RESEARCH DOSSIER
Cyber Incident Forensic Causation and Evidence Preservation
Abstract. Cyber incident forensic causation investigates the mechanisms connecting digital system disturbance, operational effects and contractually attributable loss. This monograph integrates incident reconstruction, digital evidence preservation, causal inference, insurance contract interpretation, forensic accounting and claims data analysis. It examines official technical publications, an original cyber insurance specimen, selected judicial decisions, European operational resilience legislation, public incident investigation and scholarly research. The analysis develops an explicit distinction between integrity of a preserved record and its capacity to establish a causal proposition, between event time and contractual interruption time, and between economic damage and indemnifiable loss. A worked hypothetical business interruption example derives attributed and covered amounts under alternative timing and income assumptions. The synthesis identifies how evidence availability changes causal identifiability, while actual legal outcomes remain specific to issued policy wording and governing law. The study is a targeted interdisciplinary synthesis; it neither estimates a proprietary claims population nor asserts a new named theory.
Keywords: Cyber insurance; digital forensics; incident reconstruction; insured loss; evidence preservation; causal inference; timeline; business interruption; clock uncertainty; forensic readiness; coverage causation; counterfactual loss; cloud dependency; claims adjustment; operational resilience.
Introduction
The scientific object of cyber incident forensic causation is the reconstruction of the processes by which a digital disturbance produces consequences that can be identified, measured and evaluated under an insurance contract. Evidence preservation is considered here in relation to that object: it sustains the observations necessary to compare explanations of loss and to determine the time, system, entity and mechanism to which a consequence belongs. A reliable record of bytes supports this investigation only when the circumstances of generation, collection and interpretation establish what the bytes can legitimately demonstrate.
A claim presents at least three distinct questions. The technical question concerns changes in systems and the mechanisms that produced them. The economic question concerns the difference between the financial outcome observed after those changes and the outcome that would reasonably have occurred under a defined no-incident or alternative-incident scenario. The contractual question concerns which of those attributed consequences fall within the operative promise, including its terms, definitions, exclusions and applicable law. These questions interact throughout a claim but retain independent truth conditions. Demonstrated malicious access can coexist with a lack of material interruption. Proven interruption can coexist with exclusion of a particular expense. Payment can occur under a compromise even while an initial route of intrusion remains unresolved.
This dossier belongs to the Science of Cyber Insurance. Its boundary is the causal reconstruction of insured damage, with forensic integrity as a precondition of usable observations. The independent subject of legal admissibility, custody doctrine, and comprehensive evidence management belongs to the related dossier on digital forensic evidence integrity and chain of custody. Threat actor identification enters only where the identity or purpose of an actor materially changes a causal or contractual proposition. Claims concerning a specific organization require its policy wording, actual systems, insured entities, jurisdictions, operational records and financial data. This research does not possess those confidential records.
The research is a targeted comparative synthesis, not a systematic review claiming exhaustive retrieval of all literature. It uses selected official technical publications, a published insurance wording, two United States judicial opinions and one subsequent procedural order, EU operational resilience legislation, an official public event investigation, and selected published research on time reconstruction, causal reasoning and cyber incident measurement. The source ledger identifies the access level for each source and the proposition checked. Abstract-only material supports only the corresponding limited statement. The insurance example in the later chapters is constructed to make the calculations inspectable. Its numbers have no empirical or underwriting significance. The cutoff for source examination is November 18, 2026. The comparative legal design establishes transferable distinctions without inferring uniform law across jurisdictions.
Conceptual and Historical Foundations
From incident identification to forensic causation
Digital forensics developed around the problem of recovering interpretable traces of actions from information systems. An insurance investigation integrates the same evidentiary problem with an additional obligation to explain financial harm. Kent et al. (2006), in their forensic techniques guide, identify multiple computer and network evidence environments, including operating systems, applications, traffic and files. Their IT orientation matters to insurance research because it accommodates both hostile incidents and operational problems. The existence of a malfunction belongs within the investigation from its outset; a destructive event need not be treated as hostile before the evidence supports that conclusion.
The Internet Engineering Task Force's evidence collection guidelines (Brezinski & Killalea, 2002) document early operational concerns about volatile evidence, analysis copies and traceable handling. The historical importance of this document lies in its articulation of evidence loss as an investigation problem. Its examples reflect the technical environment of 2002; neither its example utilities nor its hash algorithm examples establish a contemporary default. Kent and Souppaya's (2006) log management guide moves the focus toward programmatic enterprise logging. The historical trajectory illustrated by these publications is from episodic collection toward preparedness across the system lifecycle. This is a selective documentary history, not a claim about the earliest invention of digital forensics.
The April 2025 incident response community profile (Nelson et al., 2025) places response within cybersecurity risk management and supersedes the 2012 revision. It provides an institutional setting for integrating investigations with recovery and prevention. For insurance, the integration has two implications. First, investigators are often studying a system that is changing because responders seek to protect customers and restore service. Second, the evidential capacity available at the time of loss partly reflects prior investments in observability and response planning. Insurance risk engineering therefore has a legitimate interest in both reduction of actual damage and reduction of avoidable uncertainty about damage.
Units of observation and units of decision
An artifact is a recorded object or state from which a proposition about an event may be inferred. An event is a specified change or action in a bounded system. A causal pathway is an ordered relation connecting changes through mechanisms that may be corroborated, challenged or qualified by observations. A loss item is a measured consequence assigned to an entity, category, accounting period and causal explanation. A coverage decision applies a particular contract and law to a set of established or stipulated facts.
Each unit can be many-to-many. One artifact may bear on multiple events; one event may produce many independent artifacts. Several technical events may constitute one insurance occurrence, or a single technical event may produce claims under several policies. A decision about occurrence count is therefore not resolved by counting alerts. Two notifications about identical activity may indicate duplicate telemetry rather than two attacks; two separate organizations may lose income as a result of a shared provider incident while each has a separate retention. Technical incident identifiers and policy occurrence identifiers should remain independently recorded.
Distinctions within causal language
The origin of a vulnerability, the event that activated it, the condition that enabled propagation, the mechanism of disruption and the measure of consequent financial harm answer different questions. The label root cause may be useful in a post-incident report if the purpose of the analysis is explicit. A patch defect can explain susceptibility; an unauthorized action can explain exploitation; the configuration of dependencies can explain propagation; an isolation decision can explain an interval of unavailability. More than one explanation may correctly describe different links of the same chain.
A necessary condition is one without which the specified consequence would not have occurred under the assumed background conditions. A sufficient set is a collection of conditions under which the consequence occurs within the stated model. A causal contribution may change the duration or extent of harm without being necessary for its existence. A background condition may explain the vulnerability of a system without satisfying the contract's peril definition. The analytical vocabulary should preserve these distinctions before legal causation is considered.
The methodological value of the distinction lies in the precision of the question asked of a forensic report. A report that calls absent multifactor authentication the root cause may be answering a preventive security question. The insurer may instead require evidence that access associated with a particular account caused corruption of an insured database, or that a later shutdown was undertaken in response to that access. The answer to the preventive question cannot establish these downstream relations by semantic substitution.
Forensic Integrity
Integrity and the evidential meaning of a trace
A forensic object supports causal analysis only through the proposition it is capable of bearing. A validated digest of an exported audit file establishes consistency of the compared bytes with the recorded digest, conditional on a sound process and trustworthy reference. It cannot alone establish that the service initially recorded a true event, that the file contains all relevant events, or that the user associated with an identifier is the person who performed an action. Integrity of preservation and reliability of observation must be assessed separately.
Guttman et al. (2022) identify the distinctive difficulties of digital evidence preservation for media and digital objects. The present inquiry uses that observation to ask which facts become unrecoverable when specific traces expire or when analysis alters their metadata. This is particularly consequential where operational restoration overwrites the affected environment. Acquisition plans should identify the inferential question for each object: authorization, execution, sequence, affected data, service interruption, mitigation, or financial consequence.
Different sources have different functions. Identity logs may establish an authentication or authorization event according to a provider's documented semantics. Endpoint observations may show a process or a file change. Application transaction records may establish whether a customer operation reached a defined completion state. Network observations may show communication between endpoints but lack access to encrypted contents. Configuration history may establish that a control changed. A support ticket may document what a responder observed or believed, without directly establishing the event described. Financial records may show expenditure, not its technological cause. Preservation must retain the relationships among these observations rather than simply maximize the number of stored files.
Preservation decisions and loss of identifiability
Brezinski and Killalea's (2002) order of volatility gives the investigation a defensible starting point: evidence susceptible to disappearance receives priority, with analysis conducted on separate copies where applicable. The precise sequence depends on the environment and on competing continuity and safety objectives. Live acquisition itself may change the state of a system; physical shutdown may terminate both malicious activity and the surviving record of it. In cloud environments, comprehensive storage copying may be unavailable to a tenant. A defensible contemporaneous record identifies the decision, the responsible role, the relevant time, what could be obtained and what risk made other acquisition impracticable.
A preservation decision should therefore be explained by reference to the consequences of missing evidence. If a host is reimaged before investigators retain execution records, the claim may still support interruption through independent transaction failure logs. The route from an external communication to execution may remain unresolved. If the existence of a covered interruption is independently established and the unknown route does not affect any material term, uncertainty is confined to mechanism. If a particular exclusion requires distinguishing unauthorized access from an authorized configuration change, the same missing record might be decisive.
Cloud evidence and distributed authority
Herman et al. (2024) describe cloud forensic readiness as a purpose of the NIST reference architecture. The insurance interpretation is that forensic capacity follows the division of operational authority. A tenant may observe its application and identity activity while the provider controls infrastructure telemetry, internal key-management history, service boundaries and aspects of incident recovery. Shared tenancy also imposes confidentiality limits on what can be disclosed. A provider's assertion and a tenant's observation should each be attached to their originating perspective.
Preincident arrangements can clarify which records are generated, which parties retain them, whether their availability varies by service tier, which time bases are used, how an incident export can be requested and what redactions apply. These arrangements influence the possibility of reconstructing causation. They also affect the expected costs of investigation and the time required to resolve disputed indemnity.
A provider's status notification that a service suffered disruption may support a common-event hypothesis. It does not establish that all customers lost income. Customer-level dependence, failover capacity, use of the service, geographic routing and alternate workflows explain differences in consequences. Evidence from the service level and evidence from the insured level should therefore be recorded as linked observations at different levels of analysis.
Preservation of financial and operational originals
A cyber claim may require preservation of order records, settlement files, production schedules, inventory movement, payroll, invoices, customer refunds, service-level data and documented recovery actions. Their significance arises from their temporal and functional relationship to the technology. An invoice for emergency support may cover restoration of affected systems, unrelated replacement of old equipment, post-incident improvements and routine work within the same billing period. The cost must be decomposed by activity where material. A report assigning the full invoice to the cyber event without activity-level support introduces uncertainty into the financial reconstruction regardless of the fidelity of the original invoice.
Financial data can also be affected by the incident itself. A damaged accounting database may contain missing or duplicated entries, and an offline process may create transactions that are posted only after recovery. Consequently, the investigator should compare independent operational measures with reconciled financial evidence and document the completeness of the resulting reconstruction. The preservation of raw conditions supports later correction when new records become available.
Absence, contradiction and corroboration
A negative finding is meaningful when the investigation can establish the detection capability relevant to the proposition. Absence of an entry for a particular action has different significance when logging was enabled, functioning, comprehensive for that action, and retained across the relevant interval than when the event was outside the scope of the log. A statement that no evidence was observed should identify the source and interval searched. A statement that an event did not occur needs an additional account of why the record would reliably have captured it.
Corroboration requires attention to lineage. A dashboard alert, an analyst report summarizing that alert, and a ticket created from the report may be three records derived from one observation. The evidential gain from their agreement is smaller than the gain from concordance with an independently generated application transaction failure. Contradictions should remain in the research record; resolving them requires evidence or a justified reinterpretation of their meaning. Deleting an inconsistent row from a reconstructed timeline because it conflicts with the favored narrative converts a test of the narrative into support manufactured by selection.
Incident Timeline
Event time, record time and business time
The timeline of a cyber incident is a structured estimate of the ordering and timing of relevant events. Its temporal units are defined by the decisions that depend on them. A clock reading, an event's occurrence, a log's reception by an aggregator, a responder's first awareness, a business interruption and a policy notice may all have different times. Combining them into a chronology without specifying their meanings makes apparent contradictions and artificial precision likely.
A minimal event history distinguishes an earliest evidenced adverse action; the earliest documented material operational effect; the first indication recognized by the insured or a service provider; the initial response; the point at which propagation ceased or was reasonably contained; partial and full functional recovery; notice to the insurer; and later discovered consequences. The earliest evidenced action is an observation boundary rather than necessarily the real start of the incident. The point of discovery depends on the policy's language and the persons or entities whose knowledge is relevant. Restoration of technical availability may precede or follow restoration of ordinary economic activity.
Vanini et al. (2024) examine timestamp interpretation through time anchors. Their repository abstract describes clock skew, malfunction and tampering and illustrates the use of anchoring events to evaluate clock correctness. This supports the requirement that investigators consider the origin of reported times. The following interval arithmetic is an original instructional application of elementary measurement uncertainty; it is not a claimed method or result of that paper.
Clock normalization and partial order
Let s be a timestamp reported by a source, let b be its estimated clock offset relative to a common time base (source clock minus reference clock), and let u be a bound on the remaining error, including resolution and synchronization uncertainty where relevant. A simple normalized event interval is [s - b - u, s - b + u]. The model assumes an offset that is sufficiently stable over the interval under consideration. A clock that steps, drifts materially or changes timezone requires a segmented or wider representation. For a timestamp that records ingestion rather than occurrence, delay uncertainty must be introduced separately.
Consider two synthetic observations. A host reports 10:04 UTC, with its clock estimated to be two minutes fast and an uncertainty of one minute. Its corrected interval is 10:01 to 10:03 UTC. A gateway reports 10:03 UTC with zero estimated offset and a two-minute uncertainty, giving 10:01 to 10:05 UTC. The overlap prevents strict ordering on clock values alone. A reliable record demonstrating that the gateway explicitly acknowledged an earlier host request could establish order without fixing an exact instant. Absent such evidence, the sequence remains partially ordered.
Time-zone labels belong in the record. A converted UTC timestamp should be accompanied by its original literal value and the conversion assumptions, including daylight-saving ambiguity where applicable. A logging pipeline that truncates milliseconds can move events into the same reported second. A sequence counter or known application dependency can preserve relative order despite that loss of resolution. Investigators should represent unobserved intervals and time ranges explicitly rather than convert each observation to one invented point.
Building and challenging causal sequences
A chronology supplies the necessary setting for comparing mechanisms. Consider an insured application that stops processing payments after an identity alert. The alert might correspond to compromised access that led to a configuration change. It might also reflect ordinary administrative work occurring during a separately scheduled deployment. The recorded proximity of the alert to the outage is consistent with either scenario. Evidence about the actual configuration state, change authorization, deployment timing and app-level failure mechanism is required to discriminate between them.
Represent each proposed causal connection as a question with expected observations. If a configuration change disabled database access, transactions using that dependency should show an appropriate failure state or interruption associated with the change. If the underlying database failed independently, symptoms may precede the configuration action or affect a broader set of clients. If a responder deliberately isolated the system after the first alert, pre-isolation and post-isolation failures should be distinguished. The purpose is not to search only for corroboration, but to find observations capable of making an alternative explanation less plausible.
The chronology should retain provenance at every edge between events: source identifiers, the event semantics, time bounds, the reason for assigning an order, the competing interpretations and their relevance to loss. This is a scientific record of reasoning, not a claim that a timeline graph automatically proves a cause.
Public case: the Summer 2023 Exchange Online intrusion
The Cyber Safety Review Board's (2024) Exchange Online intrusion review describes a compromise of cloud mailboxes in May and June 2023, anomaly detection on June 15 and associated alerts on June 16 involving audit information available to the detecting organization's enhanced logging service. The review also reports that Microsoft had not determined how or when the signing key had been obtained as of the report. The statements describe the public report's evidentiary state at its date; they do not establish the state of subsequent investigations. The report is a public incident reconstruction, not an insurance claim file.
The case illustrates separation between an observed effect and knowledge of its antecedents. A system can record unauthorized access in a way that provides useful evidence of compromise while leaving the acquisition route of an enabling key unknown. Customer observability depends on the records available to that customer. Multiple affected organizations can share a technological antecedent without having identical records, business interruption, liability or insurance recovery.
The insurance relevance is conditional. In a hypothetical privacy or incident-response claim, documented access may support an investigation expense or an affected-data inquiry if the corresponding insuring agreement applies. A business interruption claim would additionally require evidence that defined operations were interrupted and income was lost. The publicly established existence of the intrusion cannot supply any of these missing private policy facts. It would be an evidentiary error to treat the public report's affected population as a population of insured claims.
Causal effects of response and recovery
An incident timeline is altered by interventions made in response to the incident. Containment can prevent subsequent damage but impose immediate operational interruption. An emergency failover may reduce lost income yet create incremental costs. Restoring a service in a restricted state can reduce disruption before full capacity is reached. The investigation should attribute observed operational changes both to the original incident and to documented interventions, with dependencies between the two recorded.
The 2026 ransomware risk management profile (Souppaya et al., 2026) distinguishes the encryption of information from theft-related extortion and covers readiness, response and recovery. This distinction is useful for separating evidence of interruption from evidence of disclosure. Encryption may produce a loss of availability; theft may produce privacy or confidentiality consequences without a material outage. A single extortion episode can implicate several different loss categories. Evidence of an extortion demand is relevant to threat and response, but the claimed facts about the underlying theft or destruction require independent assessment.
Restoration should be reconstructed as a changing performance state rather than as a single timestamp if capacity returns gradually. Technical uptime, ability to complete business processes, backlog clearance and financial recovery are separate measures. A forensic timeline linked to operational throughput can distinguish the duration of material interruption from the later period of impaired demand or deferred transaction processing.
Coverage Causation
Translating evidence into the operative insurance contract
Contractual causation is a rule-governed evaluation of the relationship between established facts and the insured peril, loss, conditions, exclusions and limits. The contract determines the legal relevance of a technical event. A claim file must identify the policy period, named insureds, insured interests, operative declarations, applicable endorsements, definitions, insuring agreements, exclusions, conditions, aggregation provisions and limits. Evidence about a provider outage must be evaluated against any dependent interruption wording rather than simply against first-party interruption terms.
The scientific method is to formulate propositions at a level specific enough for legal evaluation: an identified system was materially interrupted; its service to an insured operation fell below the defined standard; that interruption was attributable to a particular kind of event; loss accumulated during an evidenced period; the policy assigns that loss to an applicable module. Each proposition needs support appropriate to its subject. A technical expert can explain system behavior and evidence. An economist can reconstruct lost contribution and incremental expense. The interpretation of ambiguous language and the applicable legal standard require separate legal analysis.
Original specimen wording: interruption and dependent interruption
The 2022 CyberAcuView insurance wording provides a concrete primary contractual example. Its first-party business interruption module links income loss and defined extra expenses to material interruption resulting from a security breach or system failure. Its restoration-period definition uses restoration to prior functionality and service with reasonable speed, subject to a 180-day maximum. The waiting period starts at material interruption. A separate dependent business interruption module addresses specified provider systems. The published wording is an adaptable specimen, not evidence that a particular insurer issued these terms.
The example establishes the evidentiary importance of distinguishing initial compromise from material interruption. The insurer's financial obligation depends on the defined interruption and the contract's time and financial parameters, not merely on the first intrusion timestamp. A forensic conclusion of infection at midnight may be entirely compatible with the commencement of a waiting period at noon, when a previously functioning business system becomes materially interrupted. The same wording contains specific exclusions concerning provider events and voluntary intentional shutdown. Consequently, the reason, scope and timing of isolation matter; the research cannot infer that all reasonable containment actions are insured under this specimen. Issued wording, endorsements and applicable law govern that question.
The specimen's proof-of-loss and income calculation conditions make financial context part of the causal inquiry. They bring estimated performance without interruption, continuing expenses and subsequent recovery of income into the evaluation. The general lesson is that the financial counterfactual must correspond to contractual valuation rules. Two otherwise similar firms can have different insured loss for the same duration of outage because their policy wording, margins, mitigation, waiting periods and retentions differ.
Direct loss and intervening actions: Indiana
In G&G Oil v. Continental Western (2021), the Indiana Supreme Court considered a ransomware loss claimed under a commercial crime computer-fraud provision. The insured made a payment in bitcoin to regain access to its systems. The relevant language required loss directly resulting from computer use to fraudulently cause a transfer. The court found the direct-result requirement satisfied but concluded that neither party was entitled to summary judgment; the matter was remanded. The opinion does not establish that every ransomware ransom is covered. Its ruling concerns the particular provision and procedural record.
The analytical importance is the treatment of human intervention within a causal chain. The insured's decision to make a transfer followed the disabling of its systems and consultation with others. Whether the decision breaks or preserves a legally relevant causal connection cannot be determined merely from the existence of a human action. The policy's wording and the governing legal test decide how that sequence is evaluated. The case also demonstrates why decision makers should keep separate the evidence supporting an event's occurrence, the interpretation of one contractual element, and the resolution of all other elements necessary for a final indemnity decision.
Exclusion causation and actor attribution: New Jersey
The Appellate Division in Merck v. ACE American (2023) examined the application of a particular hostile or warlike action exclusion in all-risk property policies to losses caused by the 2017 NotPetya malware. It affirmed the finding that the insurers had not established application of that exclusion under the circumstances. Its reasoning turned on the wording and the historical context of an exclusion associated with military action; the court did not announce the full scope of every potential cyber operation under war exclusions. The subsequent New Jersey dismissal order (2024) dismissed the Supreme Court appeals with prejudice following a stipulation. It is a procedural order, not a further merits interpretation.
The case isolates the analytical gap between evidence of malicious origin, evidence of a relationship to a sovereign, and the legal meaning of a policy exclusion. A forensic investigator may resolve technical attribution while the applicability of the exclusion remains a contractual issue. Conversely, contract language may define an excluded event through an attribution procedure that requires information beyond the insured's logs. Modern cyber-specific endorsements cannot be treated as replicas of the property wording examined in Merck. The insurer must demonstrate whichever legally relevant propositions the actual terms require.
Together the Indiana and New Jersey opinions show that a chronological technical narrative cannot replace contract interpretation. They are selected United States cases illustrating distinct issues, not a representative comparative sample of all courts or a statement of universally binding rules. Neither opinion provides evidence about the frequency with which coverage disputes are settled or about ordinary ultimate payout amounts.
Concurrent mechanisms and allocation of shared harm
Multiple causal pathways can contribute to one outage. A compromised account and an independent equipment failure might each have the capacity to disrupt the same service. A defective update may trigger a fault whose duration is extended by the failure of a backup system. A provider incident may affect one operation while a local containment action affects another. The causal problem is to identify whether the conditions operate as alternatives, interactions, sequential contributors or simultaneous sufficient causes.
A conventional counterfactual contrast may fail to assign a unique fraction of loss in a case of overdetermination. Suppose two independently sufficient disturbances would each have stopped the same operation for the first hour. Removing either disturbance individually leaves the hour of downtime unchanged. Yet both disturbances contributed to the actual state of the system. Allocating half of the loss to each would be an additional convention rather than a consequence of the observed facts. Where one event is covered and another excluded, the operative law and contractual causation language are required to make the allocation decision. An investigator should report the feasible causal histories and avoid assigning an unsupported universal percentage.
An insurer may need to distinguish the consequences of a qualifying cyber event from physical damage, routine maintenance, preexisting impairment and third-party services. Even when one activity addresses several problems, the expense can sometimes be attributed through its incremental purpose, alternative costs and the extent to which it accelerates recovery. If the evidence cannot support a unique allocation, the range of plausible values and the contractual consequences of each should be shown. Precision unsupported by records is not a scientific solution to an allocation dispute.
First-party loss, third-party liability and recovery
First-party loss concerns the insured's own loss under the applicable promise, whereas liability coverage requires a separate examination of the insured's obligations toward others, the relevant claim or circumstance, the conduct or event giving rise to liability, and the costs and damages recognized by the policy. Evidence that data was accessible can be relevant to a privacy claim without proving that every person suffered a compensable injury. The affected population, applicable obligations, actual claims and recoverable categories must be assessed under the law and wording in force.
Subrogation or other recovery rights may make evidence of third-party responsibility important even when coverage of the insured's own loss can be determined without resolving negligence. Documentation of contractual responsibilities, causally linked losses and mitigation costs supports later allocation or recovery. Expected recoveries should be distinguished from realized recoveries. A disputed vendor indemnity is not money already received. Careful records prevent counting the same loss as an insurer indemnity, a vendor reimbursement and a separate unrecovered economic loss.
Regulatory records and coverage boundaries
In the European Union, the original text of Digital Operational Resilience Act (2022), Article 17, requires in-scope financial entities to record ICT incidents and significant cyber threats and establish processes for identifying and addressing underlying causes. Article 18 describes impact criteria including downtime, data consequences and economic impact. Article 19 concerns reporting major incidents; Article 13 provides for review after qualifying disruptions of core activities, including forensic analysis where appropriate. Article 11(8) addresses accessible continuity and recovery activity records.
These statutory provisions are relevant to the operational resilience of financial entities within the Regulation's scope; they do not define coverage written by insurers for customers. The analysis uses the original enacted articles, not a comprehensive review of subsequent technical standards, exemptions, amendments or national implementation. A supervisory classification and a private insurance trigger may intersect but have different legal purposes. A regulator's incident record may become evidence for a claim without determining the claim's contractual outcome.
Evidence Validation and Sensitivity Analysis
Causal inference and identifiable questions
The causal inference literature gives a disciplined vocabulary for separating observation from intervention and for stating assumptions. Pearl's (2010) introduction to causal inference emphasizes structural reasoning, conditional assumptions, interventions, counterfactuals and mediation, as summarized in the accessible abstract. Hernan and Robins (2020) organize Causal Inference: What If into methods of increasing complexity, including longitudinal analysis. These references motivate cautious use of conventional counterfactual reasoning here; the formulas below are standard mathematical illustrations and do not claim to reproduce the results of either work.
Define an outcome Y as lost contribution in a stated interval for an identified insured operation. Let Y(1) denote the outcome under the incident mechanism of interest and Y(0) denote the outcome under a specified alternative, usually absence of that mechanism while other relevant background conditions are held fixed. The contrast Y(1) - Y(0) is the attributed effect under those assumptions. For a single realized incident the investigator observes at most the realized outcome, and estimates the relevant alternative using other evidence. The definition alone supplies no estimator.
The alternative scenario should be operationally meaningful. It might preserve a scheduled maintenance period, seasonal demand changes, known unrelated equipment failures and preexisting capacity constraints. Removing the attack while assuming all other impairments also disappear would inflate the attributed effect. Conversely, post-incident demand changes caused by the attack may belong to the causal chain; controlling mechanically for them could obscure an effect. The analyst must identify which outcome and which contrast the insurance wording requires.
Causal diagrams can help explain whether a variable is a background cause, mediator, joint effect or independent disruption. For example, a major incident can cause both a long outage and extensive investigation. A naive association between investigation expense and outage duration cannot establish that investigation causes the outage. Severe incidents may generate both outcomes. Restricting analysis to claims referred for expensive investigation can change the distribution of observed severity and mechanism. The relationship between available evidence and the event studied is part of the observation process, not a neutral sample.
Validation of the observation process
Represent R as an indicator that the relevant trace is retained and accessible. Observations conditional on R = 1 need not represent all events if R depends on the incident, the organization or the severity of its effects. An attacker or a recovery process may cause loss of records precisely where the causal event occurred. Logging tier, capacity and maturity also influence observation. This possibility is an instance of informative missingness. Without an adequate observation model the proportion of confirmed mechanisms among documented claims cannot be interpreted as the population proportion.
An investigation should state, for each hypothesis, which sources are expected to record its effects, which sources were actually collected and searched, and what limitations govern the interpretation of missing items. Independent lines of corroboration deserve explicit consideration. When two explanations remain consistent with every accessible observation, the evidence establishes a set of plausible histories, not a unique history. Further collection is justified if a plausible additional observation could change a material decision.
A validation exercise should distinguish reproducibility of extraction, reliability of inference and robustness of the coverage outcome. A second analyst may reproduce an artifact's extracted timestamp while disagreeing about its semantic meaning. Both analysts may agree on the technical history while disagreeing about an economic baseline. Two competing histories can yield the same contractual outcome. Such cases are analytically important: resources spent resolving a technologically interesting uncertainty may have little effect on the specific insurance decision, although the information may remain valuable for recurrence prevention or third-party recovery.
Worked hypothetical reconstruction
The following example is entirely synthetic. It concerns one insured retailer, a single hypothetical first-party interruption policy, one transaction service, amounts in US dollars and an interval measured in hours from zero. It assumes a verified qualifying security event and an independently documented scheduled maintenance interruption. It ignores tax, salvage, liability, dependent interruption, policy aggregate exhaustion, coverage disputes about deliberate shutdown and other modules. These assumptions define the exercise rather than assert actual specimen or market terms. It uses an eight-hour straight waiting period that excludes income loss in the first eight hours from the commencement of the qualifying interruption; a single $20,000 monetary retention then applies to the combined eligible loss, followed by a $150,000 payment limit.
Assume the unimpeded contribution margin is $10,000 per hour, net of variable expenses. The actual operations lose all of that contribution from hour 0 to hour 12, half from hour 12 to hour 20, and none thereafter. Independent scheduled maintenance would have stopped operations without the attack from hour 4 to hour 6; no deferred sales occur in the initial calculation. The total observed reduction relative to uninterrupted operations is 12 x $10,000 + 8 x $5,000 = $160,000. In the no-attack scenario, the scheduled maintenance independently causes 2 x $10,000 = $20,000 of loss. Attributed income reduction over the whole interval is $140,000.
The waiting period ends at hour 8. Between hours 8 and 12 all contribution is lost, totaling $40,000. Between hours 12 and 20 half is lost, totaling $40,000. Scheduled maintenance is wholly within the waiting period, so it does not reduce the $80,000 post-wait attributed income loss again. Assume $15,000 of documented extra expense incurred after hour 8 is eligible and that eligibility is fully established under the hypothetical contract, including any expense cap. The combined eligible amount is $95,000. Applying the retention gives a hypothetical payment of $75,000, below the limit. The whole-interval attributed loss of $140,000 is thus distinct from the total eligible amount and the final payment. If the $15,000 is included as incremental economic cost, the corresponding whole-interval economic harm is $155,000.
For a time sensitivity, keep the operational impairment fixed but assume the materially qualifying interruption can be established only from hour 2. Under an otherwise identical policy the waiting period ends at hour 10. Income loss in hours 10 to 12 is $20,000 and in hours 12 to 20 is $40,000; with the same eligible expense and retention, payment is $55,000. Changing the qualifying start from hour 0 to hour 2 changes the payment by $20,000. The example isolates sensitivity to contractual timing; it does not claim the uncertain start actually has these values in a real claim.
For a baseline sensitivity, return to the hour-0 start and reduce the expected contribution margin to $8,000 per hour. Assume the percentage impairments remain identical and the extra expense remains eligible at $15,000. The observed reduction is $128,000; maintenance accounts for $16,000; whole-interval attributed income reduction is $112,000. After the eight-hour wait the eligible income loss is $64,000 and payment is $59,000. If later performance demonstrates $10,000 in recovered contribution that is attributable entirely to the covered post-wait segment and the assumed contract deducts it, the original $75,000 payment falls to $65,000. Recovered contribution must be subtracted once, from the appropriate loss segment.
These results may be reproduced directly from the stated intervals, rates and contractual operations. They illustrate why time measurement, counterfactual baseline, expense eligibility and financial terms must be considered jointly. A claim consisting of $160,000 in recorded interruption reductions cannot be adjusted accurately by attaching a generic coverage percentage. The relevant economic and contractual operations have an order and depend on evidence at different levels.
Probability and limits of inference
Let H1 and H2 denote two mutually exclusive and, for the simplified exercise, exhaustive causal hypotheses. If defensible prior odds and conditional evidence probabilities can be specified, posterior odds equal prior odds multiplied by the likelihood ratio P(E given H1) / P(E given H2). This is ordinary Bayes' rule. For an illustrative prior odds of 1 to 1 and a likelihood ratio of 4, posterior odds are 4 to 1 and the posterior probability of H1 is 0.8 under those assumptions. This number does not represent a legal standard or an empirically calibrated forensic confidence.
If two evidence items are dependent because they are generated by the same sensor pipeline, multiplying their marginal likelihood ratios as if independent can overstate support. If the investigator cannot defend the prior or calibrate the conditional likelihood, numerical confidence should not be fabricated. Qualitative descriptions of observed, strongly supported, compatible, contradicted and unresolved propositions can be useful when accompanied by explicit reasons. Decision makers can then assess which uncertainty is material under the applicable standard without mistaking a label for a validated probability.
Portfolio and measurement implications
Romanosky's (2016) study of cyber incident costs analyzes recorded events spanning 2004 through 2015 and describes the public-source selection process. The research distinguishes several incident, litigation and financial metrics and acknowledges the descriptive nature of the dataset. Its relevance is methodological: measured loss distributions depend on what becomes known and recorded, and exposure comparisons require appropriate denominators. The figures from that historical sample are not used here as contemporary cyber insurance severity estimates.
A claim-level database should identify policy exposure, actual events, notified claims, claim dispositions, paid and incurred loss, expenses, coverage modules and the stage of investigation. Incident frequencies per organization-year, reported claim frequencies per policy-year and accepted indemnity frequencies per policy-year answer different questions. A change in forensic logging or reporting can increase observed incidents without changing the actual event frequency. A portfolio showing fewer claims of uncertain causation may reflect better evidence, stricter acceptance criteria or a different composition of incidents. Distinguishing these mechanisms requires data not available in this study.
For a simple synthetic demonstration, suppose there are 100 actual incidents among 1,000 equally exposed organization-years. If 60 are detected in year A and 80 in year B under otherwise identical conditions, the observed rate rises from 6% to 8% although the actual 10% rate is unchanged. This is a measurement illustration, not a detection-rate estimate. The same logic applies when preservation practice changes the proportion of claims with resolved cause. Without accounting for ascertainment, claims experience may confuse quality of observation with change in risk.
Portfolio dependence adds a separate dimension. A single provider event may generate correlated claims across customers with different policy modules and different exposure to the provider. Common mechanism classification helps identify potential accumulation; insured exposure, limits, deductibles and recoverable loss must still be measured separately. If uncertain-cause events are omitted from aggregation, tail exposure can be understated. If every coincident outage is assigned to one vendor event without support, aggregation can be overstated. Both errors affect underwriting, reserving and reinsurance interpretation. No empirical dependence parameter is estimated in this research.
Integrated Scientific Synthesis and Research Agenda
The relationship between evidence and indemnification
The integrated analysis yields a practical scientific conclusion: the evidentiary requirement should follow the causal and contractual proposition rather than the volume of digital records available. A technically exhaustive investigation can still leave the insured's financial counterfactual unresolved. An incomplete technical history may nevertheless establish all material facts for a particular coverage decision. The quality of the investigation is measured by the warranted relationship between the observations, the causal explanations and the loss items that must be evaluated.
This conclusion follows from three results developed above. First, digital traces carry different observation semantics and preservation protects the possibility of later inference rather than guaranteeing a true interpretation. Second, event timing separates intrusion, operational effect, awareness, response, restoration and financial consequence; these intervals cannot be collapsed without risking contractual error. Third, the same technical history can produce distinct coverage outcomes under different wording, illustrated by the specimen and the two judicial opinions. These are ordinary analytical conclusions drawn from the research corpus, not claims of a new named theory.
The scope of uncertainty should be explicit. An unknown initial access route may remain material for a control-specific exclusion, an insurer's subrogation claim or a preventive review. It may be immaterial to a loss item whose causal relation to an established qualifying event is independently supported. A decision should expose which facts have that status and why. This promotes both defensible settlements and scientifically honest reporting. It also helps prevent needless extension of an investigation whose additional findings would have no bearing on the disputed issue.
Institutional design and incentives
The expected cost of a cyber claim includes the damage itself, the expense of measuring and resolving it, and delays associated with incomplete information. Forensic readiness may affect all three through earlier observation, more targeted recovery and better evidence of facts. The direction and magnitude of those effects depend on the technology, governance, contract and event population. An empirical causal effect cannot be asserted from the existence of standards recommending preparedness.
Insurance arrangements can align incentives by making the information needs of claims assessment clear before a loss. Logging and preservation should have an intelligible relationship to system criticality and to the facts likely to matter. Evidence obligations that require records an insured cannot obtain from a provider may create disputes without improving the underlying risk. Equally, omitting attention to basic observation capability can leave an insurer reliant on retrospective narratives. The design problem is to specify attainable evidence duties with a proportionate relation to the insured peril and to enable reasonable response when emergencies constrain collection.
Preservation and privacy must be pursued together. Indiscriminate retention of sensitive payloads can create additional security and confidentiality burdens. Collection should have an identified investigative purpose, access restrictions, and an appropriate treatment of third-party information. Different organizations have different legal preservation, privacy, notification and discovery obligations. Interpretation of these duties needs their applicable jurisdiction and circumstances. The present study establishes the need to protect evidential availability while containing unnecessary exposure; it does not prescribe a universal lawful retention period.
An auditable interdisciplinary research record
A useful research record connects the source, proposition, reasoning, alternative account and consequence. Where a proposition concerns an artifact, its record should identify the original data source, the available interval, the interpretation method and relevant caveats. Where it concerns economic loss, it should specify the affected process, ledger or other supporting record, calculation units and treatment of deferred or recovered operations. Where it concerns coverage, it should identify the exact issued wording and the legal authority or interpretive assumptions used. Revisions should preserve previous inferences and their reasons for change so a later reader can distinguish correction from simple replacement.
Consider a disputed shutdown. An incident-response report can establish that a particular action was taken at a specified time in response to observed risk; transaction records can quantify the resulting change in throughput; a policy may include a shutdown exclusion, an endorsement or a different trigger. No single specialist's opinion resolves all questions. Integrating these records means coordinating compatible units and times, not erasing the independent disciplines. A chronology shared among specialists should permit each to inspect which claims are factual, which are estimates and which are interpretations of law or contract.
The same structure assists review of early claim estimates. Initial reserve assumptions may rely on a limited estimate of restoration time, an unresolved provider event and a provisional financial baseline. Later artifacts and actual costs may revise several assumptions. Documented reasons for movement support analysis of loss development. The final adjustment should distinguish estimate corrections from new damage and from a change in coverage interpretation. This matters because actuarial records serve different future decisions from the contemporaneous forensic investigation.
Research limitations and future evidence
The source corpus is deliberately limited to selected materials whose relevant statements were accessible. Official NIST publication abstracts and the repository abstract of the timestamp study establish only the limited propositions recorded in the evidence ledger; detailed implementation requirements are not inferred from abstracts. The official PDF text of the public investigation, specimen, New Jersey decisions and EU regulation was readable in extracted form, but their rendered page images were unavailable to the research environment. Citation verification used the accessible text rather than claiming examination of unavailable figures or page images. The Indiana opinion was checked through its reproduced judicial text. Later legal treatment beyond the specified New Jersey dismissal and a comprehensive current multi-country jurisprudence survey have not been established. The public claims data available to the study do not support causal effect estimation, policy-level payment distributions or calibrated forensic likelihood ratios. The historical incident study serves as a methodological example and is not treated as a contemporary insured-loss dataset.
A more extensive empirical program would require deidentified policy-level exposure and claims records with versions of the issued forms, source-specific telemetry characteristics, detection and retention intervals, incident histories, expense classifications and the evolution of causal findings. A review team could test agreement between independent analysts using the same admissible data and measure whether disagreements affect coverage classifications or only technical details. Simulated incident histories with known mechanisms could test clock correction and partial ordering. The outcome measure should distinguish reconstruction accuracy from agreement and speed of decision. A separate longitudinal design could assess whether specified readiness arrangements improve loss outcomes or merely improve documentation; it would require adjustment for security maturity, business scale, dependencies, incident type and prior risk.
Comparative legal work should pair actual wording with the relevant substantive and procedural law, including the effect of burdens of proof, exclusions, multiple causes and policy conditions. It should sample jurisdictions according to a declared design and preserve language-specific differences rather than generalize from a small set of reported disputes. Comparative insurer research should distinguish notified incidents, paid insured losses and disputes that become public. Those are separate selection mechanisms.
The central research agenda is therefore the measurement of the information-to-decision relationship: which observable traces discriminate among financially material causal histories, what uncertainty remains when those traces are unavailable, and how alternative histories map to distinct contractual outcomes. This inquiry joins forensic science, risk management, economics, actuarial science and insurance law within the canonical research object.
Conclusion
Cyber incident forensic causation establishes a reasoned connection between digital events, operational effects and insured financial consequences. Evidence preservation enables investigation of that connection by retaining observations with their temporal meaning, scope and provenance. The reliability of the record depends on more than physical or cryptographic integrity: its generation process, coverage of relevant events, interpretation, independent corroboration and relationship to the loss all affect the warranted inference.
The documentary analysis demonstrates why a cyber occurrence must be evaluated in layers. The time at which an incident begins, the time at which an insured operation becomes materially interrupted, the time at which the insured discovers a problem and the time at which functionality is restored can differ. A particular mechanism may be technically established while its legal relevance remains unresolved. A financial loss may be demonstrable while some of its components remain outside the contract. Alternative causal accounts should therefore be retained until their material consequences are resolved by evidence or explicitly bounded.
The selected insurance wording and judicial opinions illustrate that coverage causation is sensitive to the exact terms and jurisdiction. The public intrusion report illustrates that unknown antecedents may coexist with supported downstream findings. The hypothetical calculations demonstrate that uncertainty about the qualifying start, income baseline and recoveries can produce materially different values even when the broad incident narrative is unchanged. The historical empirical research supplies a further reason for separating observed events from true event incidence and from insured claims outcomes.
The primary scientific result is a coherent interdisciplinary synthesis: investigators should preserve the observations capable of resolving causal hypotheses, reconstruct the technical and economic sequence with transparent uncertainty, and submit only the established or explicitly conditional consequences to contract-specific evaluation. Its implications extend to underwriting evidence requirements, claims adjustment, reserve development, third-party recovery and portfolio accumulation. Further empirical validation is necessary before asserting quantitative relationships between evidence readiness and insured outcomes. This research establishes the analytical relationships and the methodological grounds on which such validation can proceed.
References
Brezinski, D., & Killalea, T. (2002). Guidelines for evidence collection and archiving. RFC 3227, BCP 55. Internet Engineering Task Force. DOI: 10.17487/RFC3227. Publication record
Cyber Safety Review Board. (2024). Review of the Summer 2023 Microsoft Exchange Online intrusion. U.S. Department of Homeland Security. Publication record
CyberAcuView, LLC. (2022). CyberAcuView cyber insurance wording. Base policy, first-party business interruption and dependent business interruption coverage modules, pp. 21-28. Lloyd's Market Association-hosted specimen. Publication record
European Parliament and Council of the European Union. (2022). Regulation (EU) 2022/2554 of 14 December 2022 on digital operational resilience for the financial sector (DORA). Official Journal of the European Union, L 333, 1-79. Publication record
Guttman, B., White, D. R., Williams, S., & Walraven, T. (2022). Digital evidence preservation: Considerations for evidence handlers. NIST IR 8387. National Institute of Standards and Technology. DOI: 10.6028/NIST.IR.8387. Publication record
Herman, M., Iorga, M., Salim, A. M., Jackson, R., Hurst, M., Leo, R., Mishra, A. K., Landreville, N., & Wang, Y. (2024). NIST cloud computing forensic reference architecture. NIST SP 800-201. National Institute of Standards and Technology. DOI: 10.6028/NIST.SP.800-201. Publication record
Hernan, M. A., & Robins, J. M. (2020). Causal inference: What if. Chapman & Hall/CRC. Publication record
Kent, K., & Souppaya, M. (2006). Guide to computer security log management. NIST SP 800-92. National Institute of Standards and Technology. DOI: 10.6028/NIST.SP.800-92. Publication record
Kent, K., Chevalier, S., Grance, T., & Dang, H. (2006). Guide to integrating forensic techniques into incident response. NIST SP 800-86. National Institute of Standards and Technology. DOI: 10.6028/NIST.SP.800-86. Publication record
Nelson, A., Rekhi, S., Souppaya, M., & Scarfone, K. (2025). Incident response recommendations and considerations for cybersecurity risk management: A CSF 2.0 community profile. NIST SP 800-61 Rev. 3. National Institute of Standards and Technology. DOI: 10.6028/NIST.SP.800-61r3. Publication record
Pearl, J. (2010). An introduction to causal inference. International Journal of Biostatistics, 6(2), Article 7. DOI: 10.2202/1557-4679.1203. Publication record
Romanosky, S. (2016). Examining the costs and causes of cyber incidents. Journal of Cybersecurity, 2(2), 121-135. DOI: 10.1093/cybsec/tyw001. Publication record
Souppaya, M., Barker, W., Fisher, W., & Kent, K. (2026). Ransomware risk management: A Cybersecurity Framework 2.0 community profile. NIST IR 8374 Rev. 1. National Institute of Standards and Technology. DOI: 10.6028/NIST.IR.8374r1. Publication record
Superior Court of New Jersey, Appellate Division. (2023, May 1). Merck & Co., Inc. v. ACE American Insurance Co., 475 N.J. Super. 420, Nos. A-1879-21 and A-1882-21. Publication record
Supreme Court of Indiana. (2021, March 18). G&G Oil Co. of Indiana, Inc. v. Continental Western Insurance Co., 165 N.E.3d 82, No. 20S-PL-617. Publication record
Supreme Court of New Jersey. (2024, January 23; filed January 26). Merck & Co., Inc. v. ACE American Insurance Co., Nos. A-62/63, 088274. Order of dismissal. Publication record
Vanini, C., Hargreaves, C. J., van Beek, H., & Breitinger, F. (2024). Was the clock correct? Exploring timestamp interpretation through time anchors for digital forensic event reconstruction. Forensic Science International: Digital Investigation, 49(Suppl.), 301759. DOI: 10.1016/j.fsidi.2024.301759. Publication record