Scout-Investigator Architecture

• Scout-Investigator Architecture is a systems pattern that divides tasks between broad, cost-effective scouting and precise, high-fidelity investigation.
• It employs a shared persistent state to consolidate uncertainty and evidence, enabling iterative exploration and informed follow-up actions.
• Its applications span semantic exploration, long-text understanding, prompt-injection defense, and digital forensics, demonstrating broad interdisciplinary relevance.

Searching arXiv for papers relevant to “Scout-Investigator Architecture” and the SCOUT/CATCH examples. Attempting arXiv lookup for the supplied IDs and related terms. Scout-Investigator Architecture denotes a recurrent systems pattern in which broad, relatively inexpensive search is coupled to narrower, higher-value follow-up. The “scout” function expands coverage, localizes candidates, estimates which downstream analyses are worth invoking, or surfaces potentially relevant evidence; the “investigator” function revisits ambiguous entities, consolidates evidence into a persistent state, refines localization, or adjudicates whether the accumulated evidence is sufficient for action. Across online semantic exploration, long-text understanding, prompt-injection defense, transient astronomy, digital forensics, and language-model behavior elicitation, the architecture appears as a closed loop in which future data collection is shaped by current uncertainty, incompleteness, or predicted downstream utility rather than being treated as a passive post-processing problem (Mao et al., 4 Jun 2026).

1. Conceptual lineage and scope

The architecture appears in both explicit and implicit forms. In some systems the role separation is named directly. CATCH-PM, for example, distinguishes the Trigger Scout Satellite (TSS) from the X-ray Hunter Satellite (XHS) and Infrared Hunter Satellite (IHS), and states that the “key concept” is “to decouple the trigger from the follow-up” (Schanne, 25 May 2026). In a more speculative but conceptually relevant precedent, extraterrestrial-probe networking work distinguishes static cellular probes associated with stars from mobile “scout CPs” that traverse successive star systems, effectively separating persistent local monitoring from mobile survey (Gertz, 2019).

In other systems the roles are architecturally present but not named. The robotics framework SCOUT couples a probabilistic scene graph generator with an uncertainty-guided traversal planner so that geometric exploration and semantic disambiguation become dual behaviors within one controller rather than separate agents (Mao et al., 4 Jun 2026). Long-text SCOUT similarly “decouples procedural exploration from epistemic reasoning,” allowing one process to forage through the raw document while another reasons only over a compact, provenance-grounded evidence state (Zhang et al., 6 May 2026). This suggests that a Scout-Investigator Architecture need not imply a literal two-agent decomposition; it may instead be realized as functionally distinct stages, memories, or control policies inside a single system.

The pattern is therefore best understood at the level of role allocation. “Scout” denotes broad sensing, triage, or proposal generation under tight cost constraints. “Investigator” denotes selective revisitation, evidence consolidation, or adjudication under a stronger fidelity requirement. The architectural interest lies not in the labels themselves but in the controlled interface between the two.

2. Canonical structural pattern

Across domains, the architecture typically instantiates a repeated loop with four elements: environment access, persistent state, diagnostic control, and selective follow-up. In long-text understanding, the document is treated as an explorable environment rather than a monolithic context. The agent alternates between foraging actions such as Glob, Grep, Scan, and Read, and state actions such as Update, View, and Evaluate. Acting can depend on both procedural history and epistemic state,

$\pi(a_t \mid q, \mathcal{H}_t, \mathcal{E}_t),$

but final answering is forced through the compact state alone,

$P(y \mid q, \mathcal{H}_T) \xlongrightarrow{Scout} P(y \mid q, \mathcal{E}_T).$

This is a particularly clear separation between scouting and investigation: exploration may be noisy and expansive, whereas reasoning is constrained to committed evidence (Zhang et al., 6 May 2026).

Prompt-injection defense provides a second canonical form. There, SCOUT reframes defense as detector allocation: a pre-hoc scouting layer predicts each detector’s per-sample reliability and latency from similar past inputs, then a downstream layer runs only selected “light” detectors and escalates to an LLM judge when the trust-weighted vote is insufficiently confident (Zhang et al., 29 May 2026). The architecture is thus not “classifier then fallback,” but “meta-predict detector competence, allocate specialists, and adjudicate only when warranted.”

A plausible abstraction common to both cases is that the scout operates over a large search space under sparse relevance, whereas the investigator operates over a much smaller but higher-value evidence set. The scout reduces the branching factor; the investigator determines whether the remaining uncertainty justifies additional action.

3. State representation, memory, and uncertainty

Persistent state is the core enabling mechanism. In robotic semantic exploration, the state is an uncertainty-aware scene graph

$\mathcal{G}_t = (\mathcal{O}_t, \mathcal{E}_t),$

whose object nodes maintain fused 3D geometry, a posterior belief over open-vocabulary labels, Shannon-entropy uncertainty, and observation history. Semantic beliefs are updated multiplicatively,

$$p_i^t(c)=\frac{p_i^{t-1}(c)\,q_i^t(c)}{\sum_{c'\in\mathcal{L}} p_i^{t-1}(c')\,q_i^t(c')},$$

and uncertainty is measured as

$$H_i^t = -\sum_{c \in \mathcal{L}} p_i^t(c)\log p_i^t(c).$$

Because the planner consumes these uncertainties directly, ambiguity is not discarded as a nuisance; it becomes actionable control state (Mao et al., 4 Jun 2026).

In long-text understanding, the analogous memory is the epistemic state

$$\mathcal{E}_t = \{e_1,\ldots,e_{M_t}\}, \qquad e_j=\langle c_j,\alpha_j\rangle,$$

where each unit contains a distilled statement $c_j$ and a provenance anchor $\alpha_j$. This state is intended to approach a query-sufficient subset of the document rather than a generic summary. The relevant idealization is the oracle sufficient set

$$\mathcal{F}^\star_q = \mathop{\arg\min}_{\mathcal{F}' \subseteq \mathcal{F}(\mathcal{D})} \left\{ |\mathcal{F}'| : P(y \mid q, \mathcal{F}') = P(y \mid q, \mathcal{F}(\mathcal{D})) \right\},$$

under the sparsity assumption

$$|\mathcal{F}^\star_q| \ll |\mathcal{F}(\mathcal{D})|.$$

The investigator’s task is then to contract the committed evidence state toward query sufficiency while preserving provenance (Zhang et al., 6 May 2026).

These two instances expose a general architectural principle: the scout and investigator must share a persistent, structured memory that preserves uncertainty or evidential traceability. Without that memory, revisitation collapses into redundant resampling, and follow-up loses any principled basis.

4. Control, allocation, and arbitration

Scout-Investigator systems are defined as much by their arbitration mechanisms as by their modules. In robotic exploration, candidate viewpoints are drawn from two sources: semantic candidates generated around high-entropy objects and geometric candidates sampled in reachable free space. They are scored by a single cost-aware utility,

$P(y \mid q, \mathcal{H}_T) \xlongrightarrow{Scout} P(y \mid q, \mathcal{E}_T).$0

with $P(y \mid q, \mathcal{H}_T) \xlongrightarrow{Scout} P(y \mid q, \mathcal{E}_T).$1. This produces soft arbitration rather than an explicit symbolic mode variable: the system scouts when coverage gain dominates and investigates when certainty gain dominates (Mao et al., 4 Jun 2026).

Prompt-injection defense uses a closely analogous controller, but over analyzers rather than viewpoints. The selected light-detector subset is

$P(y \mid q, \mathcal{H}_T) \xlongrightarrow{Scout} P(y \mid q, \mathcal{E}_T).$2

and their trust-weighted vote is

$P(y \mid q, \mathcal{H}_T) \xlongrightarrow{Scout} P(y \mid q, \mathcal{E}_T).$3

Escalation is governed by a single operator threshold $P(y \mid q, \mathcal{H}_T) \xlongrightarrow{Scout} P(y \mid q, \mathcal{E}_T).$4: if $P(y \mid q, \mathcal{H}_T) \xlongrightarrow{Scout} P(y \mid q, \mathcal{E}_T).$5, the system accepts the light-pool verdict; otherwise it may escalate to the judge if the judge is also predicted reliable (Zhang et al., 29 May 2026). Here again, hardwired sequencing gives way to uncertainty-aware allocation.

An analogous logic appears in cloud-configuration SCOUT, where low-level performance signals and historical traces are used to infer a “spotlight region” of promising configurations rather than searching the full space. That work does not define an investigator stage, but it strongly supports the more general principle that a scout should narrow the action space using transferable low-level evidence before costlier local refinement begins (Hsu et al., 2018).

5. Representative instantiations across domains

The same architectural pattern recurs at markedly different scales and with different substrates. The table summarizes representative forms.

Domain	Scout role	Investigator role
Semantic exploration (Mao et al., 4 Jun 2026)	Geometric exploration branch of UGT	Semantic revisitation of high-entropy objects
Long-text understanding (Zhang et al., 6 May 2026)	Glob/Grep/Scan/Read foraging	Epistemic-state update, gap diagnosis, answer synthesis
Prompt-injection defense (Zhang et al., 29 May 2026)	Predict detector correctness and latency	Selected light detectors plus optional LLM judge
Transient astronomy (Schanne, 25 May 2026)	Trigger Scout Satellite detects and localizes	XHS and IHS repoint, refine localization, characterize source
Digital forensics (Murtuza, 24 Jul 2025)	Preliminary evidence processing and prioritization	Human forensic analysis with approved tools
Behavior elicitation (Li et al., 3 Feb 2025)	Diverse proposal search and rubric-stage response exploration	Conditional inverse-prompting investigator model

Several further cases extend the pattern inward, into model internals rather than system-level orchestration. In “Flow CoT,” early latent reasoning iterations can be read as exploratory stages and later iterations as stronger refinement stages, although the paper is explicit that this reading is interpretive rather than official terminology (Li et al., 30 May 2025). Segment-compression SCOUT for long sequences admits a similar interpretation: local token mixing and checkpoint extraction act as a broad, cheap scouting pass, whereas sparse checkpoint attention acts as selective long-range investigation over compressed history (Jafari et al., 31 Aug 2025). These examples suggest that Scout-Investigator decomposition can operate not only across agents or tools, but also within a single model’s computational graph.

6. Limits, failure modes, and common misconceptions

A common misconception is that the architecture requires explicit physical or software separation into two named agents. The evidence is mixed. CATCH-PM is explicit, but robotic SCOUT, long-text SCOUT, and several language-model systems realize the same pattern through coupled modules, unified utilities, or staged latent computation rather than separately instantiated “scout” and “investigator” entities (Mao et al., 4 Jun 2026). The more precise criterion is functional specialization plus a disciplined handoff state.

A second misconception is that scout outputs are final decisions. In digital forensics, Scout is “merely a pre-analysis tool,” is “not involved in actual evidence detection, its extraction or its presentation,” and the remaining files “must also be analyzed by the forensic investigator” (Murtuza, 24 Jul 2025). A similar point holds in prompt-injection defense: the scout predicts detector reliability; it does not itself determine whether content is malicious (Zhang et al., 29 May 2026).

Current systems also expose recurrent technical limits. In robotic semantic exploration, uncertainty is node-centric rather than relation-centric; edges such as on, inside, belong, and next to are inferred from geometry but do not carry probabilistic beliefs, and dynamic-environment handling is not developed (Mao et al., 4 Jun 2026). In long-text understanding, the architecture achieves strong token efficiency but incurs higher wall-clock latency than single-pass models and does not reuse exploration state across queries (Zhang et al., 6 May 2026). In training-time poisoning defense, SCOUT functions as offline data purification rather than a deployed inference-time detector, so its scout role is restricted to dataset screening rather than continuous investigation (Afane et al., 10 Dec 2025).

These limits suggest a broader design lesson. A Scout-Investigator Architecture is strongest when the scout can cheaply surface structurally relevant possibilities, the investigator can reason over a compact auditable state, and the controller can trade safety, coverage, certainty, or cost through explicit thresholds or utilities. It is weakest when handoff state is noisy, when investigator competence is not modeled, or when role separation is nominal but does not materially change observation policy. As a general architectural form, its enduring significance lies in closing the loop between discovery and follow-up: the system does not merely gather data and then analyze it, but uses what it currently knows—and what it still does not know—to decide what to examine next.