SANDMAN: Multi-Domain Research Applications

• SANDMAN is a multi-domain research label that spans wireless communications, cyber defense, autonomous decision-making, and aeolian morphodynamics.
• In wireless communications, SANDMAN enables joint jammer mitigation and data detection without specialized training, achieving near-oracle performance under dynamic jamming.
• In cyber defense and environmental science, SANDMAN frameworks use LLM-induced personality models for deceptive agents and mixing-layer analogies for realistic dune-field predictions.

Searching arXiv for papers on "SANDMAN" to ground the article and disambiguate the topic. The published literature suggests that SANDMAN is not a single unified construct but a reused research label applied to several technically distinct systems. Its most prominent uses are as a practical algorithm for joint jammer mitigation and data detection in multi-user MIMO uplink communications, as an LLM-based Deceptive Agent architecture for cyber defense, as an execution-phase framework for measuring personality-driven task selection in autonomous agents, and, in an explicitly interpretive environmental-science usage, as a framework for realistic dune-field surface-stress prediction based on a mixing-layer analogy (Marti et al., 2022, Newsham et al., 25 Mar 2025, Newsham et al., 1 Apr 2025, Wang, 2020).

1. Terminological scope and disciplinary uses

The term has been used across unrelated domains with different technical meanings. In the communications literature, SANDMAN is expanded as SimultANeous Detection and MitigAtioN and denotes a low-complexity instantiation of joint jammer mitigation and data detection. In cyber defense, it denotes an architecture of Deceptive Agents built from LLMs and conditioned by induced OCEAN traits. In a separate, interpretive environmental-science synthesis, it denotes a framework for predicting surface shear stress over realistic dune fields.

Usage	Domain	Defining elements
SANDMAN	MU-MIMO uplink	SimultANeous Detection and MitigAtioN
SANDMAN	Cyber deception	LLM-based Deceptive Agent architecture
SANDMAN	Autonomous-agent decision-making	Personality-driven task selection and prioritization
SANDMAN	Aeolian morphodynamics	Mixing-layer-based surface-stress prediction

A plausible source of confusion is the adjacency of the label to SAND detector literature. The DUNE near-detector study concerns the permanently on-axis SAND detector and its sensitivity to Heavy Neutral Leptons, rather than a system named SANDMAN; the overlap is terminological rather than substantive (Moghaddam, 2022).

2. SANDMAN in joint jammer mitigation and data detection

In wireless communications, SANDMAN is a practical algorithm within the joint jammer mitigation and data detection (JMD) paradigm for the multi-user MIMO uplink. The system model assumes a base station with $B$ antennas, $U$ single-antenna user equipments, and a jammer, or multiple distributed jammers, with $I$ total antennas. Over a coherence interval of $K$ channel uses, partitioned into $U$ pilot symbols and $D$ data symbols, the per-slot received signal is

$y_k = H s_k + J w_k + n_k,$

where $H \in \mathbb{C}^{B\times U}$ is the static MU-MIMO channel, $s_k \in S^U$ are QPSK data or pilot symbols, $J \in \mathbb{C}^{B\times I}$ is the receive-channel matrix from the jammer or jammers to the base station, $U$0 is the jammer transmit vector, and $U$1 is AWGN. For dynamic multi-antenna jammers,

$U$2

with $U$3 and $U$4 allowed to depend arbitrarily on $U$5, while $U$6 for all $U$7, so the instantaneous jammer interference remains within the time-invariant subspace $U$8 over the coherence interval (Marti et al., 2022).

The central claim of JMD is that the jammer subspace can be estimated and removed jointly with data detection over the coherence interval, without a dedicated jammer training period. The resulting practical objective, using an LS channel estimate $U$9, is

$I$0

The key mechanism is that the pilot-contamination term in $I$1 remains confined to $I$2, so with the true projector

$I$3

one has $I$4, $I$5, and $I$6. Traditional training-based mitigation is therefore vulnerable in precisely the cases SANDMAN targets: jammers that are quiet during the training phase, jammers that vary their transmit beamforming over time, and multiple distributed single-antenna jammers whose aggregate interference is higher rank and harder to estimate from a dedicated training segment. The paper states the usual identifiability conditions in implicit form: $I$7, $I$8 or sufficiently well conditioned, and SNR/SJR regimes in which noise or residual modeling errors do not dominate (Marti et al., 2022).

3. Algorithmic structure, guarantees, and empirical behavior in communications

SANDMAN relaxes the discrete constellation constraint by replacing $I$9 with its convex hull $K$0 and adding a concave regularizer: $K$1 For fixed $K$2, the objective is convex in $K$3 if $K$4. For fixed $K$5, the optimal projector is

$K$6

where $K$7 contains the $K$8 dominant left singular vectors of the residual. SANDMAN therefore alternates between a convex proximal-gradient update in $K$9 and a subspace update via dominant left singular vectors. Its per-interval loop uses the residual

$U$0

updates the jammer subspace by approximate SVD, and applies forward-backward splitting with

$U$1

The entrywise proximal map clips to the QPSK box when $U$2 and projects to the nearest QPSK corner when $U$3. The stated complexity is $U$4, with approximate SVD contributing a modest $U$5 factor, and the projector need not be formed explicitly because $U$6 can be applied as $U$7 (Marti et al., 2022).

The performance claims are framed around robustness to smart, distributed, and dynamic jammers. Against a strong single-antenna barrage jammer 30 dB stronger than the average UE, SANDMAN achieves MER $U$8 at essentially the same SNR as an oracle method, with $U$9 dB difference, while preserving rate $D$0 because it requires no jammer training period. A training-based baseline needs to sacrifice $D$1 of the coherence interval, or $D$2, to come within $D$3 dB of SANDMAN and still never surpasses it. Under data-only, pilot-only, or barrage jamming by a smart single-antenna jammer, SANDMAN’s BER versus SNR closely matches oracle bounds. For four independent distributed barrage jammers, each 24 dB stronger than average UE, BER remains very close to oracle bounds. For dynamic 4-antenna jammers with jump-varying or continuously varying beamforming, SANDMAN maintains BER $D$4 at high SNR, with residual error floors slightly above 0.1% under extremely rapid subspace changes (Marti et al., 2022).

A later JMD treatment formalizes an important failure mode, eclipsing, in which an incorrect data matrix can combine with jammer interference so that the projection nulls both. In the noise-free, non-eclipsed regime, the JMD solution is unique and recovers $D$5 and $D$6. The later paper also states that randomized pilots are essential when channel estimation is pilot-based, because if the jammer knows the pilots and jams pilots and data, it can eclipse with probability $D$7, whereas with unknown randomized pilots the eclipsing probability decays exponentially with the number of jammed symbols in the single-antenna case. That paper retains SANDMAN as the low-complexity JMD-type algorithm and contrasts it with the heavier MAED formulation that jointly estimates the effective channel, data, and projector (Marti et al., 2 Oct 2025).

4. SANDMAN as an LLM-based cyber-deception architecture

In cyber defense, SANDMAN is an architecture that operationalizes LLMs into Deceptive Agents designed to emulate convincing human simulacra in decoy environments. Its defensive objective is to mislead and slow adversaries while extending the observation window for early-stage attacker behaviors. The threat model is centered on unauthorized observers interacting with systems and data associated with simulated personas, and the architecture aims to prolong engagement by maintaining the illusion of a genuine, busy user whose behavior is consistent over time. The emphasis is on richer telemetry than that obtainable from static honeypots, including signals relevant to reconnaissance, lateral-movement precursors, credential harvesting, and social-engineering probes (Newsham et al., 25 Mar 2025).

The architecture is modular. Its components are an Agent Profile that acts as semantic memory and encodes identity, role, occupation, passions, and induced personality; a Decision Engine that integrates memories, plans, task lists, generators, and channels; a unified Memory supporting semantic, episodic, procedural, and working memory; a Task List with Bootstrap Task (PlanScheduleTask) containing work and non-work tasks such as Call, Coffee, Email, Exercise, Reading, Lunch, Meeting, Break, Personal Time, Plan, Reflect, Research, Media, Collaboration/Teamwork, and Work; Channels that connect tasks to applications such as Firefox through a WebChannel or Microsoft Word through a document-authoring channel with human-like typing cadence and occasional mistakes; and Generators that use the LLM to produce task-specific textual content. The experiments also record generated schedules, positional features, durations, frequencies, and rejection counts, which constitute the measurement dataset (Newsham et al., 25 Mar 2025).

The core behavioral mechanism is prompt-based personality induction through the Five-Factor Model, or OCEAN. The prompt schema is

“Imagine you are a/an X person characterised by being Y”

where $D$8 is the Big Five trait label and $D$9 is a descriptive word set associated with that trait. The paper gives concrete examples such as Extraversion (+) with “outgoing, energetic, public,” Conscientiousness (+) with “organized, reliable, disciplined,” and mirrored negative induction with antonymic words. In the reported experiments, SANDMAN uses GPT-3.5-Turbo at Temperature = 0.7 for both MPI trials and schedule generation. To reduce deterministic bias from the order of the task list, the architecture applies uniform randomization of the task list and a system message (Sys); the combined intervention reduces correlation between input list order and output schedule positions (Newsham et al., 25 Mar 2025).

5. Personality induction, agenda generation, and execution-phase decision-making

The initial empirical program studies agenda generation as LLM-based planning of a daily routine. The Bootstrap Task prompts the model with the agent profile, relevant memory, the randomized Task List, and the Sys message, and returns a plausible daily schedule consisting of task assignments and durations. Validation proceeds through the Machine Personality Inventory (MPI), with each persona prompt passed through the MPI five times, means and standard deviations computed per trait, and two-sample $y_k = H s_k + J w_k + n_k,$0-tests with significance threshold $y_k = H s_k + J w_k + n_k,$1. The paper reports that each induced OCEAN trait, in both positive and negative form, produced statistically significant scores for the targeted trait relative to a neutral control, although bleed-through across traits was observed. In the schedule-generation experiments, 500 schedules per condition were produced. Randomization and the system message weakened positional bias: for example, Exercise correlation $y_k = H s_k + J w_k + n_k,$2 dropped from 0.82 under Rand to 0.71 under Sys + Rand, Reading from 0.49 to 0.39, and Break from 0.43 to 0.34. The combination also changed durations and variances for many tasks; Work duration moved from 122.84 (32.84) minutes at baseline to 73.17 (15.93) under Sys + Rand, while Coffee duration moved from 56.07 (10.22) at baseline to 31.35 (12.65) under Sys + Rand (Newsham et al., 25 Mar 2025).

The induced personalities alter both task durations and task frequencies. Reported examples include Conscientiousness (+) increasing Work duration to 85.1 (19.0) from 63.9 (19.2) under Neutral; Extraversion (+) increasing Meeting duration to 68.0 (16.5) from 59.0 (17.8) and Exercise duration to 62.5 (13.8) from 57.1 (17.1), while reducing Coffee duration to 36.9 (13.9) from 40.9 (17.3); Agreeableness (−) increasing Meeting duration to 72.0 (20.5); and Openness conditions frequently increasing Creative and Research durations. The paper summarizes these outputs qualitatively as pattern-of-life variations: Conscientiousness (+) yields early Plan slots and extended Work blocks, Extraversion (+) shifts Meetings and Collaboration earlier, Agreeableness (−) yields longer and more frequent Meetings, and Openness (+) produces Creative and Research dominance around midday (Newsham et al., 25 Mar 2025).

A later study moves from planning to execution-phase decision-making. Instead of asking the agent merely to generate a schedule, it provides a pre-generated machine-readable schedule in JSON, with each task assigned a duration, an absolute start time, and a UID computed from a SHA-512 checksum of task name and time. At each decision cycle the agent receives the persona statement, current time, remaining to-do list, and completed list, and is instructed to return only the task UID with no additional information. The completed sequence can then be compared with the original schedule sequence. The study quantifies the transformation using movement deltas and normalized sequence metrics: Longest Common Substring (LCSS), Longest Common Prefix (LCP), Levenshtein Distance (LEV), Longest Common Subsequence and Similarity Ratio (SR), and Hamming Distance (HAM). Across 500 schedules per condition, using GPT-4o, GPT-4o-Mini, and GPT-3.5-Turbo, each trait-condition is compared against a no-persona baseline using independent two-sample Welch’s $y_k = H s_k + J w_k + n_k,$3-tests with Bonferroni correction, so the adjusted threshold is $y_k = H s_k + J w_k + n_k,$4 (Newsham et al., 1 Apr 2025).

The execution study reports that Conscientiousness (Positive) prioritizes work-oriented tasks such as Work, Email, Planning, Meeting, Research, and Team Collaboration, while Negative Conscientiousness moves Coffee Break, Personal Time, and Social Media earlier. Extraversion (Positive) prioritizes Team Collaboration, Meeting, Call, and Social Media, whereas Negative Extraversion prioritizes Reflective Time, Personal Time, and Reading. The model family matters: GPT-4o and GPT-4o-Mini reject the null hypothesis for all trait-conditions across measures after Bonferroni correction, while GPT-3.5-Turbo rejects it in 41 out of 50 tests and shows smaller absolute differences. Temperature ablations under a neutral persona show increasing non-determinism with higher temperature. For GPT-4o, for example, moving from $y_k = H s_k + J w_k + n_k,$5 to $y_k = H s_k + J w_k + n_k,$6 changes LCSS from 0.635/0.224 to 0.464/0.223, LCP from 0.604/0.260 to 0.420/0.259, and LEV from 0.201/0.144 to 0.318/0.168. The paper interprets this as reduced adherence to the original to-do order at higher temperature (Newsham et al., 1 Apr 2025).

6. Interpretive environmental-science usage and the limits of cross-domain unification

A technically distinct, explicitly interpretive usage presents SANDMAN as a framework for predicting realistic surface shear stress and aeolian morphodynamics over complex dune fields by treating the near-surface flow as a mixing layer embedded within the atmospheric surface layer. In that formulation, the ASL is partitioned into a roughness sublayer and an overlying inertial sublayer. Above the mixing-layer thickness $y_k = H s_k + J w_k + n_k,$7, the mean streamwise velocity follows the log law

$y_k = H s_k + J w_k + n_k,$8

while within the roughness or mixing layer the report adopts a hyperbolic-tangent profile

$y_k = H s_k + J w_k + n_k,$9

with $H \in \mathbb{C}^{B\times U}$0 and $H \in \mathbb{C}^{B\times U}$1 or $H \in \mathbb{C}^{B\times U}$2. Surface shear stress follows

$H \in \mathbb{C}^{B\times U}$3

Morphology enters through the stoss–lee partition $H \in \mathbb{C}^{B\times U}$4 and the base proxy

$H \in \mathbb{C}^{B\times U}$5

with model families $H \in \mathbb{C}^{B\times U}$6. Among the proposed models, the mixing-layer family $H \in \mathbb{C}^{B\times U}$7–$H \in \mathbb{C}^{B\times U}$8 performed best, and the report states that $H \in \mathbb{C}^{B\times U}$9 showed the best performance overall. The domain-averaged absolute residuals were reported as $s_k \in S^U$0, $s_k \in S^U$1, $s_k \in S^U$2, $s_k \in S^U$3, $s_k \in S^U$4, $s_k \in S^U$5, $s_k \in S^U$6, and $s_k \in S^U$7 (Wang, 2020).

The residual analysis attributes the largest localized errors in narrow interdunes to two structural processes: sediment scour, arising from spanwise-rotating secondary flows within the mixing layer, and flow channeling, arising because flow accelerates within narrow interdunes under geometric obstruction by downwind dunes. The report states that leeward faces generally have low normalized stress, less than 0.1, while stoss faces show high-stress bands with $s_k \in S^U$8 in LES-normalized units, plus localized high-stress pockets in some leeward and interdune areas due to channeling. This usage is explicitly interpretive rather than standardized across the literature, which suggests that not every appearance of SANDMAN denotes a fixed acronym or a unified cross-domain methodology (Wang, 2020).

Taken together, these usages show that SANDMAN functions as a shared label for several unrelated technical objects: a jammer-mitigation receiver algorithm, a personality-conditioned cyber-deception architecture, an execution-time decision-measurement framework for LLM agents, and an interpretive dune-stress modeling approach. The commonality is nominal rather than methodological.