This review was conducted as a scoping review based on the Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews (PRISMA-ScR) guidelines (Tricco et al., 2018), which are suitable when a scoping review aims to map out the conceptual landscape rather than to estimate a pooled effect. The four areas targeted were: (1) intelligence fusion and multi-source fusion; (2) formal uncertainty representation; (3) confidence calibration or state space tracking in analytical reasoning; and (4) structured analytical techniques relevant to multi-source assessment. The databases searched were Google Scholar, IEEE Xplore, ScienceDirect, SpringerLink, and Taylor & Francis Online. Search terms covered various combinations, such as “intelligence fusion,” “multi-source fusion,” “structured analytical technique,” “Dempster-Shafer,” “Bayesian fusion,” “confidence aggregation,” “weakest-link,” and “adversarial uncertainty.” The first search resulted in 312 hits. In all, eighty-nine papers were chosen for full-text review after screening of titles and abstracts. Following the full-text screening, forty-seven documents were retrieved for synthesis, which included references to multi-source uncertainty, confidence aggregation, or state space tracking in any tradecraft or security context. The research was not considered if it was limited to an environment where sensor fusion was used in non-adversarial engineering context. The PRISMA-ScR flowchart is included in Appendix A.7.

The review does not support absolute novelty claims. A scoping review maps the existing territory; it cannot exhaustively prove the absence of prior work. The claim this review supports is a relative contribution claim: within the reviewed corpus, no framework was found that simultaneously and explicitly integrates conservative weakest-link aggregation, hierarchical source compartmentalisation, and traceable assessment evolution in a formally specified intelligence-fusion context. The DRVF extends the existing literature by combining these three features in one formally specified, testable structure.

Analysis of competing hypotheses is widely cited in the methodological literature and familiar to many Western-trained analysts, although its operational prevalence varies across organisations and is difficult to establish from publicly available evidence (Heuer and Pherson, 2019). Software-assisted implementations of ACH now allow analysts to adjust source-validity weights interactively, which addresses some of the gap DRVF targets at the procedural level. The limitation relevant here is that ACH does not encode source-tier distinctions or confidence trajectories as algebraic properties of its structure. Mandel and Barnes (2014) found that structured methods improve forecast accuracy under controlled conditions, but that benefits are domain-contingent. This finding supports both the value of formal structure and the need to test structure against performance. A recurring limitation across the Scholastic Assessment Test (SAT) literature is the absence of a formal mechanism for tracking analytical state transitions. When source reliability shifts mid-analysis, ACH provides no structural constraint on how that shift should propagate through the combined picture. Recent works by Rieber (2004) on confidence-calibration methods and Chang, et al. (2018) on structured techniques in operational settings have advanced the field, but the specific gap that DRVF addresses the formal tracking of confidence trajectories under degrading source reliability remains.

Bayesian networks are powerful when priors are defended and dependency structure is stable (Hall and Llinas, 2001; Liggins et al., 2009; Pearl, 1988). Their application to multi-source intelligence fusion is well documented. Adversarial Bayesian formulations incorporate strategic manipulation of evidence inputs, and robust prior specifications reduce sensitivity to prior misspecification (Friedman and Zeckhauser, 2012; Muñoz-González and Lupu, 2017). These extensions represent genuine progress. Their limitation for the present purpose is not adversarial fragility per se but the combination of requirements: even adversarially robust Bayesian formulations do not simultaneously enforce a strict weakest-link confidence floor and preserve source-tier distinctions as algebraically separable compartments within one formally specified structure.

The DST represents epistemic uncertainty under incomplete evidence and is extensively applied in information fusion contexts (Castanedo, 2013; Dempster, 1967; Sentz and Ferson, 2002; Shafer, 1976). The tension with the present paper’s requirements arises from Dempster’s combination rule, which normalises conflict by redistributing mass away from the empty set, producing combined confidence estimates that exceed what the weakest credible source would justify. Both Murphy’s (2000) averaging rule and Smets’ (1990) conjunctive combination rule address conflict management without Dempster’s normalisation step, and Yager’s (1987) modification assigns conflicting mass to a global ignorance term. These are substantive improvements. Their limitation is the same as for adversarial Bayesian extensions: none enforces a strict source-tier floor as an algebraic property of the combination structure itself.

Dynamic epistemic logic provides formal tools for representing how belief states change after public or private updates (van Ditmarsch et al., 2007). DEL is conceptually proximate to the trajectory requirement in DRVF: it tracks how beliefs evolve, not merely what they are. Its limitation for operational deployment is practical. DEL models are highly expressive and formally rigorous, but adapting them to multi-source intelligence analyst workflows at the speed required by operational environments is not demonstrated.

Four conclusions emerge from the reviewed literature. First, structured analytical techniques remain operationally valuable but do not formalise confidence trajectory under degrading source reliability. Second, probabilistic frameworks provide mathematical rigour but rest on assumptions fragile in adversarial environments, and their adversarial extensions do not jointly enforce all three analytical requirements. Third, DST supports incomplete evidence but does not enforce a strict weakest-link floor as an algebraic property, even in its modified forms. Fourth, dynamic epistemic approaches represent state transitions formally but have not been operationalised for multi-source intelligence fusion. From these conclusions, the relative contribution of this paper can be stated: The DRVF extends the existing literature by integrating, in a single formally specified framework, (a) conservative weakest-link aggregation, (b) hierarchical source compartmentalisation, and (c) traceable assessment evolution. No framework combining all three features in an intelligence-fusion context was identified in the reviewed corpus. The synergy of the combination matters because each property, taken alone, is achievable with simpler structures; it is their simultaneous algebraic enforcement within one framework that constitutes the contribution. These distinctions are summarised in Table 1.

The problem class addressed by DRVF is a restricted one: multi-source intelligence fusion with (a) unstable or differentially degrading sources, (b) potential for adversarial manipulation of one or more evidence streams, and (c) need for an auditable record of assessment evolution. The three conditions that define this problem class are not exceptional; they describe the standard operating environment for all-source intelligence analysis in most major agencies. The lack of something is not an operational practice but a formal mechanism that makes some protection constraints structural properties and not procedural directions. Three analytical needs emerge from this problem class: (1) do not let weak sources be overwhelmed by strong sources (conservative aggregation); (2) do not compress sources into a single number (compartmentalisation); and (3) document the analytical state over time to allow for inspection of assessment evolution (trajectory tracking). The DRVF meets requirements (1) and (2) formally, and requirement (3) is met by the structured update protocol outlined in Section 5. The correspondence is not a necessary consequence but a deliberate option: other options are conceivable and might be just as successful with different assumptions.

Consider the following operational scenario, which serves as the worked example throughout this section. An analyst is assessing whether an adversary is preparing a coordinated cyber-disinformation campaign, a characteristic hybrid-crisis pattern in which cyber operations and information warfare converge. SIGINT provides technical indicators of preparatory network activity at high reliability (tier 1). A HUMINT source partially corroborates operational intent but is only partially validated (tier 2). OSINT provides abundant circumstantial evidence, but the domain is actively subjected to adversarial manipulation (tier 3). This pattern concurrent multi-domain operations, with at least one stream under active manipulation is typical of the hybrid-crisis environments that motivated the framework’s development. A standard averaging rule would allow a strong SIGINT signal to lift aggregate confidence above what the HUMINT and OSINT evidence justifies. DRVF is constructed to prevent that outcome.

The DRVF is a proposed method for combining intelligence from different sources that deliberately prevents combined confidence from exceeding what the weakest important piece of evidence allows. It also maintains information from different sources in separate inspectable compartments so that analysts can always see which evidence came from where and how the overall assessment changed as new information arrived. The DRVF is a formally specified framework for conservative multi-source intelligence fusion. It combines a filtered-ring algebraic structure, preserving source-tier distinctions as algebraically separable entries, with an assessment-tier aggregation rule, which enforces the weakest-link confidence floor. It is proposed as a candidate analytical tool, not as a universal replacement for the existing methods.

The term “formal” means that the framework’s properties, the conservative floor, the source isolation conditions, and the update rules are expressed as mathematical propositions with proofs, not as procedural guidelines. An analyst following the framework cannot inadvertently violate the conservative floor because it is enforced by the algebraic structure of the assessment tier. The term “conservative” means that the combined confidence estimate is bounded by the least reliable active substantive source entry. “Framework” denotes a structured set of integrated components, the filtered-ring representation, the assessment-tier rule, the source-mapping rubric, and the isolation criterion designed to work together as a coherent system for a defined analytical purpose. DRVF is not a decision-support system, a software tool, or an organisational process. It is a mathematical specification of how evidence combination should behave under stated conditions.

Let R be a non-zero, non-commutative ring equipped with a descending filtration {R_k} ≥ 0 such that R₀ = R, R_K+1 ⊆ R_K, and R_kR_m ⊆ R_k+m. The filtration is interpreted as a reliability ordering: lower-index elements represent higher-reliability propositions; higher-index elements represent weaker support. This is a modelling convention, not a theorem about epistemic states.

The filtered-ring structure is chosen over simpler alternatives for two specific reasons. First, the compatibility condition ϑ(R_k) ≥ k aligns algebraic operations with reliability tiers by construction: any product of ring elements automatically respects the reliability grading without requiring the analyst to recheck tier constraints at each step. An ordered set or vector representation does not provide this automatic propagation. Second, the matrix ring M₃(R) preserves internal distinctions between source streams, including corroborative and contradictory cross-stream relations, in a way that cannot be done by a scalar or vector representation. These two properties together, automatic tier-graded propagation and source-pair distinction, justify the algebraic overhead relative to simpler formalisms (Anjom Shoa, 2016).

Intelligence propositions are encoded as elements of the matrix ring M₃(R). The three dimensions correspond to three principal intelligence streams. In the worked example, these are HUMINT, SIGINT, and OSINT. This is a motivated modelling choice tied to the specific operational context; it is not a universal claim about source taxonomy. A primary source class, as used here, refers to a collection discipline whose reliability characteristics may degrade independently of other disciplines. SIGINT, Communications Intelligence (COMINT), and Measurement and Signature Intelligence (MASINT) may share collection infrastructure, and a compromise affecting that infrastructure could degrade all three simultaneously. The three-class assumption is a modelling simplification: it treats the intelligence problem as involving three principal reliability tracks. If a specific operational context requires finer-grained source distinctions, the matrix dimension should be revised accordingly (see Section 6.6 Boundary Conditions). Diagonal entries a_ii represent source-specific assessments for stream i. Off-diagonal entries a_ij represent corroborative or contradictory relations between streams i and j. A zero off-diagonal entry means no cross-validation has been established.

A formal framework that cannot be instantiated from operational data is a mathematical exercise, not a usable tool. The following rubric specifies how an analyst maps raw evidence to filtration levels. The rubric has three components: source reliability grading, corroboration weighting, and matrix entry assignment. Diagonal entries receive the filtration tier assigned to the corresponding primary source stream under the reliability grading above. Off-diagonal entries a_ij (representing cross-stream corroboration or contradiction) are assigned the filtration tier of the weaker of the two contributing pieces of evidence, consistent with the weakest-link principle for corroboration. A zero entry indicates that no cross-validation relationship has yet been established. All matrix assignments, together with the evidentiary basis for each tier, are recorded in the analytical log.

Source reliability grading. Tier 1 (filtration level 1): Source has documented operational reliability across prior assessments; intelligence has been independently corroborated by at least one source from a different collection stream; collection method is robust to adversarial manipulation under current conditions.

Tier 2 (filtration level 2): Source has partial reliability documentation; intelligence has been partially corroborated or is consistent with independent indicators; collection method may be vulnerable to limited adversarial manipulation.

Tier 3 (filtration level 3): Source has limited or unknown reliability; information from source is not corroborated by other streams or contradicted; collection method is known to be subjected to active adversarial manipulation.

Tier 4 (filtration level 4): Source has been discredited partially, collection channel has been compromised, or intelligence is inconsistent with independent indicators. Tier 4 movement is explicit analyst judgement and documented. In between cases, the case is rounded up using a conservative approach: if a source falls into two tiers, the higher filtration index (weaker reliability) is assigned. This rubric is planned to be assessed for inter-rater reliability in the confirmatory evaluation study and should be considered as a proposed instrument that needs to be validated prior to its confirmatory use.

Corroboration weighting. An off-diagonal entry between streams i and j is assigned at the filtration level of the weaker corroborating piece of evidence, not the average. If both SIGINT (tier 1) and HUMINT (tier 2) point to the same indicator, the corroboration entry is assigned tier 2. This is the rubric-level instantiation of the weakest-link principle for cross-stream corroboration.

Define a quasi-valuation ϑ: R → ℕ ∪ {∞} satisfying: ϑ(0)= ∞; ϑ(ab) ≥ ϑ(a) + ϑ(b); and ϑ(a + b) ≥ min{ϑ(a),ϑ(b)}. For a non-zero element a, ϑ(a) provides the algebraic infrastructure of the framework. For A = (a_ij) ∈ M₃ (R), define ϑ(A) = min_ij ϑ(a_ij), the quasi-valuation provides the algebraic infrastructure of the framework. The compatibility condition ϑ(R_k) ≥ k ensures that algebraic operations on ring elements are automatically constrained by their reliability tiers: any product of elements from R_k and R_k is guaranteed to lie in R_k+m, which means the tier information propagates correctly through addition and multiplication without requiring manual verification. The addition and multiplication properties of the quasi-valuation (proved in Appendix A.2) guarantee that the tier structure is preserved under evidence combination. These properties are the mathematical foundation that makes the framework’s structural constraints enforceable.

The weakest-link norm itself is implemented by a separate but related function. Define the filtration tier of a non-zero ring element a as τ(a)= k if a ∈ R_k R_k+1; define τ(0) as undefined (zero entries carry no tier information). For a matrix A ∈ M₃(R), define the assessment tier as:

T(A) = max{τ(a_ij): a_ij ≠ 0}

The maximum is taken over all non-zero, tier-assigned entries in the active matrix. This function implements the weakest-link aggregation norm directly: the highest (weakest) filtration tier among all active evidence entries determines the overall assessment tier. If any active entry sits at tier 3, the overall assessment cannot exceed tier 3, regardless of how many tier-1 entries are present. The assessment tier T(A) is bounded from below by the compatibility condition: since τ(a) ≤ ϑ(a) for any tier-assigned element a, we have T(A) ≤ maxϑ(a_ij) for all tier-assigned active entries. The quasi-valuation provides an algebraic upper bound on the assessment tier, and in the special case where τ(a_ij) = ϑ(a_ij) for all entries, the bound is tight.

The relationship between ϑ and T is worth stating precisely. The quasi-valuation ϑ operates at the level of individual ring elements and enforces the compatibility condition that makes the algebraic structure work. The assessment tier T operates at the level of the matrix and enforces the weakest-link norm that makes the framework conservative. Both are necessary: ϑ without T would provide algebraic infrastructure without a conservative aggregation rule; T without ϑ would provide a conservative rule without the algebraic guarantees that ensure tier information propagates correctly through operations on the matrix.

This conservative norm appears in intelligence doctrine under various formulations. The formal contribution is not the weakest-link norm itself (which experienced practitioners already recognise), but its formalisation in a structure that permits algebraic tracking, mathematical comparison across assessments, and formal testing of properties. A structure that enforces the norm by construction is harder to violate inadvertently than one that relies on analyst discipline alone.

A substantive objection to the weakest-link approach must be addressed directly. Converging independent sources can and do justify higher aggregate confidence than any single source alone. This is the logic underlying corroboration-based inference, and it is epistemologically sound under conditions of genuine source independence. If all five independent witnesses of varying reliability report observing the same event, it is reasonable to be more confident than the least reliable witness alone would justify. The reviewer who raised this point is correct as a matter of general epistemology.

The weakest-link norm in DRVF is not proposed as a universal epistemological principle. It is proposed as a domain-specific procedural constraint appropriate to a particular failure mode: the risk that coordinated deception or undetected source compromise inflates aggregate confidence by providing apparent corroboration that is in fact manufactured. In such environments, the independence assumption that justifies corroboration-based confidence-building does not hold. The five witnesses may be reporting the same story because the same deception operation has reached all of them. The conservative floor is a safeguard against this specific failure mode, not a general claim about how evidence should always be combined.

This distinction has a practical implication. In environments where adversarial manipulation is not a concern, where sources can be verified as genuinely independent, or where the cost of conservatism (missed signals and delayed judgements) outweighs the cost of overconfidence, other aggregation philosophies may be more appropriate. Bayesian combination, for example, is well suited to such environments. The DRVF is designed for the boundary case where the analyst has a reason to believe that one or more evidence streams may be compromised and cannot reliably determine which ones. This is not an unusual condition in intelligence analysis, but it is not a universal one either. The boundary conditions are stated in Section 6.6.

A substantive question raised in review deserves a direct response here: can the addition of weak but non-contradictory evidence rationally reduce the confidence warranted by an otherwise unchanged body of strong evidence? The answer that DRVF gives is conditional, and the condition turns on a distinction that the framework already encodes operationally but should now state explicitly. Weak evidence that fails a validity screen is dropped, not retained; weak evidence that passes the screen is retained at its assigned tier and bounds the aggregate. The conservative floor reflects residual deception risk that cannot be eliminated by scrutiny alone.

The two-stage gate operates as follows.

Stage 1 (validity screen): Before any piece of evidence enters the matrix, it is subjected to the standard validation procedures that experienced analysts already apply source triangulation, chain-of-custody verification, plausibility assessment against independent indicators, and explicit documentation of any doubts. Evidence that fails this screen is not entered into the matrix at all; it is logged in a separate excluded-evidence record with the reason for exclusion. This is the operational realisation of the reviewer’s instinct that weak or suspect evidence should be subjected to scrutiny and dropped if invalid. DRVF does not override this step; it presupposes it.

Stage 2 (tier assignment and floor enforcement): Evidence that passes the validity screen is assigned a filtration tier under a rubric in Section 4.4 and entered into the matrix. Once entered, the assessment-tier rule T(A) = max{τ(a_ij)} applies. The conservative floor at this stage is not a second validity judgement; it is a residual-risk bound. The screened-in weak evidence may still be reporting the same fabricated story that strong sources are reporting, because the screen cannot prove independence. The floor reflects the analyst’s inability to verify independence, not a verdict that the weak evidence is itself invalid. This is the substantive difference between “this evidence is suspect, drop it” (a Stage 1 outcome) and “this evidence is acceptable but of low reliability, let it bound the aggregate” (a Stage 2 outcome). The matrix structure of DRVF keeps these two judgements distinct, while a scalar weakest-link rule would conflate them.

This reformulation addresses the reviewer’s concern directly. The framework already provides for dropping invalid evidence: the rubric’s tier-4 designation and the source-isolation operation in Section 4.7 implement exactly this. What the framework does not do, and should not do, is allow weak-but-valid evidence to be silently overridden by stronger sources. The risk that weak evidence corresponds to a coordinated deception channel cannot be ruled out by scrutiny; it can only be bounded. The floor is that bound.

A substantive objection to the framework’s design must be addressed directly. In real intelligence analysis, if one source is compromised, experienced analysts do not collapse the entire assessment to the lowest available tier; they isolate the weak source and rely on stronger streams. A naive implementation of the weakest-link norm would preclude this, forcing analytical paralysis when any source is weak.

This objection has force against a scalar weakest-link rule. It does not apply to the matrix structure of DRVF. The matrix preserves source-tier information at the entry level. An analyst who identifies that the OSINT stream is compromised can exclude the relevant OSINT diagonal entry from the active matrix state and recalculate T over the remaining entries. This is a compartmentalised isolation operation, not a collapse.

Source isolation is not left to unconstrained analyst judgement. The following formal criterion governs when isolation is warranted: a source entry a_ij may be isolated if and only if (a) the entry has been assigned tier 4 under the rubric in Section 4.4, indicating partial discrediting or channel compromise, or (b) an explicit, documented evidentiary trigger, such as adversary-confirmed channel penetration or source recantation, has been recorded in the analytical log. Isolation on the basis of analyst’s personal preference or solely to raise the aggregate confidence floor, without a documented triggering condition, is impermissible within the framework. The algebraic treatment of source isolation, including the proof that T(A(–i)) ≤ T(A), is provided in Appendix A.4.

The update protocol is structured in such a way that it encodes the changes in the matrix state as new evidence comes in. The analyst will then do the following: (1) determine the target matrix entry associated with the relevant source stream; (2) apply a filtration level to that entry using the rubric in Section 4.4; (3) update the target matrix entry and any other matrix entries that would be impacted if the source stream were to be upgraded or downgraded; (4) recalculate the assessment tier T for all the entries that are assigned to it; and (5) record the type of evidence used (confirming the current tier, source stream upgraded, source stream downgraded, corroboration established, corroboration withdrawn, or assessment held pending), and the matrix state that results from the update in the analytical log. The log is the traceable history of the development of assessment. From the initial state of the matrix A₀ and the log, it’s possible to reconstruct the matrix state at any step by applying the sequence of operations that were recorded.

An optional visualisation of the update path was explored in earlier versions of this work. That visualisation used a hexagonal adjacency structure to encode six transition types as directions on a grid. Testing whether this visualisation reduces cognitive load or improves analyst performance is an empirical question. The visualisation has been moved to Appendix A.5 and is presented there as a provisional design, not as a required component of the framework.

The theoretical claim of this paper is that a filtered-ring, assessment-tier framework can formalise conservative intelligence fusion in a way that simultaneously enforces the weakest-link confidence floor and preserves source-tier distinctions algebraically. This claim is theoretical: the framework is formally specified, its key algebraic properties are proven in Appendix A, and a worked example demonstrates its behaviour under the specific scenario conditions for which it is designed. The claim does not extend to empirical superiority over ACH, DST, or Bayesian networks; that question requires controlled testing under the pre-registered protocol available from the corresponding author.

A formally specified framework is a necessary but not a sufficient condition for practical utility. Whether DRVF performs better than the existing methods under the conditions for which it is designed is an empirical question. This paper does not answer it. The contribution is complete in the sense that the theoretical claim that such a framework can be formally specified with algebraically enforceable properties has been discharged. The contribution is incomplete in the sense that the empirical question it raises has not been answered.

Triangulation and DRVF solve similar but different analytical issues. Triangulation is a validation technique the analyst examines whether other sources come to the same conclusion, and accepts the conclusion as valid unless it is found that the other sources converge. It generates a binary or categorical convergent assessment (convergent, partially convergent, and divergent). DRVF is a mechanism for confidence tracking: it is a record of the change of the aggregate confidence as the reliability of the sources changes over time and yields a graded, formally bounded estimate that can be updated incrementally.

Three specific differences distinguish DRVF from triangulation as conventionally practiced. First, triangulation does not encode source-tier distinctions; it treats convergent sources as mutually confirming without preserving their individual reliability levels as separable, inspectable quantities. DRVF maintains these distinctions in the matrix structure. Second, triangulation does not produce a confidence metric that can be compared across assessment cycles; it answers the question “do sources agree?” but not the question “how confident should we be, given how source reliability has changed since the last assessment?” DRVF is designed to answer the second question. Third, triangulation does not enforce a conservative floor; convergent sources can jointly produce a higher confidence estimate than any individual source warrants if the analyst is not explicitly guarding against manufactured corroboration. DRVF enforces this guard by construction. The two approaches are complementary, not competing: triangulation can operate within a DRVF framework as a corroboration-checking procedure.

The weakest-link rule is operationally naive. The most substantive objection to DRVF is that the max-based assessment tier imposes an operationally unrealistic constraint. If one source is weak or compromised, real analysts do not discard the entire assessment; they isolate the weak source and rely on stronger material. Applied naively to a scalar assessment, the weakest-link rule would force combined confidence to collapse to the weakest tier regardless of whether other streams provide strong independent support. This objection has force against a scalar weakest-link rule. It does not apply to the matrix structure as specified. As demonstrated in Section 5 (step S₃), the framework supports formal source isolation under a documented criterion: a compromised source entry can be excluded from the active assessment, and T is recalculated over the remaining entries. Whether analysts find the isolation operation natural and executable under time pressure is an empirical question for the evaluation protocol.

The mathematical formalism is unnecessary. Another objection is that the filtered-ring and assessment-tier structures add algebraic overhead without corresponding practical benefit. A weighted average with a min-floor or a simple fuzzy-logic system could perform the same function. This objection deserves a direct response. A weighted average with a min-floor satisfies the conservative norm on the aggregate but does not preserve source-tier information as algebraically separable entries: once the average is computed, the individual stream reliabilities are collapsed. The matrix structure keeps them distinct and trackable. The compatibility condition ϑ(R_k) ≥ k guarantees that if an element in a ring is reliable, any algebraic operation performed on it is automatically guaranteed to be reliable as well; this requirement does not need to be rechecked by the analyst at every step of fusion. The theoretical case is that these additional properties do not give rise to simpler structures, and thus are formally guaranteed; whether these additional properties are worth the cognitive overhead depends on empirical evidence. The simple-models literature (Dawes, 1979; Gigerenzer and Brighton, 2009) shared this concern and it is recognised as a legitimate challenge.

The formal specification is unnecessary complexity. The third objection is that intelligence analysts do not need formal mathematical specifications; they need clear procedures, good training, and sound professional judgement. This objection has practical force. Most of the intelligence organisations do not operate with formal mathematical frameworks and produce usable assessments without them. The response is that the DRVF is not intended to replace professional judgement but to supplement it with structural constraints that are difficult to enforce through procedural guidelines alone. Whether the benefit of these constraints outweighs the cost of the additional complexity is precisely the question the pre-registered evaluation study is designed to answer.

An exploratory pilot study (N = 23) was conducted to assess protocol readiness for the planned confirmatory evaluation. The pilot used intelligence-scenario materials and the source-reliability rubric described in Section 4.4. Participants were intelligence professionals with operational analytical experience. The pilot was not designed to test the framework’s performance against alternatives; its purpose was to identify whether the protocol, training materials, and scenario texts were sufficiently clear for participants to engage with the DRVF structure meaningfully.

The pilot produced the following qualitative observations, which should be interpreted with caution, given the small sample and the absence of inferential statistics. First, participants who received the standard training materials (a two-page written explanation of the framework) had difficulty applying the matrix structure to the scenario materials. This suggested that training time needed to be extended and that visual scaffolding (a prepared matrix template with pre-labelled axes) would be necessary. Second, several participants found the concept of “tier 4” and the source-isolation criterion confusing when presented in abstract form but understandable when illustrated with a concrete example. Third, the scenario texts were revised to make the evidentiary basis for tier assignments more explicit, after participants indicated that they were uncertain about which tier to assign to certain evidence items. Fourth, no participant objected to the conservative floor principle itself; the resistance was to the operational complexity of the matrix representation, not to the norm it enforced. These observations informed three protocol refinements incorporated into the pre-registered full-study design: extended DRVF training time, visual scaffolding requirements, and scenario wording revisions.

Cell sizes in the pilot (n = 7-8 per condition) preclude any inferential use. Inclusion of underpowered directional statistics in a theoretical specification paper risks creating a misleading impression of empirical validation. The pilot’s role was to identify protocol refinements, which it did. Pilot data and refinement details are available from the corresponding author on request.

The reviewer’s observation that the DRVF as specified resembles a Paris Fashion Week show, impressive on the runway but unobvious in its translation into everyday analyst practice, is well taken. The framework’s mathematical formalism is a specification, not an operational procedure. The question of how an analyst without graduate-level mathematics training would actually use DRVF in a working assessment deserves a direct answer, and the answer is that the operational pathway through the framework is narrower than the formal specification suggests.

There are three observations that will help to address the question of operationalisation. First, the algebraic proofs in Appendix A prove that the structural properties of the framework are true; they do not need to be checked by an analyst when using the framework. An analyst using DRVF does not have to reprove that T(A(–i)) ≤ T(A), whereas an analyst using a Bayes’ network does not have to reprove Bayes’ theorem. The mathematics is not something the analyst does in the analysis; rather, it is a guarantee on the structure. Second, the operational pathway through DRVF consists of three concrete steps: (i) assign a filtration tier to each piece of evidence, using the rubric in Section 4.4; (ii) place the tier in the appropriate diagonal or off-diagonal position in the matrix; (iii) read off T(A) as the maximum tier over the entries that are active. These steps are not symbolic algebraic steps but categorical judgements that have a rubric that has been documented. Third, software implementation of the matrix algebra is very simple: a small python or R script (less than 200 lines) can implement tier assignment, matrix update, T(A) calculation, and generation of audit-log. In addition to the confirmatory study, a reference implementation is published.

On the Science, Technology, Engineering, and Mathematics (STEM)-readiness question, the reviewer correctly notes that the UK intelligence analysts, on average, do not hold STEM degrees at the national average rate. This matters and is acknowledged. The UK Government Communications Headquarters (GCHQ) and the Joint Intelligence Committee (JIC) have historically trained analysts in structured analytical techniques, ACH, red team analysis, and key assumptions check without requiring university-level mathematics. DRVF’s operational demands are within the same range: applying the rubric is a categorical judgement; entering a tier into a 3×3 matrix is a tabular task; reading the maximum is a comparison. None of these operations requires the analyst to manipulate the underlying ring-theoretic objects. The cognitive demand is comparable to that of ACH’s matrix scoring, which is taught in introductory analyst training across North Atlantic Treaty Organization’s (NATO) member services (Heuer and Pherson, 2019). The pilot feasibility study (Section 6.4) confirmed that the principal obstacle was not mathematical capacity but familiarity with the rubric’s tier distinctions; participants who received extended training with concrete examples engaged with the framework successfully.

On time budget, a back-of-envelope estimate based on the pilot is as follows. Initial matrix setup for afresh assessment of three streams, initial tier assignments for 8-12 evidence items, rubric consultation, and audit-log initialisation takes approximately 25-35 minutes. A per-update cycle as new evidence arrives tier assignment for the new item, matrix update, T(A) recalculation, log entry takes approximately 5-10 minutes. Source isolation under the Section 4.7 criterion verification of trigger condition, isolation operation, log entry, and recalculation takes approximately 10-15 minutes. These estimates assume a trained analyst working without software support; a software-assisted implementation would compress per-update time to less than 2 minutes by automating matrix bookkeeping. The estimates are within the range that operational all-source analysts routinely allocate to ACH or Bayesian belief network updating for comparable scenarios.

The framework is therefore not a Paris Fashion Week artefact in any operationally relevant sense. The algebraic specification is rigorous because the structural guarantees it provides conservative floor preservation, source-tier compartmentalisation, and auditability need to be provable, not because the analyst must re-derive them. The operational pathway through the framework is a sequence of categorical judgements comparable in difficulty with the existing structured analytical techniques. What DRVF adds operationally is the conservative floor as a structural constraint that cannot be inadvertently violated, not a body of mathematical machinery that the analyst must master. The mathematical sophistication is in the specification; the operational practice is in the rubric.

A software reference implementation is in preparation and will be released as an open-source tool alongside the confirmatory study. The implementation will provide: a guided tier-assignment interface grounded in the Section 4.4 rubric; automatic matrix update and T(A) recalculation; a structured audit log; and an isolation-operation workflow that enforces the Section 4.7 criterion. The tool is intended to make DRVF usable for analysts without requiring any direct engagement with the underlying algebra. This is the practical operationalisation pathway, and it is the route through which the framework, if empirically validated, would enter working analytical practice.

The framework is valid under three assumptions: (a) the number of primary source classes is analytically meaningful for the context (the current specification uses three, but this is a modelling choice, not a universal claim); (b) conservative propagation is a normatively appropriate constraint for the specific failure mode being guarded against (manufactured corroboration in an adversarial environment); and (c) a record of assessment evolution is operationally useful. Outside these conditions, the framework should be modified rather than defended as universal. DRVF is not appropriate for stable multi-sensor environments where priors are well specified (Bayesian methods are better suited), for tasks where speed is paramount and cognitive overhead cannot be accepted, or for contexts with more than three primary source classes without architectural revision.

The author’s background as an operational intelligence and security professional informed the scenario design with ecological validity that would otherwise be difficult to achieve. It also created the risk that the problem was framed in ways that favour the proposed solution. The blind coding procedure, pre-registration, and open materials commitment described in the evaluation protocol were designed specifically to mitigate this risk. Future independent replication by researchers without practitioner backgrounds would provide the strongest available test of whether the architectural properties, not the author’s assumptions, drive any observed performance differences. All scenario texts, the rubric, coding instructions, and anonymised response data will be deposited in a public repository upon publication of the confirmatory study.

The framework’s motivating domain warrants explicit discussion. Hybrid crises, the concurrent deployment of military, cyber, information-warfare, and economic instruments by state or non-state actors create a specific and acute version of the general fusion problem that DRVF addresses. Three features of hybrid crises make them particularly demanding environments for multi-source confidence tracking. First, the number and diversity of relevant collection disciplines expand rapidly: a hybrid crisis may simultaneously require assessment of SIGINT, HUMINT, OSINT, financial intelligence, and geospatial intelligence, each with different reliability characteristics and different vulnerability profiles to adversarial manipulation. Second, the adversarial actor has an incentive and a capacity to manipulate specific evidence streams: disinformation operations directly target OSINT inputs, while cyber operations can degrade or fabricate technical indicators. Third, the tempo of operations in a hybrid crisis compresses the decision cycle, reducing the time available for careful source evaluation and increasing the risk that a single compromised stream will contaminate the aggregate picture through apparent corroboration. The two declassified episodes cited in Section 1, while predating the contemporary “hybrid warfare” debate, exhibit precisely these structural features: multiple collection disciplines, active manipulation of at least one stream, and time pressure that discouraged the type of slow and careful source-by-source evaluation that might have prevented the overconfident assessments that followed.

It is important to reiterate that DRVF is not specific to hybrid crises. The three structural conditions the framework requires unstable source reliability, adversarial manipulation potential, and auditability requirements are present in many intelligence contexts. The framework’s mathematical properties hold, regardless of the domain in which it is applied. Hybrid crises are the motivating domain because they concentrate all three conditions in a form that makes the need for a formal conservative mechanism most visible and most consequential. The NATO Hybrid Centre of Excellence (2022) has noted that strategic-communications challenges in hybrid environments are a recognised gap; the present framework addresses the analytical rather than the communications dimension of that gap, but the underlying problem distinguishing genuine from manufactured corroboration across multiple evidence streams is the same.