belcak, Greg Heinrich, Shizhe Diao, Yonggan Fu, Xin Dong, Saurav Muralidharan, Yingyan Celine Lin, and Pavlo Molchanov. Small language models are the future of agentic AI. arXiv preprint arXiv:2506.02153, June 2025. doi: 10.48550/arXiv.2506.02153. NVIDIA Research position paper. [8] Roy Schwartz, Jesse Dodge, Noah A. Smith, and Oren Etzioni. Green AI. Communications of the ACM, 63(12):54–63, 2020. doi: 10.1145/3381831. [9] Elham Tabassi. Artificial intelligence risk management framework (ai rmf 1.0). Nist ai 100-1, National Institute of Standards and Technology, January 2023. [10] National Institute of Standards and Technology. Artificial intelligence risk management framework: Generative artificial intelligence profile. Nist ai 600-1, NIST, July 2024. [11] ISO/IEC. ISO/IEC 42001:2023 — information technology — artificial intelligence — management system. International Standard, December 2023. URL https://www.iso.org/ standard/81230.html. 9
[12] European Parliament and Council of the European Union. Regulation (eu) 2024/1689 of 13 june 2024 laying down harmonised rules on artificial intelligence (AI Act). Official Journal of the European Union, L 2024/1689, 12 July 2024, 2024. URL http://data.europa.eu/ eli/reg/2024/1689/oj. [13] Andres Herrera-Poyatos, Javier Del Ser, Marcos Lopez de Prado, Fei-Yue Wang, Enrique Herrera-Viedma, and Francisco Herrera. A framework for responsible AI systems: Building societal trust through domain definition, trustworthy AI design, auditability, accountability, and governance. arXiv preprint arXiv:2503.04739, 2025. URL https://arxiv.org/abs/ 2503.04739. [14] Serhii Zabolotnii, Viktoriia Holinko, and Olha Antonenko. From black-box confidence to measurable trust in clinical AI: A framework for evidence, supervision, and staged autonomy. IEEE Instrumentation and Measurement Magazine, September 2026. doi: 10.48550/arXiv. 2604.26671. URL https://arxiv.org/abs/2604.26671. arXiv:2604.26671. Special Issue "A Measure of Trust in Healthcare" (guest editors: O. Postolache, A. Solanas); submitted April 2026, under review. Update volume/issue/pages before camera-ready. [15] Andrii Shcherban. A method for controlling the drilling process at oil and gas fields using agentic artificial intelligence. Ukrainian Utility Model Patent Application, August 2025. Filed 19 August 2025; applicant: A. Shcherban. Status: utility model application. [16] Andrii Shcherban. Agentic AI for upstream operational intelligence: A unified framework. U.S. Copyright Office (deposit copy submitted), March 2026. Submitted 15 March 2026; application number pending final registration confirmation. Update before camera-ready. [17] JCGM. International vocabulary of metrology — basic and general concepts and associated terms (VIM), 3rd edition. JCGM 200:2012, 2012. [18] Eduardo Mosqueira-Rey, Elena Hernández-Pereira, David Alonso-Ríos, José BobesBascarán, and Álvaro Fernández-Leal. Human-in-the-loop machine learning: a state of the art. Artificial Intelligence Review, 56(4):3005–3054, 2023. doi: 10.1007/s10462-022-10246-w. [19] JCGM. Evaluation of measurement data — guide to the expression of uncertainty in measurement (GUM). JCGM 100:2008, 2008. [20] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks. In Proceedings of the 34th International Conference on Machine Learning (ICML), volume 70 of PMLR, pages 1321–1330, 2017. arXiv:1706.04599. [21] Gaurav Menghani. Efficient deep learning: A survey on making deep learning models smaller, faster, and better. ACM Computing Surveys, 55(12):259:1–259:37, 2023. doi: 10. 1145/3578938. [22] Yi Dong, Wenjie Huang, Vikas Bharti, Victoria Cox, Alec Banks, Sen Wang, Xingyu Zhao, Sven Schewe, and Xiaowei Huang. Reliability assessment and safety arguments for machine learning components in system assurance. ACM Transactions on Embedded Computing Systems, 22(3):39:1–39:48, 2023. doi: 10.1145/3570918. [23] Council of Europe / CEPEJ. European ethical charter on the use of artificial intelligence in judicial systems and their environment. Strasbourg, 3–4 December 2018, 2018. URL https: //rm.coe.int/ethical-charter-en-for-publication-4-december-2018/16808f699c. [24] Council of Europe / CEPEJ. Operationalisation of the CEPEJ ethical charter — assessment tool. Cepej(2023)16rev, Council of Europe, June 2025. Revised 5 June 2025.
10
[25] Varun Magesh et al. Hallucination-free? assessing the reliability of leading AI legal research tools. Journal of Empirical Legal Studies, 22:216, 2025. [26] Daan Kolkman, Floris Bex, Nidhi Narayan, and Michelle van der Put. Justitia ex machina: The impact of an AI system on legal decision-making and discretionary authority. Big Data & Society, 11(2), 2024. doi: 10.1177/20539517241255101. [27] High Qualification Commission of Judges of Ukraine. The European Union funded project “Support to Justice Sector Reforms in Ukraine (PRAVO-JUSTICE III)”. https://vkksu.gov.ua/en/page/european-union-funded-project-supportjustice-sector-reforms-ukraine-pravo-justice-iii, 2024. Implementing agency: Expertise France S.A.S.; duration: January 2024 – June 2026. [28] EU Project Pravo-Justice. Database of legal positions of the Supreme Court: Uniformity of jurisprudence online. https://www.pravojustice.eu/en/post/baza-pravovih-pozicijverhovnogo-sudu-yednist-sudovoyi-praktiki-onlajn-en, April 2024. Implemented by Expertise France; funded by the European Union.
11