Architecture-Centric Cross-Layer ISAC for 6G: Operational Contracts, Uncertainty, and End-to-End Validation

1. Introduction

Integrated sensing and communication (ISAC), one of the six IMT-2030 usage scenarios [1], shares spectrum, hardware and signal processing between communication and sensing. Its technical core includes waveform and transceiver co-design, estimation, detection, localization and communication–sensing tradeoffs [2,3,4,5,6,7].

ISAC has attracted significant industrial and research interest. The research frontier is now cross-layer, as demonstrated by [7]: sensing observations influence association, scheduling, mobility, learning, computation placement, security decisions, and application actions, while those functions determine which observations can be acquired and delivered in time [8,9,10,11].

This architecture-centric survey complements the industrial-academic perspective of Wymeersch et al. [7] by comparing representative studies through consistent questions about cross-layer interfaces. It emphasizes whether reported outputs remain interpretable, timely, and admissible at the receiving function. Section 2 defines the coverage, evidence interpretation, and relation to prior surveys.

A locally successful ISAC component may become unusable when its output is fused, acted upon, or executed elsewhere. This survey applies six operational contracts across seven functional layers to compare reported outputs, assumptions, and downstream requirements, tracking uncertainty and supporting evidence. It links physical estimation and communication metrics to network, service, intelligence, computation, security and validation functions; examines compatibility among the Cramér–Rao bound (CRB), mutual information, detection, task loss and latency; and formulates research questions on uncertainty transfer, sensing-session control, freshness-aware computation, hardware-constrained learning, privacy and reproducible evaluation.

The remainder is organised as follows. Section 2 defines the survey scope, evidence interpretation, and relation to prior work. Section 3 introduces the operational framework. Section 4, Section 5 and Section 6 cover physical, network/service, and intelligence/computation functions. Section 7 treats propagation, hardware, security, and trust. Section 8 covers standards and validation. Section 9 presents the contract audit and research agenda, and Section 10 provides the paper’s conclusion.

2. Survey Scope, Evidence Interpretation, and Positioning

2.1. Scope and Coverage

This is a selective, architecture-centric survey. Table 1 summarizes its coverage. We include studies where models, outputs, decisions, or validation conditions affect another ISAC function. We examine propagation, hardware, security, and standards through their consequences for model validity and admissibility.

The representative audit in Table 11 compares nine selected examples spanning local design, coupled estimation or control, impairment modelling, measured data, prototypes and networked sounding. For each example, the six operational contracts distinguish the reported capability from requirements introduced by subsequent fusion, control, or service use. The comparison identifies connections and limitations within the stated models and experiments; it does not estimate the prevalence or rank the importance of gaps across the field.

Table 1. Technical coverage and deliberate boundaries of the survey.

Area	Covered in this survey	Outside the principal scope
Physical layer and metrics	Signal models, waveform and beamforming classes, estimation and detection criteria, fundamental tradeoffs, and hardware assumptions that affect the use of sensing information by higher layers.	Exhaustive waveform catalogues, repeated algorithm derivations, and comparisons that do not change a cross-layer interface or operational conclusion.
Network and sensing services	Cooperative architectures, synchronization, association, fusion, information age, sensing-session control, processing location, mobility support, and application exposure.	Conventional network optimisation with no sensing dependency and clause-by-clause protocol description unrelated to an end-to-end sensing service.
Intelligence and computation	Model-driven and constrained learning, sequential and distributed control, representation selection, offloading, queues, latency, memory, energy, and task-oriented performance.	General AI, edge-computing, or semantic-communication studies without an explicit ISAC observation, control decision, or downstream task.
Propagation, hardware, security, and validation	Propagation regimes that change model validity; RF and converter impairments; privacy and trust; standards, datasets, prototypes, testbeds, and reproducibility.	Standalone surveys of individual enabling technologies, generic wireless security, and conformance procedures that do not affect the ISAC interfaces considered here.

Table 1. Technical coverage and deliberate boundaries of the survey.

Area	Covered in this survey	Outside the principal scope
Physical layer and metrics	Signal models, waveform and beamforming classes, estimation and detection criteria, fundamental tradeoffs, and hardware assumptions that affect the use of sensing information by higher layers.	Exhaustive waveform catalogues, repeated algorithm derivations, and comparisons that do not change a cross-layer interface or operational conclusion.
Network and sensing services	Cooperative architectures, synchronization, association, fusion, information age, sensing-session control, processing location, mobility support, and application exposure.	Conventional network optimisation with no sensing dependency and clause-by-clause protocol description unrelated to an end-to-end sensing service.
Intelligence and computation	Model-driven and constrained learning, sequential and distributed control, representation selection, offloading, queues, latency, memory, energy, and task-oriented performance.	General AI, edge-computing, or semantic-communication studies without an explicit ISAC observation, control decision, or downstream task.
Propagation, hardware, security, and validation	Propagation regimes that change model validity; RF and converter impairments; privacy and trust; standards, datasets, prototypes, testbeds, and reproducibility.	Standalone surveys of individual enabling technologies, generic wireless security, and conformance procedures that do not affect the ISAC interfaces considered here.

For terminology, JRC/DFRC denotes integration centered on waveform, aperture, RF-chain, or transceiver design; ISAC/JCAS denotes broader network or service integration; ISCC requires an explicit computation, placement, offloading, or processor-allocation decision; and task-oriented ISAC evaluates the system through a declared downstream loss, risk, or control objective. These categories may overlap in systems that integrate several decision levels.

2.2. Evidence Interpretation

Validation strength affects how you can interpret a reported result. Evidence is described at the level of the specific claim: E0 denotes an analytical formulation without numerical evaluation; E1, ideal simulation; E2, impairment-aware or measurement-calibrated simulation; E3, offline measurement or a post-processed hardware experiment; E4, a real-time single-node prototype; and E5, a real-time multi-node or networked testbed. These categories identify the evaluation setting; they do not provide a universal ranking of realism or scientific quality. A paper may support different claims in different settings. Independent reproduction and standards alignment require separate supporting evidence and are not inferred from an E0–E5 designation.

Analytical and simulation results can establish performance under a declared model; they do not alone validate synchronization, calibration stability, converter dynamic range, queueing, real-time execution, interoperability or reproducibility. Prototype interpretation likewise depends on waveform, bandwidth, geometry, hardware, processing platform, trial conditions and baseline. Evidence maturity qualifies conclusions and identifies validation gaps.

2.3. Relation to Prior Work and Cross-Layer Insights

Foundational surveys cover waveform, signal-processing and fundamental-limit aspects [2,3,4,5,6]. Wymeersch et al. [7] already connect Hexa-X-II priorities, key performance indicators (KPIs), key value indicators (KVIs), simulations and proof-of-concept activities across physical design, hardware and higher-layer functions. Luo et al. [11] cover layered technologies, standards, prototypes and testbeds; Wen et al. [9] organise ISCC models and resource allocation; and González-Prelcic et al. [8] connect the sixth-generation wireless communication (6G) vision with techniques, data and quantitative examples.

Building on these views, Table 11 consistently compares reported outputs with receiving-function requirements for uncertainty, update timescales, decisions, and operating constraints. Requirements for transfer to another function or operating setting are interpreted within each source’s model, scope and supporting evidence.

3. Operational Framework

The framework compares the information supplied and required at functional interfaces. Its six contracts organise the literature analysis; they do not define a new protocol stack or formal contract theory.

3.1. Six Contracts Across Seven Functional Layers

Table 2 applies the six operational contracts to the seven functional layers. For each layer, the survey asks six questions: C1 what object enters and leaves; C2 which uncertainty or confidence accompanies it; C3 at which timescale it is updated; C4 which variables are controlled; C5 how achieved performance is compared; and C6 which physical, computational, security, legal, or standard conditions make the action admissible. C5 and C6 are deliberately separated: latency is an evaluation measure when comparing implementations, whereas a hard deadline removes an operating point from the feasible set.

Propagation and hardware primarily constrain model validity and admissibility. Their timescale and decision entries therefore include slower configuration and calibration choices — carrier band, array model, platform state and calibration schedule — alongside the faster decisions made by other functions.

Table 2. Six operational contracts applied across seven functional layers.

Layer	Exchanged object and uncertainty (C1–C2)	Timescale and decision (C3–C4)	Evaluation and admissibility (C5–C6)
Physical observation	Decoded data, detections, parameter estimates or sufficient statistics; covariance, posterior, detector operating point, and residual calibration.	Sample, symbol, frame, and coherence-block scales; waveform, beam, receiver, detector, and estimator.	Rate/BLER, achieved error, detection at fixed false alarm, ambiguity and energy; constrained by numerology, power, spectrum, RF dynamic range, and synchronization.
Propagation/hardware validity	Channel/target model and its declared validity; blockage, visibility, calibration, oscillator, converter, and self-interference residuals.	Coherence to trajectory/calibration scales; band, array model, platform, surface state, and calibration schedule.	Model mismatch, identifiability, coverage, and calibration burden; constrained by propagation physics and hardware capability.
Network coordination	Local estimates or features, timestamps, dependence information, associations and resource state; fused belief and information age.	Slot, frame, scheduling and handover scales; serving set, resource allocation, sharing level, and fusion rule.	Utility, distortion, age, fairness, fronthaul and consistency; constrained by synchronization, signalling, capacity and scalability.
Intelligence	Observations, beliefs, queue/task state and model-validity indicators; action, calibrated confidence, abstention and fallback.	Inference step, training round and episode; model, policy, adaptation and threshold.	Task loss, calibration, robustness and constraint violations; constrained by action feasibility, memory, privacy and control-loop timing.
Computation/task	Representation, model graph, deadline, processor and queue state; placement, split, allocation and completed result.	Task period, queue slot and offloading epoch.	Mean/tail latency, deadline violation, task loss, energy and memory; constrained by radio/MEC capacity, data locality and hard deadlines.
Security/trust	Credentials, provenance, threat indicators and model updates; authenticated/sanitised object, trust state and recovery action.	Packet, event, session, model-update and audit scales.	Detection, integrity, privacy leakage, robustness and recovery; constrained by consent, authorization, privacy budget, attestation and safety policy.
Standards/validation	Use cases, KPIs, procedures, datasets, prototypes and metadata; interoperability and reproducibility evidence.	Trial, release, certification and maintenance cycles.	Confidence intervals, replication variance and interoperability; constrained by document status, regulation, test capability and reporting requirements.

Table 2. Six operational contracts applied across seven functional layers.

Layer	Exchanged object and uncertainty (C1–C2)	Timescale and decision (C3–C4)	Evaluation and admissibility (C5–C6)
Physical observation	Decoded data, detections, parameter estimates or sufficient statistics; covariance, posterior, detector operating point, and residual calibration.	Sample, symbol, frame, and coherence-block scales; waveform, beam, receiver, detector, and estimator.	Rate/BLER, achieved error, detection at fixed false alarm, ambiguity and energy; constrained by numerology, power, spectrum, RF dynamic range, and synchronization.
Propagation/hardware validity	Channel/target model and its declared validity; blockage, visibility, calibration, oscillator, converter, and self-interference residuals.	Coherence to trajectory/calibration scales; band, array model, platform, surface state, and calibration schedule.	Model mismatch, identifiability, coverage, and calibration burden; constrained by propagation physics and hardware capability.
Network coordination	Local estimates or features, timestamps, dependence information, associations and resource state; fused belief and information age.	Slot, frame, scheduling and handover scales; serving set, resource allocation, sharing level, and fusion rule.	Utility, distortion, age, fairness, fronthaul and consistency; constrained by synchronization, signalling, capacity and scalability.
Intelligence	Observations, beliefs, queue/task state and model-validity indicators; action, calibrated confidence, abstention and fallback.	Inference step, training round and episode; model, policy, adaptation and threshold.	Task loss, calibration, robustness and constraint violations; constrained by action feasibility, memory, privacy and control-loop timing.
Computation/task	Representation, model graph, deadline, processor and queue state; placement, split, allocation and completed result.	Task period, queue slot and offloading epoch.	Mean/tail latency, deadline violation, task loss, energy and memory; constrained by radio/MEC capacity, data locality and hard deadlines.
Security/trust	Credentials, provenance, threat indicators and model updates; authenticated/sanitised object, trust state and recovery action.	Packet, event, session, model-update and audit scales.	Detection, integrity, privacy leakage, robustness and recovery; constrained by consent, authorization, privacy budget, attestation and safety policy.
Standards/validation	Use cases, KPIs, procedures, datasets, prototypes and metadata; interoperability and reproducibility evidence.	Trial, release, certification and maintenance cycles.	Confidence intervals, replication variance and interoperability; constrained by document status, regulation, test capability and reporting requirements.

3.2. Cross-Layer Trace and Metric Discipline

Figure 1 traces information and uncertainty across ISAC functions. Estimation error, detection reliability, model confidence, security risk, latency variation, and experimental uncertainty require explicit statistical, physical, or decision models for downstream use or transformation; they are not directly interchangeable.

Rate, achieved distortion, Bayes risk, task loss, latency, energy, memory, and deadline violation may coexist in a matched-constrained optimization or a Pareto analysis. CRB, sensing mutual information, Signal-to-clutter-plus-noise ratio (SCNR), ambiguity sidelobes, a single Receiver operating characteristic (ROC) point, and uncalibrated model confidence are typically bounds or surrogates. They become operationally comparable only after the estimator or decision rule, state distribution, resource budget, and mapping to downstream loss are fixed [6,12,13,14].

4. Physical-Layer Foundations and Metric Compatibility

The physical layer creates the observations and communication outputs consumed by the rest of the ISAC architecture. This section establishes signal-model assumptions, distinguishes achieved performance from bounds and surrogates, and compares waveform and spatial designs under matched resources [4,15,16].

Waveform comparisons match bandwidth and spectral masks, frame and CPI duration, radiated energy, guard and reference overhead, aperture/ RF chains, coding/BLER targets, amplifier operating points, and receiver complexity. Practical NR resource-grid and reference-signal assumptions are considered in relation to the current Release-20 RAN study, 3GPP TR 38.765 [17].

4.1. Signal Models and Assumptions

Consider an $N_t$-antenna transmitter serving K communication users while illuminating Q point targets. Let $\mathbf{s}\left(t\right)={[s_1\left(t\right),\dots ,s_K\left(t\right)]}^T$, $\mathbf{W}=[{\mathbf{w}}_1,\dots ,{\mathbf{w}}_K]$, and ${\mathbf{x}}_s\left(t\right)\in {\mathbb{C}}^{N_t}$. For the sensing array, ${\mathbf{a}}_t\left(\varphi \right)\in {\mathbb{C}}^{N_t}$ and ${\mathbf{a}}_r\left(\theta \right)\in {\mathbb{C}}^{N_r}$. A representative narrowband transmit signal is

$$\mathbf{x}\left(t\right)=\mathbf{W}\mathbf{s}\left(t\right)+{\mathbf{x}}_s\left(t\right),$$

(1)

where $\mathbb{E}\left[\mathbf{s}\right(t\left)\right]=\mathbf{0}$, $\mathbb{E}\left[\mathbf{s}\left(t\right){\mathbf{s}}^H\left(t\right)\right]={\mathbf{I}}_K$, and the dedicated sensing component is zero mean with covariance ${\mathbf{Q}}_s=\mathbb{E}\left[{\mathbf{x}}_s\left(t\right){\mathbf{x}}_s^H\left(t\right)\right]⪰\mathbf{0}$. For the baseline used below we further assume $\mathbb{E}\left[{\mathbf{x}}_s\left(t\right){\mathbf{s}}^H\left(t\right)\right]=\mathbf{0}$, so that

$${\mathbf{R}}_x\triangleq \mathbb{E}\left[\mathbf{x}\left(t\right){\mathbf{x}}^H\left(t\right)\right]=\mathbf{W}{\mathbf{W}}^H+{\mathbf{Q}}_s.$$

If payload and sensing components are intentionally correlated or data dependent, the corresponding cross-covariance terms must be retained rather than absorbed into ${\mathbf{Q}}_s$.

A flat-fading communication user with ${\mathbf{h}}_k\in {\mathbb{C}}^{N_t}$ observes

$$y_k\left(t\right)={\mathbf{h}}_k^H\mathbf{x}\left(t\right)+n_k\left(t\right),$$

(2)

where $\mathbb{E}\left[\right|n_k\left(t\right){|}^2]={\sigma }_k^2$. Rate and coded-error claims under this flat-fading model require the noise distribution and the channel knowledge available to the receiver to be specified [18,19].

Imperfect transmitter channel knowledge can be represented as ${\mathbf{h}}_k={\widehat{\mathbf{h}}}_k+{\mathbf{e}}_k$, with a declared conditional error distribution, for example ${\mathbf{e}}_k\mid {\widehat{\mathbf{h}}}_k\sim \mathcal{CN}(\mathbf{0},{\mathbf{C}}_{e,k})$. An outage-constrained alternative to the known-channel SINR constraint is

$$Pr\left\{{\mathrm{SINR}}_k(\mathbf{W},{\mathbf{Q}}_s;{\widehat{\mathbf{h}}}_k+{\mathbf{e}}_k)<{\gamma }_k|{\widehat{\mathbf{h}}}_k\right\}\le {ϵ}_k,0\le {ϵ}_k<1.$$

The design variables are selected from the available channel estimate, and the probability is evaluated under the stated error law [20]. Receiver channel-estimation errors must separately be included in the decoding model when coded-error probability or an achievable rate is claimed.

A general bistatic MIMO sensing receiver with $N_r$ antennas can be represented, under a stop-and-hop point-target approximation, as

$$\begin{array}{cc}\hfill {\mathbf{y}}_s\left(t\right)={\displaystyle \sum _{q=1}^Q}& {\alpha }_q{\mathbf{a}}_r\left({\theta }_q\right){\mathbf{a}}_t^H\left({\varphi }_q\right)\mathbf{x}(t-{\tau }_q)e^{j2\pi {\nu }_{q}t}\hfill \\ \hfill & +\mathbf{c}\left(t\right)+\mathbf{i}\left(t\right)+{\mathbf{i}}_{\mathrm{leak}}\left(t\right)+\mathbf{n}\left(t\right).\hfill \end{array}$$

(3)

Here ${\alpha }_q$, ${\tau }_q$, ${\nu }_q$, ${\theta }_q$, and ${\varphi }_q$ denote target response, delay, Doppler frequency, and angles of arrival and departure. The complex target response includes radar cross section (RCS) and propagation loss [21]. For a fixed far-field geometry and declared normalization, it can be written as ${\alpha }_q={\kappa }_q\sqrt{{\sigma }_{\mathrm{RCS},q}}{e}^{j{\psi }_q}$, where ${\sigma }_{\mathrm{RCS},q}$ is the RCS in ${\mathrm{m}}^2$, ${\kappa }_q$ contains propagation and normalization factors not already included in the array responses, and ${\psi }_q$ is the residual phase. RCS depends on frequency, aspect, polarization, and monostatic/bistatic geometry. A nonfluctuating model holds it fixed over the stated observation interval; a fluctuating model, such as a Swerling model, must specify its mean, distribution, and temporal correlation [22].

In a monostatic far-field system, the transmit and receive directions are geometrically linked. Under the radial convention, $R_q=c{\tau }_q/2$ and $v_{r,q}=\lambda {\nu }_q/2$ where, $v_{r,q}$ denotes the radial velocity of target q relative to the radar in $\mathrm{m}/\mathrm{s}$, ${\nu }_q$ is its Doppler frequency shift in hertz, and $\lambda$ is the carrier wavelength in metres. In a bistatic/multistatic system, the directions are generally distinct; delay constrains the total transmitter–target–receiver path length, and bistatic Doppler is a geometry-dependent range-rate projection.

The terms coloured clutter $\mathbf{c}\left(t\right)$, external interference $\mathbf{i}\left(t\right)$, transmitter-origin leakage ${\mathbf{i}}_{\mathrm{leak}}\left(t\right)$, and noise $\mathbf{n}\left(t\right)$ in Equation (3) require declared deterministic/random nuisance models and covariance or probability laws before interpreting CRB, SCNR, GLRT, CFAR, or ROC results. In simultaneous-transmit/receive monostatic operation, ${\mathbf{i}}_{\mathrm{leak}}\left(t\right)$ represents residual self-interference; in a separated bistatic surveillance receiver, it includes direct-path interference from the transmitter. RIS-assisted configurations can additionally introduce reflected-path interference [23]. These components must be separated from external interference to avoid double counting. The deterministic unknown point-target ${\alpha }_q$, fluctuating target with a specified prior, and an extended/distributed target response yield different likelihoods, Fisher information matrices, and detectors [21].

An extended target can instead be approximated by $L_q$ scattering centres belonging to the same physical object [21]. Under the same narrowband approximation, a corresponding extension is

$$\begin{array}{cc}\hfill {\mathbf{y}}_s\left(t\right)=& {\displaystyle \sum _{q=1}^Q}{\displaystyle \sum _{\ell =1}^{L_q}}{\alpha }_{q\ell }{\mathbf{a}}_r\left({\theta }_{q\ell }\right){\mathbf{a}}_t^H\left({\varphi }_{q\ell }\right)\mathbf{x}(t-{\tau }_{q\ell })e^{j2\pi {\nu }_{q\ell }t}+{\mathbf{d}}_s\left(t\right),\hfill \\ \hfill {\mathbf{d}}_s\left(t\right)=& \mathbf{c}\left(t\right)+\mathbf{i}\left(t\right)+{\mathbf{i}}_{\mathrm{leak}}\left(t\right)+\mathbf{n}\left(t\right).\hfill \end{array}$$

The returns can have closely spaced delays and arrival/departure angles; their separability depends on bandwidth, aperture, SNR, and processing. A common Doppler across scattering centres requires an appropriate rigid-motion approximation; rotation or deformation may require a time-varying phase model beyond this constant-Doppler approximation.

The model assumes approximately constant delay and Doppler over the coherent processing interval (CPI). Acceleration, range migration, extended targets, diffuse scattering, wideband Doppler scaling, and frequency-dependent array responses require extensions [15,16]. Coherent processing further assumes that the sensing processor knows the realization of the transmitted reference waveform. For a co-located transmitter, this is normally available internally; for bistatic or distributed payload-aided sensing, the mechanism must be declared explicitly—for example, a shared pilot/reference signal, reliably decoded payload, signalling of the data realization, or a receiver that marginalizes unknown data symbols. Fundamental random-waveform ISAC regions can change when this knowledge assumption changes [18].

For wideband/OFDM operation the communication and sensing links should be represented on the same resource grid. For subcarrier m and OFDM symbol n,

$$y_{k,m,n}={\mathbf{h}}_{k,m,n}^H{\mathbf{x}}_{m,n}+z_{k,m,n}.$$

(4)

Under a sufficient-CP and small-Doppler approximation, a corresponding sensing baseline is

$${\mathbf{y}}_{m,n}^{\left(s\right)}={\displaystyle \sum _{q=1}^Q}{\alpha }_q{\mathbf{a}}_r\left({\theta }_q\right){\mathbf{a}}_t^H\left({\varphi }_q\right){\mathbf{x}}_{m,n}{e}^{-j2\pi m\Delta f{\tau }_q}{e}^{j2\pi {\nu }_{q}n{T}_{\mathrm{sym}}}+{\xi }_{m,n},$$

where $\Delta f$ is the subcarrier spacing and $T_{\mathrm{sym}}$ the OFDM symbol interval. The CP-OFDM model requires:

$${\tau }_{\max }<T_{\mathrm{CP}},{\left|\nu \right|}_{\max }\ll \Delta f.$$

For monostatic radial sensing, ${\tau }_{\max }=2D_{\max }/c$ and ${\left|\nu \right|}_{\max }=2v_{\max }/\lambda$, so that the corresponding sufficient conditions become

$$T_{\mathrm{CP}}>{\displaystyle \frac{2D_{\max }}{c}},\Delta f\gg {\displaystyle \frac{2v_{\max }}{\lambda }}.$$

Otherwise, ${\xi }_{m,n}$ must include the structured inter-symbol interference (ISI) and inter-carrier interference (ICI) [24,25]. These conditions define the validity of the simplified model rather than absolute sensing-range or velocity limits, since beyond-CP processing can explicitly model and mitigate the resulting interference. Wideband arrays may additionally require beam-squint or frequency-dependent manifolds.

4.2. Waveforms, Ambiguity, and Resolution

For a scalar complex baseband waveform $x\left(t\right)$, the unnormalised narrowband ambiguity function is

$${\chi }_x(\tau ,\nu )=\int x\left(t\right)x^{*}(t-\tau )e^{-j2\pi \nu t}dt.$$

(5)

If a normalized ambiguity function or sidelobe metric is reported, the normalization—for example by the waveform energy $E_x=\int {\left|x\left(t\right)\right|}^{2}dt$—must be stated. Equation (5) assumes narrowband Doppler-shift; high fractional bandwidth or wideband Doppler scaling requires an appropriate delay–scale/wideband ambiguity model.

First-order waveform resolution scales are more safely stated in delay and Doppler as

$$\Delta \tau \approx {\displaystyle \frac{1}{B}},\Delta \nu \approx {\displaystyle \frac{1}{T_{\mathrm{CPI}}}}.$$

For monostatic radial sensing, $\Delta R\approx c/\left(2B\right)$ and $\Delta v_r\approx \lambda /\left(2T_{\mathrm{CPI}}\right)$; bistatic/multistatic position and velocity resolution must instead be transformed through the appropriate geometry/Jacobian [24]. These relations describe classical resolution-cell scaling, not guaranteed estimator error, and they do not preclude model-based super-resolution under stronger sparsity, identifiability, and SNR assumptions. Unambiguous delay/range and Doppler/velocity also depend on sampling, subcarrier spacing, pulse repetition, CP/guard structure, frame duration, and receiver processing.

Table 3 compares waveform families under the matched resources listed at the start of Section 4. Sensing references may be deterministic pilots known to the receiver, known or decoded payload, or unknown payload requiring marginalisation or joint inference. Comparisons must identify the case and distinguish realised ACF/ambiguity, its mean, mean squared magnitude, variance, PSLR and ISLR, which describe different aspects of data-dependent responses [30].

Communication-centric chirp spread spectrum (CSS) provides a further candidate for sensing reuse. DM-CSS combines even/odd chirp components with phase and up/down-chirp choices; DM-TDM-CSS combines up- and down-chirped symbols and supports coherent and noncoherent communication detection [28,29]. These studies establish communication tradeoffs. Any sensing benefit still requires evaluation of the modulated waveform’s ambiguity, reference availability, and achieved detection or estimation performance under matched resources.

Synchronization requirements also depend on architecture. Co-located monostatic transmitters and receivers can share a clock and local oscillator, although residual timing, phase noise, and RF calibration errors remain. Separated bistatic nodes require a shared reference or explicit estimation of relative timing, carrier-frequency offset, and phase for coherent processing [31]. Noncoherent fusion relaxes carrier-phase alignment, but still requires adequate timestamps, geometry, and association.

OFDM remains attractive because of its cellular/WLAN integration, but CP sufficiency, phase noise, CFO, high-Doppler ICI, amplifier back-off, and payload-dependent sidelobes are not secondary implementation details [25,32,33]. Certain OTFS receivers can jointly model structured delay–Doppler interference instead of treating every ISI/ICI term purely as disturbance, but that benefit depends on pulse shape, fractional delay/Doppler, frame structure, and receiver design [34]. Ambiguity/PSLR/ISLR is itself a waveform/receiver surrogate: improved sidelobes do not establish improved $P_D$ or RMSE without a declared detector or estimator and operating point [33]. A change in the classical CRB must instead be evaluated from the likelihood, parameterization, nuisance treatment, and regularity assumptions; calculating that bound does not require selecting a particular estimator [21].

4.3. MIMO Beamforming and Optimization Guarantees

Let ${\mathbf{R}}_x=\mathbf{W}{\mathbf{W}}^H+{\mathbf{Q}}_s$ and let $\mathbf{A}\in {\mathbb{C}}^{G\times N_t}$ stack far-field transmit steering vectors as rows, so that the power toward grid direction g is $p_g={\mathbf{a}}_t^H\left({\theta }_g\right){\mathbf{R}}_x{\mathbf{a}}_t\left({\theta }_g\right)$. For a desired nonnegative power pattern ${\mathbf{p}}_d\in {\mathbb{R}}_{+}^G$, one illustrative fully digital baseline is

$$\begin{array}{cc}\hfill \underset{\mathbf{W},{\mathbf{Q}}_s⪰\mathbf{0}}{\min }& {∥diag\left(\mathbf{A}{\mathbf{R}}_x{\mathbf{A}}^H\right)-{\mathbf{p}}_d∥}_2^2,\hfill \end{array}$$

(6a)

$$\begin{array}{cc}\hfill \mathrm{s}.\mathrm{t}.& {\displaystyle \frac{|{\mathbf{h}}_k^H{\mathbf{w}}_k{|}^2}{{\sum }_{j\ne k}{\left|{\mathbf{h}}_k^H{\mathbf{w}}_j\right|}^2+{\mathbf{h}}_k^H{\mathbf{Q}}_s{\mathbf{h}}_k+{\sigma }_k^2}}\ge {\gamma }_k,\forall k,\hfill \end{array}$$

(6b)

$$\begin{array}{cc}\hfill & tr\left({\mathbf{R}}_x\right)\le P_{\max }.\hfill \end{array}$$

(6c)

Equation (6a) performs power-beampattern matching. The alternative full matrix objective $\parallel \mathbf{A}{\mathbf{R}}_x{\mathbf{A}}^H-{\mathbf{B}}_d{\parallel }_F^2$, with declared ${\mathbf{B}}_d={\mathbf{B}}_d^H⪰\mathbf{0}$, performs spatial-covariance matching, including cross-direction correlations. Here $\mathbf{A}$ is an angular grid; an angle–range grid is valid only when a range-dependent near-field or wideband array manifold is explicitly introduced.

Equation (6b) assumes known ${\mathbf{h}}_k$ and single-user detection with uncancelled sensing interference represented by second-order power. Cancellation, a decoded common stream, or joint detection changes the constraint. Denoting ${\log }_2(1+\mathrm{SINR})$ capacity requires justified receiver/distribution assumptions, particularly for deterministic or non-Gaussian sensing interference. Robust/chance-constrained variants must expose CSI uncertainty or posterior.

Beam-pattern, CRB, SCNR, detection, and waveform-ambiguity control objectives have different target, clutter, nuisance, and receiver assumptions [21,35]. SCNR/detection designs must specify the receiver combiner/filter and clutter/interference covariance.

Implementation may also require per-antenna power, hybrid/analog or phase-only structure, constant-envelope/modulus operation, spectral-mask/EVM, steering/target uncertainty, RF-chain and hardware constraints [33,36]. Finite-sample or RF-chain implementation of ${\mathbf{Q}}_s$ also requires a feasible rank, waveform length, and modulus/quantisation. Semidefinite relaxation (SDR) provides a relaxation bound and needs rank recovery unless tightness is proved; SCA/AO typically offer conditional stationarity guarantees; manifold optimization preserves non-Euclidean constraints but is generally local; and exact mixed-integer branch-and-bound can achieve global optimality at prohibitive cost. Guarantees are formulation-specific [35]. Report runtime with problem dimension, hardware and update interval, distinguishing offline waveform/codebook design from per-slot beam updates.

4.4. Estimation, Detection, Localization, and Tracking

Let $\mathbf{y}\in {\mathbb{C}}^{N_y}$ denote samples, processed observations, or sufficient statistics for the declared likelihood where available. The real parameter vector, $\eta \in {\mathbb{R}}^d$ contains the real-valued unknown parameters of the sensing problem, including, as required, delay, Doppler, angles, position, velocity, and the real/imaginary target-response components as required. Complex formulations require equivalent real augmentation or a consistent complex/Wirtinger Fisher information matrix (FIM).

For a declared conditional density $p(\mathbf{y}\mid \eta )$, the FIM is

$${\mathbf{J}}_{\eta }={\mathbb{E}}_{\mathbf{y}\mid \eta }\left[{\nabla }_{\eta }\log p(\mathbf{y}\mid \eta ){\nabla }_{\eta }^T\log p(\mathbf{y}\mid \eta )\right].$$

(7)

Under the classical regularity conditions, if $\widehat{\eta }$ is unbiased and ${\mathbf{J}}_{\eta }$ is nonsingular, $Cov\left(\widehat{\eta }\right)⪰{\mathbf{J}}_{\eta }^{-1}$. If $\eta ={[{\theta }^T,{\xi }^T]}^T$ contains parameters of interest $\theta$ and nuisance parameters $\xi$, then, when the required block is invertible, the equivalent FIM is

$${\mathbf{J}}_{\mathrm{eff}}={\mathbf{J}}_{\theta \theta }-{\mathbf{J}}_{\theta \xi }{\mathbf{J}}_{\xi \xi }^{-1}{\mathbf{J}}_{\xi \theta },$$

rather than simply selecting and inverting the parameter-of-interest block of the full FIM. A singular FIM requires an identifiable subspace or reparametrization; a Moore–Penrose inverse is not automatically a bound for non-identifiable parameters. Classical, Bayesian, constrained, biased, posterior, and position-domain bounds have distinct assumptions [6,21]. Global Bayesian bounds can be more informative under low SNR, ambiguity, or threshold effects.

For binary target presence, a generic observation model is

$${\mathcal{H}}_0:\mathbf{y}=\mathbf{v},{\mathcal{H}}_1:\mathbf{y}=\mathbf{s}\left(\eta \right)+\mathbf{v}.$$

(8)

When the observation distributions under ${\mathcal{H}}_0$ and ${\mathcal{H}}_1$ are fully specified, the Neyman–Pearson criterion maximizes the probability of detection $P_D$ subject to a prescribed upper bound on the false-alarm probability $P_{FA}$. Composite hypotheses require the GLRT, CFAR, Bayesian, or other detector, and its nuisance/background-estimation procedure. Deterministic and fluctuating models produce different ROC behaviors. Multi-target evaluation must specify whether association is known, resolved jointly, or uncertain, and account for resolution, track initiation/deletion, and false/missed detections.

Localization transforms delay, time-difference, angle, Doppler, or power measurements through geometry. The interface preserves measurement covariance/ posterior, coordinate frame, TX/RX geometry, synchronization residuals, and the Jacobian or position-domain information. Multistatic range and range-rate constraints require the appropriate geometry transformation [24].

Tracking additionally requires a dynamic state model, update/prediction interval, missed-detection process, and information age. Posterior/PCRB or achieved filter covariance must match that model and interval; the exported state carries the timestamp and validity/prediction horizon. Consequently, one-shot CRB or $P_D$ alone does not establish tracking performance under mobility, uncertain association, mismatch, delay, or calibration drift.

4.5. Fundamental Tradeoffs and Conditions for Meaningful Joint Comparison

An ISAC tradeoff is defined by its channel/state law, coding and sensing strategy, waveform/reference knowledge, receiver, resource, admissible designs, and sensing criterion [6,18]. A numerical Pareto set from a selected design family is achievable within that family; a fundamental-region claim requires achievability/converse analysis. Rate must be specified as capacity, achievable rate, coded spectral efficiency, goodput, or BLER-constrained finite-blocklength service rate. Finite-blocklength effects are relevant to short-deadline services [19,37]. Achievable tradeoffs can be explored by sweeping weights in a normalized weighted-sum objective, or by optimizing one metric while varying a constraint on the other. For example, Bazzi and Chafii combine communication and sensing mutual information through weighted-sum scalarization and optimize the pilot matrix using projected gradients on the Stiefel manifold [38]. This does not turn the problem into a linear program. Weighted sums can miss unsupported Pareto points in nonconvex objective regions; varying an $ϵ$-constraint provides an alternative. With local solvers, the nondominated solutions found describe an achievable approximation, not a certified complete frontier. Table 4 applies the metric discipline of Section 3 to specific sensing-communication tradeoffs.

4.6. Channel, Hardware, Synchronization, and PHY Export Interface

Models that exploit communication–sensing correlation or shared environmental structure must represent communication and target paths, shared scatterers, clutter, blockage, and interference [15,39]. TDD communication reciprocity does not imply sensing path reciprocity, common scatterers, or stationarity. Ray-tracing, stochastic, and hybrid models require validating the statistics that support the claimed correlation.

Phase noise and CFO affect delay–Doppler responses; IQ imbalance, gain/phase error, coupling, and array calibration affect amplitude/phase consistency and angles. Converter quantization, limited dynamic range, saturation, and residual self-interference can produce distortion or detection floors; PA nonlinearity affects spectrum and ambiguity; and clock drift changes the time/phase reference across observations [32,36,40].

Simultaneous monostatic transmission/reception requires declared isolation, analog/digital cancellation, ADC headroom/saturation, and residual-interference statistics; time-separated or pulsed operation has different requirements. Distributed coherent sensing must report its synchronization architecture and residual timing offset (TO), carrier-frequency offset (CFO) and phase error [31]. Noncoherent fusion relaxes carrier-phase coherence but still requires timing, geometry, calibration, and association.

Table 5 specifies the observation, uncertainty, timing, calibration, validity, and provenance information required by downstream fusion and control.

• Cross-layer implication.

Physical-layer results support downstream fusion or control within the assumptions and metadata specified in Table 5. Calibration bias and statistical estimator uncertainty remain distinct; an exported bound must be labelled and its downstream use justified.

5. Networked and Cooperative ISAC

Networked ISAC shares observations and resources among base stations, access points, devices and edge nodes. It can improve coverage, positioning and diversity, subject to synchronization, association, freshness, fronthaul and fusion constraints [4,15]. Table 6 compares architectures; Table 7 follows the service from request and authorization through sensing-unit selection, radio allocation, processing and result exposure. This links cooperative processing and mobility control to requester identity, area/target coverage, uncertainty, freshness, privacy and access requirements.

5.1. Architecture Axes Rather Than Overlapping Labels

“Multi-cell,” “cooperative,” “networked,” and “cell-free” are not mutually exclusive categories. A more robust taxonomy uses axes: administrative ownership, serving-set formation, control centralization and programmability, coherence, exchanged representation, and fusion level. Table 6 applies these axes to common architectures [41,42,43,44].

Table 6. Network architecture axes that avoid ambiguous labels.

Example	Serving set	Control	Coherence	Exchange/fusion	Dominant bottleneck
[l]Single-cell
multi-user	One BS/AP	Centralized	Local	Local sufficient statistics or tracks	User–target interference and multiplexing.
[l]Coordinated
multi-cell	Cell association, optional CoMP	Inter-cell	Noncoherent or coherent	Estimates, features or raw samples	Inter-cell interference, signalling and synchronization.
[l]Task-specific
cooperative	Selected nodes by geometry/task	Centralized or distributed	Mixed	Raw, feature, measurement, or track data	Node selection, calibration, fronthaul and data association.
[l]Cell-free,
user-centric	Dynamic AP cluster	Centralized or distributed	Often time/phase aligned	Distributed processing and fusion	Fronthaul, scalable calibration and coherent processing.
[l]Heterogeneous
network	[l]Terrestrial,
aerial and satellite,
and edge nodes	Hierarchical	Mixed	Hierarchical beliefs/tracks	Association, orchestration, intermittent visibility and trust.

Table 6. Network architecture axes that avoid ambiguous labels.

Example	Serving set	Control	Coherence	Exchange/fusion	Dominant bottleneck
[l]Single-cell
multi-user	One BS/AP	Centralized	Local	Local sufficient statistics or tracks	User–target interference and multiplexing.
[l]Coordinated
multi-cell	Cell association, optional CoMP	Inter-cell	Noncoherent or coherent	Estimates, features or raw samples	Inter-cell interference, signalling and synchronization.
[l]Task-specific
cooperative	Selected nodes by geometry/task	Centralized or distributed	Mixed	Raw, feature, measurement, or track data	Node selection, calibration, fronthaul and data association.
[l]Cell-free,
user-centric	Dynamic AP cluster	Centralized or distributed	Often time/phase aligned	Distributed processing and fusion	Fronthaul, scalable calibration and coherent processing.
[l]Heterogeneous
network	[l]Terrestrial,
aerial and satellite,
and edge nodes	Hierarchical	Mixed	Hierarchical beliefs/tracks	Association, orchestration, intermittent visibility and trust.

5.2. Sensing Service Lifecycle, Control Functions, and Exposure

A sensing request from an application, vertical or network function identifies the area, object class or event, observation interval, desired output (detection, location, track, map or feature), target KPI, deadline or update period, and privacy or authorization class. Before radio optimisation, the network verifies requester entitlement and permission to sense the affected users, devices, and location.

The control function selects transmitting/receiving units by coverage, geometry, waveform capability, synchronization, load and trust. Scheduling combines sensing windows with communication traffic, using an end-to-end slice where isolation or differentiated QoS is required. Processing is placed in the RAN, Multi-access edge computing (MEC), or the core, depending on data volume, fusion, latency, and privacy [45]. Results carry the estimate/event, timestamp, quality indicator and application-required uncertainty or validity. Operator APIs may expose them; the Network Exposure Function (NEF) and Common API Framework (CAPIF) provide 5G precedents, but the cited ISAC specifications do not define a universal application-facing schema carrying uncertainty, provenance and validity.

Document status and scope are consolidated in Section 8. Here, Table 7 maps those standards anchors to service functions and identifies the additional requirements for uncertainty-bearing results, consent, retention and cross-domain exposure.

Table 7. Sensing-service lifecycle and the interface gaps exposed by the operational contracts.

Lifecycle stage	Principal contracts	Standards anchor	Open interface
Request and authorization	C1, C3, C6	ITU-R M.2160; 3GPP TR 22.837/TS 22.137; TR 23.700-15/TS 23.138 [1,46,47,48,49]	Common request schema for area, output, deadline, uncertainty, purpose, and privacy class.
Sensing-unit selection	C3, C4, C6	3GPP TR 23.700-14/TS 23.137; ETSI GR ISC 003 [50,51,52]	Geometry- and trust-aware selection with explicit synchronization and calibration feasibility.
Radio scheduling and sensing window	C3–C6	3GPP TR 38.765; IEEE 802.11bf [17,53]	Joint communication/sensing scheduling whose overhead and information age are visible to the requester.
Processing and fusion	C1–C5	3GPP TR 23.700-14/TS 23.137; ETSI GR ISC 002/003 [39,50,51,52]	Portable representation of covariance, cross-node dependence, timestamp, and model validity.
Result exposure	C1, C2, C5	3GPP TR 23.700-15/TS 23.138; IEEE 802.11bf [48,49,53]	Application-facing result object carrying uncertainty, provenance, and validity without exposing unnecessary raw data.
Retention, audit, and revocation	C3, C5, C6	ETSI GR ISC 004; 3GPP vertical-application work [48,49,54]	Enforceable purpose limitation, retention, deletion, revocation, and audit across RAN, edge, and application domains.

Table 7. Sensing-service lifecycle and the interface gaps exposed by the operational contracts.

Lifecycle stage	Principal contracts	Standards anchor	Open interface
Request and authorization	C1, C3, C6	ITU-R M.2160; 3GPP TR 22.837/TS 22.137; TR 23.700-15/TS 23.138 [1,46,47,48,49]	Common request schema for area, output, deadline, uncertainty, purpose, and privacy class.
Sensing-unit selection	C3, C4, C6	3GPP TR 23.700-14/TS 23.137; ETSI GR ISC 003 [50,51,52]	Geometry- and trust-aware selection with explicit synchronization and calibration feasibility.
Radio scheduling and sensing window	C3–C6	3GPP TR 38.765; IEEE 802.11bf [17,53]	Joint communication/sensing scheduling whose overhead and information age are visible to the requester.
Processing and fusion	C1–C5	3GPP TR 23.700-14/TS 23.137; ETSI GR ISC 002/003 [39,50,51,52]	Portable representation of covariance, cross-node dependence, timestamp, and model validity.
Result exposure	C1, C2, C5	3GPP TR 23.700-15/TS 23.138; IEEE 802.11bf [48,49,53]	Application-facing result object carrying uncertainty, provenance, and validity without exposing unnecessary raw data.
Retention, audit, and revocation	C3, C5, C6	ETSI GR ISC 004; 3GPP vertical-application work [48,49,54]	Enforceable purpose limitation, retention, deletion, revocation, and audit across RAN, edge, and application domains.

5.3. Resource, Interference, and Geometry Coupling

Let $\mathbf{a}$ denote discrete user/target/node/slot associations, $\mathbf{p}$ and $\mathbf{b}$ power and bandwidth, and $\mathbf{W}$ beamformers. A representative network problem is

$$\begin{array}{cc}\hfill \underset{\mathbf{a},\mathbf{p},\mathbf{b},\mathbf{W}}{\max }& \sum _k{u}_k\left(R_k\right)-\lambda \sum _q{D}_q,\hfill \end{array}$$

(9a)

$$\begin{array}{cc}\hfill \mathrm{s}.\mathrm{t}.& R_k\ge R_k^{\min },D_q\le D_q^{\max },\hfill \end{array}$$

(9b)

$$\begin{array}{cc}\hfill & \sum _k{p}_k\le P_{\max },\sum _k{b}_k\le B_{\max },\mathbf{a}\in \mathcal{A}.\hfill \end{array}$$

(9c)

The term $D_q$ must be a declared achieved distortion/risk or a surrogate used as a constraint. In coordinated multi-cell/CoMP operation, association affects not only load but also sensing geometry, observability, interference, synchronization, and fronthaul. Thus, a user-centric utility can create target-centric externalities: one transmission may illuminate several targets or interfere with several sensing receivers [43,55].

Solution classes include matching or mixed-integer optimization for association, SDR/SCA/AO for beam and power allocation, sparse or greedy cooperative-node selection, and Lyapunov or dynamic programming for time-varying queues. Their cross-layer significance depends on scaling, optimality guarantees, required global state, and whether synchronization and sharing overhead are charged to the objective.

5.4. Sensing-Assisted Communication and Mobility

Sensing can reduce beam training, predict blockage, support localization-assisted handover, forecast channel evolution, and enable proactive scheduling. These gains depend on the validity interval and uncertainty of position, velocity, blockage and scatterer estimates. A mobility controller should therefore receive a belief or prediction distribution, not only a point trajectory. Robust handover and beam selection should report outage or regret under prediction error, update delay and map mismatch rather than only ideal-state overhead reduction.

5.5. Observation Sharing and Fusion Interfaces

Raw I/Q sharing preserves maximum information but incurs high fronthaul and stringent synchronization. Range–Doppler maps or features reduce traffic but introduce algorithmic coupling and possible privacy leakage. Measurements and tracks are compact but can lose non-Gaussian or multi-modal uncertainty. Cooperative fusion must state cross-covariance handling, data association, information age, clock/carrier residuals, provenance/trust state, and whether fusion is coherent, noncoherent, feature-level, or track-level [31,54,56]. These interface variables connect directly to edge placement in Section 6.

• Cross-layer implication.

Serving-set and handover evaluations must account for information age, dependence and validity together with the synchronization, fronthaul and privacy costs of the chosen sharing representation.

6. Intelligence, Computation, and Task Execution

The lifecycle includes model-driven and constrained learning, reinforcement and federated/ distributed learning, representation selection, split inference and edge orchestration [57,58]. Evaluation covers confidence, feasibility, distribution shift, constraint violations, queues, memory, tail latency and fallback [59,60]. Task-oriented claims require a declared downstream loss or control objective; feature matching or classification accuracy alone does not establish application benefit [14,61].

6.1. A Single Lifecycle with Distinct Decisions

Intelligence selects the estimator or policy, its adaptation and the resulting action or representation, with calibrated confidence, validity, feasibility and fallback information where applicable. Computation selects the representation, model variant, placement, split and resource allocation, and reports execution status, provenance, latency, energy and memory. The task sets sensing resolution, update rate, the operational loss or control objective, and associated safety constraints. These decisions are coupled: model size, activations, and update frequency determine radio–compute demand, while processor, memory, and deadline budgets constrain the models and actions that can be executed [9,10,58].

6.2. Constrained and Model-Driven Learning

Let $\mathbf{z}$ contain observations, uncertainty, network/queue state, and task feedback, and let ${\pi }_{\theta }\left(\mathbf{z}\right)$ output a radio or control action. Because communication, sensing, and task losses can have different units or scales, let ${\tilde{L}}_c$, ${\tilde{L}}_s$, and ${\tilde{L}}_{\mathrm{task}}$ denote dimensionless losses under a declared normalisation or utility-cost mapping. A generic hard-constrained learning problem is

$$\begin{array}{cc}\hfill \underset{\theta }{\min }& \mathbb{E}\left[{\lambda }_c{\tilde{L}}_c+{\lambda }_s{\tilde{L}}_s+{\lambda }_t{\tilde{L}}_{\mathrm{task}}\right],{\lambda }_c,{\lambda }_s,{\lambda }_t\ge 0,\hfill \end{array}$$

(10a)

$$\begin{array}{cc}\hfill \mathrm{s}.\mathrm{t}.& g_m({\pi }_{\theta }\left(\mathbf{z}\right);\mathbf{z})\le 0\mathrm{a}.\mathrm{s}.,m=1,\dots ,M.\hfill \end{array}$$

(10b)

The weights express application priorities after the loss normalization has been fixed. They should not all be zero; setting ${\lambda }_c+{\lambda }_s+{\lambda }_t=1$ is a convenient convention. Candidate weights can be swept on validation data and selected using the required communication, sensing, and task performance, with the test set reserved for final evaluation. Increasing one weight gives its normalized loss greater emphasis, but does not guarantee a corresponding improvement after nonconvex training. Report the resulting component losses, constraint-violation rates, and sensitivity to the weights. The hard constraints remain independent of this tradeoff choice.

The almost-sure qualifier states the semantics of the hard constraints; expected or chance constraints require their corresponding explicit formulations. Penalty training does not certify hard power, spectral-mask, constant-modulus, discrete-association, or minimum-SINR constraints. Exact convex projection can preserve a convex feasible set, but projection onto a nonconvex or mixed-integer set may be expensive, multivalued, or nondifferentiable. Depending on the constraint structure, practical designs may therefore use constraint-preserving parameterisation, differentiable optimisation layers, approximate feasibility recovery, runtime monitors, or model-based fallback [62,63].

Model-driven or unrolled networks retain a forward model and learn uncertain components, regularisers, or iteration parameters. Their evaluation must expose sensitivity to forward-model mismatch, hardware impairments, and truncation depth.

For sequential problems formulated as a Markov decision process (MDP) or a partially observable Markov decision process (POMDP), the state/belief (e.g., channel, target covariance, queues, geometry), observation, action (beam, schedule, trajectory, sensing mode), transition law, reward, and safety costs should be defined explicitly. A weighted reward is not the same as imposing hard or probabilistic constraints; Lagrangian primal–dual training, shielded actions, risk-sensitive objectives, and safe exploration should therefore be distinguished. Multi-agent work should specify centralised training versus decentralised execution and address partial observability and agent-induced nonstationarity.

A multimodal vehicular ISAC preprint couples semantic information age, edge-computing queues, tracking PCRB, sensor activation and beamforming under a long-term energy budget through Lyapunov-guided learning [64]. It explicitly accounts for freshness and compute congestion. We treat its simulation-only evaluation as E1, with no hardware or testbed validation.

Federated and collaborative learning can reduce raw-data movement, but evaluations should report non-IID data, client selection, dropout, communicated bits, local compute, and radio-dependent aggregation. When differential privacy is claimed, the privacy mechanism and its accountant or budget should also be reported. ISAC-for-AI creates a further closed loop because the sensing policy changes the amount, quality, and bias of training data [65].

Foundation models are an emerging 2025–2026 direction rather than an established ISAC solution. Their potential lies in reusable multimodal representations, few-shot adaptation, and agentic orchestration; unresolved issues include physical consistency, data provenance, compute/energy cost, latency, privacy, calibration, and cross-site generalisation [66]. They should be benchmarked against task-specific and model-driven baselines rather than assumed superior by scale.

6.3. Placement, Representation, and End-to-End Latency

For task i, let $r_i$ denote the transmitted representation (raw I/Q, CSI, map, feature, belief, track, logit, or action), $m_i$ a model variant, and $x_{i,n}\in \{0,1\}$ the placement at node n, with ${\sum }_n{x}_{i,n}=1$. The end-to-end latency is

$$\begin{array}{cc}\hfill T_i^{\mathrm{e}2\mathrm{e}}=& T_i^{\mathrm{sen}}+T_i^{\mathrm{enc}}\hfill \\ \hfill & +\sum _n{x}_{i,n}\left(T_{i,n}^{\mathrm{ul}}+T_{i,n}^q\right)\hfill \\ \hfill & +\sum _n{x}_{i,n}\left(T_{i,n}^{\mathrm{cmp}}+T_{i,n}^{\mathrm{dl}}\right).\hfill \end{array}$$

(11)

Equation (11) is a per-task latency decomposition. Its radio-service and queueing terms may be random, so deadline reliability requires their joint distribution. For local execution, the corresponding transport terms are defined as $T_{i,n}^{\mathrm{ul}}=T_{i,n}^{\mathrm{dl}}=0$. The uplink term uses an achieved service rate that accounts for coding, packetisation, control signalling, block errors, retransmissions, and finite-blocklength reliability. Shannon capacity is not treated as a transport rate. For representation size $S_i^{\left(r\right)}$,

$$T_{i,n}^{\mathrm{ul}}={\displaystyle \frac{S_i^{\left(r\right)}}{R_{i,n}^{\mathrm{svc}}}}.$$

(12)

For an assigned task, CPU allocation $f_{i,n}>0$ and accelerator throughput ${\rho }_{i,n}>0$ give

$$T_{i,n}^{\mathrm{cpu}}={\displaystyle \frac{C_i^{(r,m)}}{f_{i,n}}},T_{i,n}^{\mathrm{acc}}={\displaystyle \frac{F_i^{(r,m)}}{{\rho }_{i,n}}},x_{i,n}=1,$$

(13)

with aggregate node-capacity constraints

$$\sum _i{x}_{i,n}{f}_{i,n}\le f_{n,\max },\sum _i{x}_{i,n}{\rho }_{i,n}\le {\rho }_{n,\max }.$$

(14)

Unselected-node ratios are not evaluated. The construction of $T_{i,n}^{\mathrm{cmp}}$ must state whether the execution graph uses CPU only, accelerator only, serial CPU–accelerator stages, or parallel stages; for example, serial stages add their execution times whereas perfectly parallel stages are bounded by their maximum. Memory includes parameters, inputs, activations, and workspace. Real deployments can also include model loading/caching, memory bandwidth, batching, accelerator contention, and thermal throttling. For deadline- or safety-critical tasks, mean queue stability alone is insufficient; tail-latency or deadline-violation metrics should also be reported [58,59,67].

Representation selection trades radio load against local compute, privacy, and task loss. Early split inference can transmit a large activation tensor; late splitting reduces communication but increases device cycles, memory, and energy. Evaluate compression by downstream loss or decision risk under the same radio and processor realisation, not just feature reconstruction error.

6.4. Cross-Layer Intelligence–Compute Optimization

Let $\mathbf{d}=(\mathbf{v},\mathbf{c},\mathbf{m})$ collect physical/network, computation and intelligence decisions. For a fixed decision, let $\omega$ represent the task state and random service and queueing conditions under a declared joint model. Define

$${\overline{T}}_i\left(\mathbf{d}\right)={\mathbb{E}}_{\omega }\left[T_i^{\mathrm{e}2\mathrm{e}}(\mathbf{d},\omega )\right],{\overline{E}}_i\left(\mathbf{d}\right)={\mathbb{E}}_{\omega }\left[E_i^{\mathrm{tot}}(\mathbf{d},\omega )\right],$$

and the deadline-violation probability

$$P_i^{\mathrm{viol}}\left(\mathbf{d}\right)={Pr}_{\omega }\left\{T_i^{\mathrm{e}2\mathrm{e}}(\mathbf{d},\omega )>L_i\right\}.$$

For tasks with finite expected latency and energy, a representative formulation is

$$\begin{array}{cc}\hfill \underset{\mathbf{d}}{\min }& \sum _i\left[{\alpha }_i{D}_i^{\mathrm{task}}+{\beta }_i{\overline{T}}_i+{\gamma }_i{\overline{E}}_i+{\delta }_i{P}_i^{\mathrm{viol}}\right],\hfill \end{array}$$

(15a)

$$\begin{array}{cc}\hfill \mathrm{s}.\mathrm{t}.& R_i\left(\mathbf{v}\right)\ge R_i^{\min },D_i^{\mathrm{sens}}\left(\mathbf{v}\right)\le D_{i,\max },\hfill \end{array}$$

(15b)

$$\begin{array}{cc}\hfill & P_i^{\mathrm{viol}}\left(\mathbf{d}\right)\le {ϵ}_i,0\le {ϵ}_i<1,\hfill \end{array}$$

(15c)

$$\begin{array}{cc}\hfill & \mathbf{d}\in \mathcal{A}.\hfill \end{array}$$

(15d)

Dependence on $\mathbf{d}$ is suppressed in the objective. The nonnegative weights ${\alpha }_i$, ${\beta }_i$, ${\gamma }_i$ and ${\delta }_i$ map their respective quantities to a common declared cost unit and are fixed across comparisons. The rate and sensing constraints use the deterministic performance measures of the declared physical-layer model. The task loss in Equation (17) and the execution statistics refer to the same task population and operating law. Here, $\mathcal{A}$ contains the applicable processor, memory, queue, privacy, security and standard constraints. The deadline constraint is probabilistic; setting ${ϵ}_i=0$ imposes it almost surely under the declared law.

Depending on the selected variables and dynamics, an instantiated problem may generally be mixed discrete–continuous, nonconvex, and/or time coupled. Evaluation should therefore be method-appropriate: feasibility or empirical violation should be reported for all methods where relevant, convergence or optimality information where applicable, sample complexity/generalisation for learned policies, and online runtime for the deployed decision method.

Let ${\xi }_i$ denote the task state, and define the full-information oracle over the same feasible action set ${\mathcal{U}}_i\left({\xi }_i\right)$ as

$$a_i^{★}\left({\xi }_i\right)\in \arg \underset{a\in {\mathcal{U}}_i\left({\xi }_i\right)}{\min }{\ell }_i(a,{\xi }_i).$$

(16)

For an executed action ${\widehat{a}}_i$, define the excess task loss

$$D_i^{\mathrm{task}}=\mathbb{E}\left[{\ell }_i({\widehat{a}}_i,{\xi }_i)-{\ell }_i(a_i^{★}\left({\xi }_i\right),{\xi }_i)\right].$$

(17)

This quantity is task- and distribution-specific; it is not a universal semantic metric. Lower excess task loss means that the executed decision approaches the declared full-information oracle benchmark [14,61]. The full-information oracle is an evaluation benchmark, not a required real-time component. Record deployed actions, task outcomes, and timing without running the benchmark on the critical path. Where ground-truth task states are available, evaluate the oracle offline on those states using the same feasible action sets, and report its solver, tolerance, runtime, and any unresolved optimality gap separately. This prevents benchmark computation from adding contention or deadline violations to the measured execution. If only a feasible approximate solution is available, identify it as a reference policy rather than an exact oracle. Any such computation used online must be included in the end-to-end latency.

6.5. Evaluation Baselines

Compare with applicable model-based solvers, complexity-matched black-box and model-driven networks, fixed representations/splits, and local-only and edge-only execution. Sequential policies also require fixed, myopic, or model-predictive baselines. Match, or explicitly normalise, pilots, bandwidth, power, sensing and task models, processor resources and deadlines. Separate offline training/data generation from online runtime; report parameter/activation memory, mean/tail latency, deadline and constraint violations, energy and relevant queue behaviour. Stress tests should cover distribution or site shift, hardware/model mismatch, non-IID data, privacy, confidence calibration, and edge-link failures when material to the claim.

E3 identifies offline measurement evidence; it does not impose a universal minimum geometric diversity or establish generalization. For a claimed deployment domain, report the sampled transmitter/receiver placements, baselines, target ranges, angles, orientations and trajectories, together with environments, blockage conditions, calibration, ground truth, and independent trials per condition. Coverage should be justified against task observability and the intended operating domain. Generalization tests should hold out relevant sites, geometries, trajectories, or users, rather than only randomly separating correlated frames from the same recording, and should report performance and uncertainty for the held-out conditions.

• Cross-layer implication.

End-to-end advantage requires feasible, timely, calibrated and recoverable actions on the deployed processor and network. Evaluation must include relevant activation memory, queueing, model-loading delay, constraint violations and fallback behaviour alongside prediction accuracy.

7. Propagation, Hardware, Security, and Trust

Propagation defines signal-model validity; hardware residuals constrain coherent processing and implementation; and security/privacy govern authorised, trustworthy use of sensing information and actions [4,15,39,54]. Near-field curvature requires spherical-wave models or validated approximations; spatial nonstationarity requires separate aperture/environment evidence [68]. RIS, FR3, optical, aerial and non-terrestrial regimes change propagation, calibration, synchronization, blockage, visibility and resources [69,70,71,72,73,74]. Converter limitations, phase noise, oscillator drift, array mismatch, PA distortion, residual self-interference and calibration ageing can bias estimates or invalidate coherence [36,40].

7.1. Propagation and Platform Regimes as Validity Conditions

Emerging bands and platforms matter because they change the validity range of earlier models. Table 8 concentrates the principal changes and the information that should accompany reported gains. In this survey, FR3 denotes approximately the 7–24 GHz upper-mid-band region. It changes the engineering balance among coverage, electrical aperture, available bandwidth, blockage, near-field behaviour, wideband beam effects, coexistence, and sensing-resource overhead; realised coverage and sensing resolution remain dependent on allocation, aperture, bandwidth, blockage, environment, and link budget [70,71]. Coverage should be defined by task-specific success events. Let ${\mathcal{E}}_c$ require an achieved communication service rate or SINR threshold with a declared BLER requirement, and let ${\mathcal{E}}_s$ require, for example, a localization-error tolerance and a result deadline, counting missing results as failures. Under a declared joint distribution of user/target positions, channels, blockage, and target responses, report

$$P_c^{\mathrm{cov}}=Pr\left({\mathcal{E}}_c\right),P_s^{\mathrm{cov}}=Pr\left({\mathcal{E}}_s\right),P_{\mathrm{joint}}^{\mathrm{cov}}=Pr({\mathcal{E}}_c\cap {\mathcal{E}}_s).$$

For detection, instead specify coverage through a required $P_D$ at fixed $P_{FA}$ and a declared target/RCS model. FR3 comparisons should match, or explicitly normalize, bandwidth, radiated power, aperture, observation time, pilot/payload overhead, and beam-squint compensation [71]. Range or area coverage is meaningful only with the corresponding geometry and success thresholds stated.

With fixed analog phase-shifter settings, intra-band beam squint denotes frequency-dependent pointing within the occupied band; inter-band squint concerns misalignment when the same analog weights are reused across separated carriers [71]. Report the frequencies, fractional bandwidth, aperture, beamformer architecture, and compensation method, since frequency-dependent steering or true-time-delay processing changes these losses.

Table 8. Propagation and platform regimes treated as model-validity conditions rather than separate catalogues.

Regime	Model change	Cross-layer consequence	Validation information to report
Near-field/XL-MIMO [68]	Spherical wavefronts and range–angle coupling; possible visibility regions/spatial nonstationarity. Beam squint and coupling are separate array effects.	Serving-set, beam, and generalisation decisions can depend on range, position, and visible subarrays rather than angle alone.	Aperture, frequency, bandwidth, calibration, model regime, and near-/far-field comparison under matched resources.
FR3 [70,71]	Band-dependent path loss, penetration, angular/delay spreads, wideband arrays, coexistence, near-field effects, and payload-aided sensing.	Coverage-oriented deployments must expose band-dependent visibility, intra-band and inter-band beam squint, channel-model validity, and sensing-resource overhead.	Cross-band measurements or calibrated simulators with equal-resource baselines and documented incumbent/coexistence assumptions.
RIS/metasurface [69]	Cascaded/direct/target paths, insertion loss, finite phase/amplitude resolution, coupling, and calibration; amplifier noise/power apply to active RIS.	Control overhead and cascaded uncertainty can erase ideal phase-only gains.	No-RIS/relay baselines with matched aperture, radiated and surface power, training, update rate, and calibration.
Optical O-ISAC [72]	Intensity/non-negativity constraints, field of view, orientation, ambient light, saturation, occlusion, and illumination/beam constraints.	Optical sensing and communication decisions depend on visibility, orientation, illumination, and receiver dynamic range.	Optical communication/sensing KPIs under declared illumination, orientation, blockage, safety, and energy conditions.
UAV/low-altitude [73]	Time-varying 3-D geometry, UAV/target motion, Doppler, and navigation uncertainty.	Trajectory, sensing updates, radio allocation, and tracking validity become horizon-coupled.	Navigation error, synchronized ground truth, trajectory assumptions, and a fixed-platform baseline.
NTN/SAGIN [74]	Long delay, high Doppler, ephemeris-driven geometry, intermittent visibility, feeder links, and segment-dependent synchronization.	Fusion, feedback, placement, and freshness depend on the active space, air, or ground segment.	Clock/ephemeris error, visibility windows, gateway capacity, link budget, and terrestrial or unsynchronised/noncoherent baselines as appropriate.

Table 8. Propagation and platform regimes treated as model-validity conditions rather than separate catalogues.

Regime	Model change	Cross-layer consequence	Validation information to report
Near-field/XL-MIMO [68]	Spherical wavefronts and range–angle coupling; possible visibility regions/spatial nonstationarity. Beam squint and coupling are separate array effects.	Serving-set, beam, and generalisation decisions can depend on range, position, and visible subarrays rather than angle alone.	Aperture, frequency, bandwidth, calibration, model regime, and near-/far-field comparison under matched resources.
FR3 [70,71]	Band-dependent path loss, penetration, angular/delay spreads, wideband arrays, coexistence, near-field effects, and payload-aided sensing.	Coverage-oriented deployments must expose band-dependent visibility, intra-band and inter-band beam squint, channel-model validity, and sensing-resource overhead.	Cross-band measurements or calibrated simulators with equal-resource baselines and documented incumbent/coexistence assumptions.
RIS/metasurface [69]	Cascaded/direct/target paths, insertion loss, finite phase/amplitude resolution, coupling, and calibration; amplifier noise/power apply to active RIS.	Control overhead and cascaded uncertainty can erase ideal phase-only gains.	No-RIS/relay baselines with matched aperture, radiated and surface power, training, update rate, and calibration.
Optical O-ISAC [72]	Intensity/non-negativity constraints, field of view, orientation, ambient light, saturation, occlusion, and illumination/beam constraints.	Optical sensing and communication decisions depend on visibility, orientation, illumination, and receiver dynamic range.	Optical communication/sensing KPIs under declared illumination, orientation, blockage, safety, and energy conditions.
UAV/low-altitude [73]	Time-varying 3-D geometry, UAV/target motion, Doppler, and navigation uncertainty.	Trajectory, sensing updates, radio allocation, and tracking validity become horizon-coupled.	Navigation error, synchronized ground truth, trajectory assumptions, and a fixed-platform baseline.
NTN/SAGIN [74]	Long delay, high Doppler, ephemeris-driven geometry, intermittent visibility, feeder links, and segment-dependent synchronization.	Fusion, feedback, placement, and freshness depend on the active space, air, or ground segment.	Clock/ephemeris error, visibility windows, gateway capacity, link budget, and terrestrial or unsynchronised/noncoherent baselines as appropriate.

7.2. Hardware Evidence

Table 9 compares experimental objects, execution modes and scope boundaries across complementary hardware platforms, without assigning an ordinal evidence ranking.

These works provide complementary evidence, not a common end-to-end benchmark. Hardware claims should therefore report, where relevant, centre frequency and occupied bandwidth; ADC/DAC effective resolution and clipping/full-scale operating point; PA operating point; oscillator and clock residuals; RF/array calibration and calibration age; residual self-interference; software/firmware version; processing location; and end-to-end latency distribution when real-time execution is claimed [36,40].

7.3. Security, Privacy, and Trust

ISAC exposes communication information, sensing signals/data/results, identities and location, models, and provenance. A security analysis should therefore state the protected asset and distinguish at least passive unauthorised sensing/eavesdropping, active over-the-air manipulation/jamming/spoofing, and compromised cooperative, edge, or learning entities, together with assumed channel/location, aperture, waveform knowledge, and system access [54,80,81,82]. Threats such as false-target injection, replay and track manipulation, Byzantine fusion, poisoned/backdoored models, and edge tampering are meaningful only relative to such capabilities.

When $R_b$ and $R_e$ are achievable instantaneous communication rates under a stated wiretap model,

$$R_{\sec }={[R_b-R_e]}^{+}$$

(18)

is a communication-confidentiality metric; outage or ergodic secrecy should be used where appropriate in fading. For multiple legitimate links, a jointly achievable secrecy-rate allocation can also be summarized by

$$R_{\mathrm{sum}}^{\sec }=\sum _{\ell \in {\mathcal{L}}_{\mathrm{leg}}}{R}_{\ell }^{\sec },$$

where ${\mathcal{L}}_{\mathrm{leg}}$ is the declared set of legitimate links. The per-link rates and their common time/bandwidth normalization must follow the same transmission, interference, and eavesdropper model, including whether eavesdroppers collude. Boljević et al. study an aggregate uplink/downlink secrecy objective under sensing and power constraints [83]; the aggregation and leakage expressions must be interpreted under that model. Sum secrecy rate should be accompanied by per-user secrecy requirements when every user’s confidentiality matters. These secrecy-rate measures do not quantify unauthorised sensing. Instead, evaluate sensing confidentiality using task-relevant quantities such as unauthorised position/velocity accuracy or resolution, detection and false-alarm performance, image resolution, or a declared estimation-information metric. If the observer’s channel, location, aperture, waveform knowledge, or estimator is uncertain, the analysis should state whether uncertainty is bounded, stochastic, or Eve-independent; a known unauthorised Fisher information matrix gives only a conditional guarantee.

A sensed target may itself be a potential eavesdropper: the illumination needed for sensing can also expose communication-bearing signals. Secure full-duplex ISAC designs therefore couple transmit/receive beamforming, user power allocation, and artificial noise while imposing communication-secrecy and sensing constraints [84]. Related work minimizes transmit power under secrecy and sensing requirements [85]. Such results depend on the assumed eavesdropper channels or directions, decoding capabilities, and residual self-interference; they do not establish unconditional confidentiality.

Privacy additionally requires purpose limitation/data minimisation, consent or applicable authorisation, transparency, access control, retention/deletion, and protection of sensing data, features, and model updates. Cooperative sensing additionally needs integrity, secure sharing, and provenance/accountability [54]. Trustworthiness is not a synonym for security or privacy: it concerns measurable or verifiable assurance in relevant properties such as accuracy, authenticity, integrity, availability, accountability, privacy, and transparency. A claimed “trust state” should therefore expose named indicators rather than collapse these properties into one opaque scalar.

• Cross-layer implication.

Downstream metadata should separately identify model regime/validity range, timestamp/ synchronisation residual, calibration version/residual, statistical estimator uncertainty, authorisation/purpose, provenance/integrity and information age. A generic confidence score obscures distinct failures: a precise estimate may be stale or miscalibrated, a physically valid result unauthorised, and an authentic result based on an invalid model. Deployment therefore requires interface and measurement validation alongside algorithm design.

8. Standards, Data Resources, and Reproducibility

ISAC standardisation spans frameworks, service requirements, architecture studies, radio procedures and interoperability conditions. Table 10 consolidates the selected documents and their status at the stated cutoff. Their roles must remain distinct: a framework or study does not itself establish an interoperable implementation or conformance profile.

Experimental evidence comes from datasets, simulators, prototypes, and networked testbeds. DeepSense 6G provides real-world multimodal sensing and communication measurements, whereas ViWi is a synthetic vision–wireless dataset framework generated using 3-D modelling and ray tracing [77,86]. Widar3.0 and WiMANS are Wi-Fi sensing benchmarks for gesture and multi-user activity recognition, respectively, and are therefore treated as adjacent sensing evidence rather than direct validation of a cellular ISAC service [87,88]. Micro-Doppler describes time-varying frequency components caused by motions within a target, such as moving limbs, beyond its bulk motion [89]. These components can provide activity features; Widar3.0 uses Doppler-derived body-coordinate velocity profiles, whereas WiMANS provides CSI recordings and labels for simultaneous multi-user activity sensing [87,88]. With multiple users, overlapping Doppler components require separation or association using available spatial, temporal, or learned structure. Evaluations should state the number of users, geometry, CSI calibration, processing window, and held-out subjects/sites. A multi-user dataset alone does not establish successful micro-Doppler-based user separation.

Hardware evidence spans waveform proof-of-concept, GNU Radio and portable THz demonstrations, and synchronised multi-node sounding. Offline datasets support algorithm evaluation, not real-time execution; prototype feasibility does not alone establish scalability, interoperability or conformance. The reporting requirements below cover configuration, trials, latency and data/software availability.

8.1. Document Status and Scope Boundaries

Table 10 distinguishes framework recommendations, feasibility reports, Stage-1 specifications, RAN studies, ETSI group reports and WLAN amendments by status, normative role and scope at the stated cutoff.

Table 10. Status, scope, and integration boundary of selected principal ISAC-related documents at the 26 July 2026 cutoff.

Document	Status/role	Use-case or procedure	Input/output or KPI provision	Scope boundary / integration issue
ITU-R M.2160-0 and 2026 WP 5D draft IMT-2030 requirements/evaluation work [1,90,91]	In-force IMT-2030 framework recommendation; draft technical-performance requirements and evaluation guidelines completed at WP 5D level by the cutoff	Defines ISAC as an IMT-2030 usage scenario and high-level capabilities; the 2026 drafts add performance requirements and RIT/SRIT evaluation guidance, including ISAC-specific test environments and channel models.	Framework-, requirement-, and evaluation-level capabilities rather than a completed IMT-2030 radio-interface specification.	These documents do not themselves define an IMT-2030 radio interface or implementation API.
3GPP TR 22.837 [46]	Release-19 feasibility/study report under change control	ISAC use cases and feasibility.	Study requirements and scenario material.	Study report; it is not a normative RAN procedure or application API.
3GPP TS 22.137 [47]	Release-19 Stage-1 service specification under change control	Normative service requirements.	Service-level requirements and KPIs.	Stage-1 scope; protocol and implementation details belong to later-stage specifications.
3GPP TR 23.700-14 / TS 23.137 [50,51]	Release-20 Stage-2 study under change control / draft specification	5GS Stage-2 architecture and sensing-service procedures.	Functions, procedures, capability and policy context under development.	NR physical-layer procedures remain within RAN scope; the TS is still in draft status at the cutoff.
3GPP TR 23.700-15 / TS 23.138 [48,49]	Release-20 vertical-application study under change control / draft specification	Use and exposure of sensing results to vertical applications.	Application requests and exposed sensing results under development.	The TS is still in draft status; cross-technology interoperability and conformance lie outside this document’s scope.
3GPP TR 38.765 [17]	Release-20 RAN1 study under change control	NR ISAC physical-layer and RAN1 aspects.	Conclusions of the completed Release-20 study and studied candidate mechanisms.	End-to-end service exposure and application-layer semantics lie outside the RAN1 study’s scope.
ETSI GR ISC 001–004 [39,52,54,92]	Published pre-standardisation group reports	Use cases and deployment; channel modelling and evaluation; architecture; security, privacy, trustworthiness and sustainability.	Scenario, architecture, modelling and threat guidance.	Guidance is not a normative interoperable result API or conformance profile.
IEEE 802.11bf-2025 [53]	Active published WLAN sensing amendment	MAC modifications and PHY service-interface enhancements for WLAN sensing.	Sensing sessions/measurements within IEEE 802.11 operation.	Scope is IEEE 802.11 WLAN sensing; cellular and vertical-service interworking are outside the amendment’s scope.

Table 10. Status, scope, and integration boundary of selected principal ISAC-related documents at the 26 July 2026 cutoff.

Document	Status/role	Use-case or procedure	Input/output or KPI provision	Scope boundary / integration issue
ITU-R M.2160-0 and 2026 WP 5D draft IMT-2030 requirements/evaluation work [1,90,91]	In-force IMT-2030 framework recommendation; draft technical-performance requirements and evaluation guidelines completed at WP 5D level by the cutoff	Defines ISAC as an IMT-2030 usage scenario and high-level capabilities; the 2026 drafts add performance requirements and RIT/SRIT evaluation guidance, including ISAC-specific test environments and channel models.	Framework-, requirement-, and evaluation-level capabilities rather than a completed IMT-2030 radio-interface specification.	These documents do not themselves define an IMT-2030 radio interface or implementation API.
3GPP TR 22.837 [46]	Release-19 feasibility/study report under change control	ISAC use cases and feasibility.	Study requirements and scenario material.	Study report; it is not a normative RAN procedure or application API.
3GPP TS 22.137 [47]	Release-19 Stage-1 service specification under change control	Normative service requirements.	Service-level requirements and KPIs.	Stage-1 scope; protocol and implementation details belong to later-stage specifications.
3GPP TR 23.700-14 / TS 23.137 [50,51]	Release-20 Stage-2 study under change control / draft specification	5GS Stage-2 architecture and sensing-service procedures.	Functions, procedures, capability and policy context under development.	NR physical-layer procedures remain within RAN scope; the TS is still in draft status at the cutoff.
3GPP TR 23.700-15 / TS 23.138 [48,49]	Release-20 vertical-application study under change control / draft specification	Use and exposure of sensing results to vertical applications.	Application requests and exposed sensing results under development.	The TS is still in draft status; cross-technology interoperability and conformance lie outside this document’s scope.
3GPP TR 38.765 [17]	Release-20 RAN1 study under change control	NR ISAC physical-layer and RAN1 aspects.	Conclusions of the completed Release-20 study and studied candidate mechanisms.	End-to-end service exposure and application-layer semantics lie outside the RAN1 study’s scope.
ETSI GR ISC 001–004 [39,52,54,92]	Published pre-standardisation group reports	Use cases and deployment; channel modelling and evaluation; architecture; security, privacy, trustworthiness and sustainability.	Scenario, architecture, modelling and threat guidance.	Guidance is not a normative interoperable result API or conformance profile.
IEEE 802.11bf-2025 [53]	Active published WLAN sensing amendment	MAC modifications and PHY service-interface enhancements for WLAN sensing.	Sensing sessions/measurements within IEEE 802.11 operation.	Scope is IEEE 802.11 WLAN sensing; cellular and vertical-service interworking are outside the amendment’s scope.

8.2. Datasets, Testbeds, and Reproducibility

Datasets, synthetic frameworks, prototypes and multi-node testbeds support different claims [77,79,86,87,88,93]. Report artifact availability per resource: publication access does not establish artifact access. Beam-prediction data do not validate cellular sensing sessions, offline activity data do not establish deadline reliability, and SDR demonstrations do not establish conformance.

The following reporting items are recommended to support reproducibility for each evidence type. A reusable dataset should report carrier/bandwidth, waveform or modality, antenna/sensor geometry, environment, targets, clocks, calibration, ground truth, preprocessing, licence or data-use constraints, and train/validation/test domain separation. A simulator should state communication, target, clutter, shared-scatterer, blockage, hardware and synchronisation models, parameter sources or validation data, software/version, random seeds, and limitations. A prototype should identify hardware, waveform/frame, sampling, antennas, clocks, cancellation, processor, real-time or offline status, trial count, excluded runs, latency, energy, uncertainty, baselines, software version, and standards relationship.

• Cross-layer implication.

The evaluation setting limits the interpretation of a result: E1 supports behaviour under an ideal model, E3 supports measured feasibility within the reported configuration, and E5 can expose network synchronization and execution effects. Independent reproduction requires evidence from a separate reproduction effort. Standards alignment requires identification of the applicable document, version and procedures; conformance additionally requires assessment against the relevant normative requirements or conformance profile. None of these properties is established by an E0–E5 designation alone.

9. Contract Audit and Research Gaps

This section audits reported information, uncertainty, decisions, metrics and constraints to identify requirements for end-to-end use. The resulting agenda addresses uncertainty transfer, freshness, trustworthy control, task evaluation, hardware validity, service exposure and reproducible validation under consistent communication, sensing, computation and security assumptions.

9.1. Representative Contract Audit

Table 11 compares nine illustrative examples, including reported interface connections and requirements for further downstream use. The selection is qualitative: it does not estimate the prevalence of contract gaps or rank paper quality. Each claim is interpreted within the model or experiment that supports it.

Table 11. Representative ISAC contract audit: reported capability and requirements for downstream use.

Study / claim	Reported output(C1)	Contract coverage(C2–C6)	Required for downstream use	Evidence / source
Pan et al.Compression and control	Compressed observations; state estimate and covariance.	Kalman covariance; per-step compression; reconstruction and estimation errors, control cost; transmission-dimension budget.	Link errors and delay for deployment; the study assumes error-free, zero-latency links.	E1[56]Secs.III–IV
Hua et al.Rate–CRB tradeoff	Rate–CRB operating point and transmit covariance.	CRB bound; coherence-block covariance design; rate and scalarised CRB; power constraint.	Achieved errors, ambiguity / threshold behaviour, and estimator validity. A bound is not achieved uncertainty.	E1[21]
de Oliveira et al.Phase noise	OFDM sensing metrics and radar image under phase noise.	Impairment model; frame/CPI waveform and oscillator parameters; peak loss, sidelobes and interference.	Measured calibration residuals and uncertainty validated for the downstream task.	E2[40]
Babu et al.Multi-cell coordination	Multi-cell precoders and CBF/CoMP configuration.	CSI-error robustness; blockwise precoding and coordination; CRB, minimum SINR and power constraints.	Fused-target covariance, cross-node dependence, and signalling / synchronization overhead.	E1[55]
Fan et al.Joint resource control	Sensor schedule, compute allocation and phase-constrained beam.	Couples compute queues and semantic AoI to spatial uncertainty and PCRB; symbolwise scheduling; energy and phase constraints.	Calibrated tracking uncertainty and feasible execution on the target processor and network.	E1[64]Secs.II–V
Xu et al.OFDM experiment	Realised dual-functional waveform and measured outputs.	Framewise waveform tradeoff; BER, beampattern and activity sensing; SDR feasibility.	Achieved estimator uncertainty, calibration residuals and timing / converter validity.	E3[75]
DeepSense 6GMultimodal dataset	Images, radar, LiDAR, GPS and beam-power labels.	Sensor specifications and acquisition sequences; offline beam-prediction evaluation; public hardware/data description.	Per-sample uncertainty and cross-modal calibration; latency and privacy requirements for online use.	E3[77]
Wild et al.Shared-hardware prototype	Communication and sensing outputs from commercial hardware.	Prototype waveform/processing choices; data rate, pedestrian detection and range evidence; shared-hardware feasibility.	Calibrated uncertainty, reproducible calibration and an operational service interface.	E4[76]
Beuster et al.Multi-node sounding	Synchronized channel measurements with RTK/INS ground truth.	Real-time node configuration and sounding; position / synchronization support and mobility; localisation-enabling measurements.	Application-specific fusion confidence and service validation beyond synchronized sounding.	E5[79]

Note. C1: output; C2: uncertainty; C3: timescale; C4: decision; C5: evaluation; C6: admissibility. Coverage describes the reported model or experiment; the next column states requirements for subsequent fusion, control or service use, not a finding that every requirement is absent from the source. E1–E5 retain the evidence settings defined in Section 2.2 and refer to the listed claim, not overall paper quality.

Table 11. Representative ISAC contract audit: reported capability and requirements for downstream use.

Study / claim	Reported output(C1)	Contract coverage(C2–C6)	Required for downstream use	Evidence / source
Pan et al.Compression and control	Compressed observations; state estimate and covariance.	Kalman covariance; per-step compression; reconstruction and estimation errors, control cost; transmission-dimension budget.	Link errors and delay for deployment; the study assumes error-free, zero-latency links.	E1[56]Secs.III–IV
Hua et al.Rate–CRB tradeoff	Rate–CRB operating point and transmit covariance.	CRB bound; coherence-block covariance design; rate and scalarised CRB; power constraint.	Achieved errors, ambiguity / threshold behaviour, and estimator validity. A bound is not achieved uncertainty.	E1[21]
de Oliveira et al.Phase noise	OFDM sensing metrics and radar image under phase noise.	Impairment model; frame/CPI waveform and oscillator parameters; peak loss, sidelobes and interference.	Measured calibration residuals and uncertainty validated for the downstream task.	E2[40]
Babu et al.Multi-cell coordination	Multi-cell precoders and CBF/CoMP configuration.	CSI-error robustness; blockwise precoding and coordination; CRB, minimum SINR and power constraints.	Fused-target covariance, cross-node dependence, and signalling / synchronization overhead.	E1[55]
Fan et al.Joint resource control	Sensor schedule, compute allocation and phase-constrained beam.	Couples compute queues and semantic AoI to spatial uncertainty and PCRB; symbolwise scheduling; energy and phase constraints.	Calibrated tracking uncertainty and feasible execution on the target processor and network.	E1[64]Secs.II–V
Xu et al.OFDM experiment	Realised dual-functional waveform and measured outputs.	Framewise waveform tradeoff; BER, beampattern and activity sensing; SDR feasibility.	Achieved estimator uncertainty, calibration residuals and timing / converter validity.	E3[75]
DeepSense 6GMultimodal dataset	Images, radar, LiDAR, GPS and beam-power labels.	Sensor specifications and acquisition sequences; offline beam-prediction evaluation; public hardware/data description.	Per-sample uncertainty and cross-modal calibration; latency and privacy requirements for online use.	E3[77]
Wild et al.Shared-hardware prototype	Communication and sensing outputs from commercial hardware.	Prototype waveform/processing choices; data rate, pedestrian detection and range evidence; shared-hardware feasibility.	Calibrated uncertainty, reproducible calibration and an operational service interface.	E4[76]
Beuster et al.Multi-node sounding	Synchronized channel measurements with RTK/INS ground truth.	Real-time node configuration and sounding; position / synchronization support and mobility; localisation-enabling measurements.	Application-specific fusion confidence and service validation beyond synchronized sounding.	E5[79]

Note. C1: output; C2: uncertainty; C3: timescale; C4: decision; C5: evaluation; C6: admissibility. Coverage describes the reported model or experiment; the next column states requirements for subsequent fusion, control or service use, not a finding that every requirement is absent from the source. E1–E5 retain the evidence settings defined in Section 2.2 and refer to the listed claim, not overall paper quality.

The audited studies connect different parts of the sensing-to-decision chain. Hua et al. [21] optimise transmit covariance to characterise a rate–CRB tradeoff under a power constraint; the CRB remains a bound whose use as an estimator uncertainty measure requires justification. Pan et al. [56] connect observation compression, Kalman-filter estimation and control cost, while assuming error-free links with no latency. Fan et al. [64] explicitly couple computation queues and semantic information age to tracking PCRB and an energy budget, with simulation-based evaluation.

These comparisons identify different downstream requirements: relating a bound to achieved uncertainty, preserving validity during transport and processing, and executing the selected action within available resources. An interface can therefore be connected within a published model while further validation is needed under delay, calibration error, or processor constraints. The audit supports these model-specific questions; it does not establish a field-wide ranking of contract gaps.

9.2. Research-Gap Matrix

Table 12 links six cross-layer interfaces to research questions and required validation evidence. Five priorities follow, with the first covering both physical-to-network interfaces and propagation/hardware validity.

First, develop calibrated physical-to-network interfaces that account for propagation and hardware validity. Second, treat sensing freshness, transport, fusion, and computation queues as one control problem. Third, make learned control selective and recoverable through uncertainty calibration, constraint monitors, and fallback. Fourth, define task losses and matched baselines that make semantic claims falsifiable. Fifth, progress from impairment-aware simulation to standard-aligned, independently reproducible multi-node validation.

For analytical work addressing these priorities, a rate–distortion–computation–latency region may be useful only after the state/task law, finite frame, service process, compute cost, and deadline event are operationally specified. These definitions are needed to interpret the region and compare the results.

10. Conclusions

Deployable ISAC requires sensing results to remain statistically meaningful, timely, feasible, trusted, and reproducible through network coordination, learned control, computation, and application exposure. This survey applies six operational contracts across seven functional layers to examine these transitions.

The representative audit illustrates recurring requirements for transferring uncertainty, validity, information age, and constraint status between functions. It also includes studies that already connect sensing, communication, estimation, or computation within declared models. These examples support examining the remaining interface requirements under matched resources and evidence appropriate to the claimed operating setting; they do not establish a field-wide ranking of gaps. End-to-end validation should therefore test whether the information and assumptions required by each receiving function remain valid through execution and application exposure.