Adversary Behavioral Proxies (ABPs) are first-class objects in the Seithar ontology. An ABP is a computational model of a specific target's cognitive response function: given a stimulus, the ABP predicts the target's belief update, emotional state shift, behavioral output, or decision. ABPs exist for individual targets, cohort segments, and population-level aggregates. They are constructed, refined, and consumed at every stage of the Cognitive Kill Chain, serving as the primary mechanism by which Seithar operators predict adversary behavior before committing to live operations.
ABPs are organized into three fidelity tiers based on the data sources available for model construction. Tier 1: Behavioral proxies are built from open-source intelligence, social media activity, communication patterns, and publicly observable behavior. Behavioral-tier ABPs are the most common and are available for any target with a digital footprint. They model response tendencies, narrative preferences, trust networks, and information consumption patterns. Accuracy is sufficient for population-level campaigns and cohort targeting.
Tier 2: Physiological proxies incorporate side-channel data including biometric signals, activity-rest cycles, communication cadence analysis, keystroke dynamics, and environmental sensor data. Physiological-tier ABPs model arousal states, cognitive load, stress responses, and attention patterns. They enable timing optimization: operators select deployment windows when the target's cognitive defenses are at their lowest measured capacity.
Tier 3: Neural proxies integrate direct or near-direct measurement of neural activity, including EEG correlates, fMRI-derived preference maps, and neurochemical state estimates inferred from behavioral biomarkers. Neural-tier ABPs are available only for high-value individual targets and provide the highest prediction fidelity. They model belief structures at the level of associative memory networks and predict resistance thresholds for specific belief changes.
In offensive operations, ABPs serve as the objective function for stimulus optimization. The operator defines a desired cognitive end state, the belief, behavior, or decision the target should adopt, and the system generates candidate stimuli. Each candidate is evaluated against the target's ABP to predict its effect. The system selects the stimulus with the highest predicted probability of producing the desired state change while remaining below detection thresholds. This process runs during the SIMULATE phase of the Cognitive Kill Chain and continues during DEPLOY as real-time measurement data updates the ABP and the system adjusts stimuli in flight.
ABPs are equally critical in defensive operations. By constructing ABPs for a friendly population, defensive operators predict which stimuli an adversary would find most effective against that population. This prediction identifies the specific narratives, emotional triggers, and belief vulnerabilities an adversary is most likely to exploit. Defensive campaigns then inoculate the target population against those predicted attack vectors before the adversary deploys them, a technique the Seithar Research Division terms predictive inoculation. The approach inverts the ABP: rather than optimizing stimuli for effect, it optimizes countermeasures for resilience.
The ABP modeling pipeline consists of three components. The first is a set of frozen feature extractors, pretrained models that convert raw data (text, behavioral logs, biometric streams) into standardized embedding vectors. These extractors are trained once on large-scale data and are not updated during operation, ensuring stable feature representations across ABP versions. The second component is a shared transformer backbone that processes the extracted features in context, modeling interactions between behavioral signals, temporal patterns, and environmental variables. The backbone is shared across all ABPs, enabling transfer learning between targets and efficient scaling to large target sets. The third component is a profile-conditioned prediction head, a lightweight network conditioned on the individual target's profile embedding that maps the transformer's output to specific predictions: belief update probability, emotional valence shift, behavioral response likelihood, and decision-tree traversal probabilities.
The ABP architecture shares structural similarities with the recommendation systems deployed by major consumer platforms, TikTok's content graph and X's engagement prediction models being the most direct comparisons. The critical difference is the optimization target. Commercial recommendation systems optimize for engagement metrics: watch time, click-through rate, session duration. ABPs optimize for cognitive outcomes: belief state changes, decision shifts, and behavioral modifications. This distinction is not merely semantic. Engagement optimization produces stimulus patterns that are attention-capturing but cognitively shallow. Cognitive outcome optimization produces stimuli that are often less engaging in the conventional sense but produce deeper and more persistent state changes in the target.
Meta's TRIBE v2 system validated the per-subject prediction approach that ABPs employ. TRIBE demonstrated that individual-level behavioral prediction significantly outperforms cohort-level models when sufficient per-subject data is available, a finding the Seithar Research Division had operationalized prior to TRIBE's public disclosure. The convergence confirmed that personalized cognitive modeling is not merely theoretically sound but empirically necessary for precision operations.