The Epistemic Boundary of Forensic Science: What Questions Can Evidence Never Answer?

Mapping the distance between what a trace can show and what an investigation wants to know

Executive Summary

A crime scene can yield a device that was active, a substance that was transferred, a document that was altered, and a body that shows a precise mechanism of injury — and still leave the most consequential question in the case untouched: why the person acted, and what the act meant to them. This is not a failure of technique. It is a structural feature of what physical and digital traces can encode. This article develops a forensic epistemology of that boundary. It distinguishes observation, description, interpretation, inference, reconstruction, explanation, intention, motive, meaning, and speculation as a graded sequence rather than a single category called "evidence," and it asks, at each step, what the trace can license and what must be supplied from outside it. Drawing on the philosophy of scientific evidence (Haack, 2014), the statistics of forensic inference (Aitken et al., 2021; Taroni et al., 2014), the criminalistics paradigm that separates identification from motive (Inman & Rudin, 2000, 2002), cognitive science on contextual bias and memory (Dror, 2020; Kassin et al., 2013; Loftus, 2005), and the governmental science-validity reviews that reshaped the field (National Research Council [NRC], 2009; President's Council of Advisors on Science and Technology [PCAST], 2016), the article proposes two original conceptual frameworks — an Epistemic Gradient of Forensic Evidence and an Epistemic Boundary Framework — and applies them to underdetermination, intention, motive, meaning, absent evidence, historical truth, digital forensics, and artificial intelligence. It closes by distinguishing limitations that better evidence can shrink from limitations that are built into what a trace is: a present-tense residue of a past event, not a record of why that event happened.

Key Findings at a Glance

• Forensic evidence is retrospective by nature: it exists in the present and is used to license claims about an unobserved past, which makes every forensic conclusion an inference, not a direct readout (Inman & Rudin, 2002).
• The 2009 National Research Council report and the 2016 PCAST report both found that several long-accepted feature-comparison disciplines lacked the error-rate studies needed to call them scientifically validated — a gap in evidence, not proof of unreliability (NRC, 2009; PCAST, 2016).
• Physical evidence can establish association and sequence far more securely than it can establish why an event occurred; classical criminalistics theory treats motive as explicitly outside its explanatory reach (Inman & Rudin, 2000).
• Mental states such as intention are, in law, almost always established through circumstantial inference rather than direct observation, because no instrument observes a mental state itself.
• Cognitive bias research shows that the same physical evidence can yield opposite conclusions depending on contextual information the examiner is exposed to, which means the evidence-to-conclusion gap is filled partly by psychology, not physics (Dror, 2020; Kassin et al., 2013).
• Absence of a trace is not proof that an event did not occur; it may only reflect a limit of survival, detection, or search (Altman & Bland, 1995).
• Large language models and other generative AI systems can integrate more data than a human examiner, but they inherit the same evidentiary ceiling and add a new risk: confident, fluent output that is not grounded in the evidentiary record at all.
• Digital evidence's apparent completeness is itself a source of epistemic risk, since metadata can be fragmented, forged, or subject to more than one valid interpretation (Casey, 2002; Greeshma, 2023).

The Forensic Paradox

Consider an investigation that is, by any operational measure, a success. Digital forensics places a specific device at a specific location at a specific minute. Trace evidence places a specific person in physical contact with a specific object. Toxicology identifies a specific compound at a specific concentration. Document examination establishes that a specific page was altered after a specific date. Every one of these findings is precise, replicable, and legally admissible. And when the report is finished, the file may still contain no answer to the question that matters most to the family, the court, and the public: why did this happen, and what did it mean to the person who did it?

This is not a rare failure mode. It is close to the default condition of evidence-rich casework. The paradox is not that forensic science has fallen short of some achievable standard — as though a better microscope or a faster sequencer would close the gap. The paradox is that presence, activity, and sequence are a different kind of fact than intention, motive, and meaning, and no increase in the resolution of the first kind of fact automatically buys any of the second kind. A trace can tell an investigator that something happened. It is a separate, much harder question — sometimes an unanswerable one — whether that trace can tell anyone why.

What follows is not an argument that forensic science is unreliable. Statistical and Bayesian approaches to evidence evaluation have made forensic inference in domains such as DNA profiling and trace comparison more rigorous and more honest about its own uncertainty than almost any other applied science (Aitken et al., 2021; Taroni et al., 2014). The argument is narrower and, in a sense, more useful: that reliability and scope are two different properties. A method can be highly reliable within a defined evidential domain and still be structurally incapable of answering a question that lies outside that domain. The task of this article is to map where that domain ends.

Ten Categories on One Gradient

Discussions of "what evidence shows" routinely collapse ten distinct epistemic categories into one. Separating them is the first move of forensic epistemology.

1. Observation

A compound is detected, a biological trace is measured, a digital artefact exists. This is the only category evidence generates without human interpretive input.

2. Description

What can be accurately stated about the observation itself — composition, location, timestamp structure, measurable pattern — without yet saying what it means.

3. Interpretation

Scientific significance assigned to the observation: a possible source, mechanism, or activity proposition.

4. Inference

A conclusion that goes beyond direct observation: a likely sequence, a probable association, a set of competing hypotheses ranked by support.

5. Reconstruction

Multiple observations assembled into an account of an unobserved past event — "ordering of associations in space and time," in the language of criminalistics theory (Inman & Rudin, 2002).

6. Explanation

Why an event may have occurred — a causal story, not merely a sequence.

7. Meaning

What the action represented to the person who performed it — its significance within their own life and context.

8. Intention

The mental state that accompanied the action — what, in criminal law, is called mens rea.

9. Motive

Why the individual wanted or chose to act — a question courts generally treat as relevant to proof of intent but not identical to it.

10. Speculation

A proposition that extends beyond what evidence and validated reasoning can support — the point past which confident language is no longer earned.

Figure 2. The Ten-Category Ladder. The path runs from directly observable (top left) to interpretive and often outside forensic scope (bottom left).

The transitions between these categories are not all equally steep. Observation to description is close to automatic. Description to interpretation already introduces a model — an assumption about how a trace relates to its source. Interpretation to inference introduces competing hypotheses. Reconstruction to explanation introduces causal claims that the physical trace alone rarely settles. And explanation to intention, motive, and meaning crosses into a domain that no forensic instrument directly measures at all, because intention and meaning are not physical properties of a trace; they are properties of a person's inner life, accessed only through what that inner life leaves in the physical record.

The Epistemic Gradient of Forensic Evidence (A proposed conceptual framework developed for analytical discussion from the literature reviewed, not an established international forensic standard.)

Figure 1. The Epistemic Gradient of Forensic Evidence. Certainty generally decreases moving down the gradient, though the rate of decrease is case-dependent, not fixed.

At the top of the gradient, a compound's identity or a fingerprint minutia count is close to directly observable, subject mainly to measurement error. Moving down, each stage adds an assumption the trace itself does not supply: that this source produced this trace and no plausible alternative source did (identification); that this sequence of traces reflects one specific temporal order and not several compatible orders (reconstruction); that this physical act reflects a specific causal chain rather than an equally consistent alternative chain (explanation); and, at the bottom, that the actor's inner state can be read off the outer act at all (intention, motive, meaning). Nothing in this gradient claims that the lower stages are unknowable in every case — only that they require a larger inferential leap, supported by more assumptions, each of which is a place the conclusion can go wrong.

1. Can Evidence Directly Observe the Past?

Forensic evidence exists in the present. The event of interest exists in the past. Every forensic conclusion is therefore an inference from a present trace to an unobserved historical event — what philosophers of science and inverse-problem theorists call a retrospective or inverse inference. Criminalistics theory states this plainly: physical evidence "provides clues to a particular course of events, but does so only indirectly," and it is left to the analyst, and ultimately the court, "to make an inference about a criminal event from the physical evidence" (Inman & Rudin, 2002). The trace is not a recording of the event; it is a residue produced by the event, and the reconstruction runs backward from residue to cause.

2. The Problem of Underdetermination

Because reconstruction runs backward, more than one historical sequence can sometimes produce the identical surviving evidence — a phenomenon closely related to what inverse-problem theory calls non-uniqueness: a forward process that maps many possible causes onto the same observed effect cannot be inverted to recover a single cause without additional assumptions. In forensic terms: a wound pattern, a transfer of fibers, or a digital timestamp may be equally consistent with several different event histories. This does not mean forensic reconstruction is arbitrary. It means the strength of a reconstruction depends on how many competing hypotheses remain compatible with the evidence, and how far the evidence has actually narrowed that set. A reconstruction can be uniquely supported, strongly supported, one of several compatible accounts, or an unsupported narrative selected for reasons that have nothing to do with the evidence. Distinguishing these four states, case by case, is a core task of scientific forensic reasoning rather than an embarrassment to be hidden.

Figure 3. Underdetermination: several distinct event histories can produce the identical surviving trace, so the evidence alone narrows the compatible set without necessarily closing it to one.

3. Can Evidence Establish Intention?

In criminal law, mens rea — the mental state accompanying an act — is almost never established by direct observation of the mind. It is established by circumstantial inference from conduct, statements, planning, and context, because a mental state is not the kind of thing a physical instrument measures. Legal doctrine is explicit that intent, absent a confession, is "a matter of circumstantial proof," inferred from what a person did, said, and planned before, during, and after an act. This does not mean forensic evidence is irrelevant to intention. It means intention is not observed directly; it is inferred, defeasibly, from a pattern of physical facts that are themselves observable — device activity, the trajectory of a weapon, the sequence of a person's movements. The correct scientific claim is not that evidence can never speak to intention, but that it speaks to intention only through inference, and that inference always carries the possibility of an innocent alternative explanation for the same pattern.

Does evidence contain intention, or does intention remain an inference made about an observed action?

The more precise position is the second: forensic evidence can support or weaken specific propositions about intention when combined with context, without ever constituting direct measurement of a mental state.

4. Can Evidence Establish Motive?

Motive — why a person wanted a particular outcome — sits a further step from the physical trace than intention. Classical criminalistics separates the two explicitly: forensic science is "fundamentally one of reconstruction... it is not concerned with, and cannot determine, why something happened (the motivation)" (Inman & Rudin, 2002). Courts frequently treat motive as legally relevant context rather than a required element of proof, precisely because a physical trace does not encode desire. Communications, financial records, and behavioural history can support a hypothesis about motive; they rarely settle it with the same confidence a chemical assay settles the identity of a compound.

5. Can Evidence Establish Meaning?

Meaning — what an act signified to the person who performed it, within their own linguistic, cultural, and personal frame — is the furthest point on the gradient from direct observation. A message, a gesture, or a purchase can carry radically different meanings depending on context that forensic examination, by itself, has no access to. This is not a defect that better instruments will fix, because meaning is not encoded in the physical properties of a trace at all; it is a property of interpretation, shared between the actor and their social and linguistic world. Forensic science can document what was said or done with high precision. Whether that language or conduct constitutes a threat, a joke, a coded reference, or an accident often depends on interpretive judgment that lies outside the forensic laboratory's competence, however precise its instruments.

6. The Unrecorded Action Problem

Can the absence of a trace establish that an act did not occur? Statisticians have long warned against this inference in a closely related context: "absence of evidence is not evidence of absence," because a negative finding may reflect a genuine absence of the phenomenon, or it may reflect inadequate power to detect it, incomplete search, or trace decay (Altman & Bland, 1995). The same logic applies directly to forensic reconstruction. Some actions leave no recoverable trace not because they did not happen, but because the method used, the surface involved, or the time elapsed did not preserve one. Treating "we found nothing" as equivalent to "nothing happened" quietly converts a limitation of search into a fact about the world.

7. The Problem of Unique Historical Truth

Exactly one sequence of events actually occurred. It does not follow that surviving evidence is sufficient to uniquely reconstruct that sequence. Historical reality and evidential reconstruction are two different things: the first is fixed and singular; the second is a model built from whatever fraction of the original information happened to survive, be collected, and be correctly interpreted. When multiple reconstructions remain compatible with the surviving evidence, forensic science's honest output is the set of compatible accounts, ranked by support — not a single narrative dressed in the language of certainty because a single narrative feels more useful to an investigation or a courtroom.

8. The Measurement-to-Meaning Gap

Consider the pipeline a piece of evidence typically travels: physical signal, measurement, feature extraction, classification, interpretation, human conclusion. Information can be lost or assumptions can be added at every stage. A signal is not the same as a measurement of it; a measurement is not the same as the feature extracted from it; a feature is not the same as the classification assigned to it; and a classification is not the same as the human conclusion a report ultimately states. Digital forensics research documents this concretely: incomplete metadata can cause a timestamp to be linked to the wrong application, and fragmented records can cause artefacts to be misclassified, even when each individual extraction step performed correctly (Casey, 2002; Greeshma, 2023). Each step in the pipeline is an opportunity for the final conclusion to claim more meaning than the original signal actually contained.

Figure 4. The Measurement-to-Meaning Pipeline. Each stage can add an assumption the raw signal did not itself contain.

9–10. Evidence, Explanation, and the Causal Boundary

Evidence supports claims about what is consistent with the observations. Explanation attempts to answer why the observations occurred. The two are frequently conflated because a good explanation feels like it has been "proven" once it fits the evidence — but fitting the evidence is a necessary condition for an explanation, not a sufficient one, since several explanations may fit equally well. Causal-inference theory formalises this distinction sharply: Pearl's account of a causal hierarchy separates mere association (what tends to occur together) from intervention (what happens if a variable is deliberately changed) and counterfactual reasoning (what would have happened under different conditions), and shows that data at one level cannot, by itself, license conclusions at a higher level (Pearl, 2009). A forensic correlation between a trace and a suspect is, at most, association-level evidence; treating it as automatically causal, without further reasoning about mechanism and alternative explanations, repeats a well-documented inferential error.

11. Probability and Knowledge

Modern forensic statistics evaluates evidence through likelihood ratios — how much more probable the evidence is under one proposition than a competing one — rather than through categorical statements of identity (Aitken et al., 2018; Taroni et al., 2014). This is a scientific advance, but it creates a communication risk: a likelihood ratio, however large, remains a statement of relative probability, not of certainty, and courtroom language has a persistent tendency to round strong probabilistic support up to categorical proof. The boundary between "very probable" and "known" is where much of the damage from forensic overstatement has historically occurred (NRC, 2009; PCAST, 2016).

12. Scientific Inference vs. Speculation

A useful progression: evidence-supported observation, scientifically justified inference, plausible but weak hypothesis, speculative proposition, narrative assertion unsupported by evidence. The transition along this chain is not a single bright line; it is a gradual weakening of support that different disciplines formalise differently — through likelihood ratios, through explicit alternative-hypothesis testing, or through transparent labelling of confidence. What every discipline shares is the obligation to say, at each stage, how far the claim has travelled from the last point the evidence itself actually supports.

13. The Narrative Completion Problem

Investigators, like all people, prefer coherent stories to incomplete ones, and this preference operates below conscious awareness. Motivated-reasoning research shows that once a preliminary theory forms, ambiguous evidence tends to be interpreted in ways that support it (Kunda, 1990), and criminal-investigation research documents the same pattern directly: an initial theory of guilt can lead investigators to filter out contradictory evidence or read ambiguous behaviour as confirmation, a pattern amplified under time pressure and public scrutiny (Ask & Granhag, 2005; Findley & Scott, 2006). Left unchecked, this "tunnel vision" fills evidentiary gaps with narrative rather than evidence — not through dishonesty, but through an entirely ordinary cognitive shortcut that forensic training must actively counteract.

14. Can AI Cross the Epistemic Boundary?

Artificial intelligence systems can integrate larger datasets and detect patterns beyond unaided human perception. It does not follow that computational power lets AI answer questions the underlying evidence never encoded. A large language model asked to reconstruct a timeline from fragmented digital artefacts will, by design, produce a fluent and confident-sounding narrative even when the underlying evidence is genuinely ambiguous — a behaviour documented across domains as "hallucination," the generation of plausible but ungrounded content, and shown to correlate with exactly the conditions forensic reconstruction often presents: negation, missing information, and the need for inference under uncertainty. Research on digital-forensic applications of large language models has found that incomplete metadata and contextual-inference bias are leading sources of misclassification even in systems built with chain-of-custody safeguards. Separately, research on linguistic assertiveness in AI systems has found a persistent gap between how confidently a model expresses a conclusion and how accurate that conclusion actually is — precisely the miscommunication risk that likelihood-ratio science was designed to prevent in human testimony. The more defensible claim is not that AI cannot help forensic interpretation, but that more sophisticated inference improves the interpretation of available evidence without creating information that was never present in the evidentiary record.

15. The Digital Evidence Illusion of Omniscience

Modern investigations can draw on device records, cloud artefacts, metadata, closed-circuit footage, transaction logs, and sensor data simultaneously. This abundance can create the impression that, given enough data sources, everything about an event becomes knowable in principle. Digital-forensics research pushes back directly: digital evidence is "often arising from a complex interplay of technologies and human actions," and this complexity, rather than being resolved by more data, frequently produces more elaborate forms of uncertainty — conflicting timestamps across systems with different clocks, deleted-and-partially-recovered records, and metadata that supports more than one internally consistent story (Casey, 2002; Greeshma, 2023). More evidence narrows some uncertainties while introducing others; it does not convert an inference problem into an observation problem.

16. Better Evidence vs. Impossible Knowledge

Not every current limitation is permanent, and treating all forensic gaps as identical erases a distinction that matters for both science and policy. It is worth separating: technological limitation (a method exists but current instruments lack sensitivity or throughput); methodological limitation (no validated method yet exists, though one is conceivable); data limitation (the relevant reference data or population statistics have not yet been collected); logical limitation (the available evidence is compatible with more than one hypothesis, and no additional measurement of the same kind would resolve the ambiguity); and epistemic limitation (the question concerns a category — a mental state, a subjective meaning — that no physical trace directly encodes, regardless of instrument sensitivity). The first three shrink with research investment. The last two do not shrink the same way, because they are not primarily instrument problems.

Figure 5. Five Types of Forensic Limitation. Technological, methodological, and data limitations shrink with investment; logical and epistemic limitations are built into what a trace is.

17. A Taxonomy of Questions Forensic Evidence Cannot Directly Answer

Category	Example question	Status
A — Internal mental states	Did the person subjectively believe their act was justified?	Not directly observable; potentially inferable indirectly from statements and conduct
B — Unrecorded events	Did a conversation occur that left no digital or physical trace?	Directly inaccessible unless corroborated by an independent trace
C — Lost historical detail	What was the exact sequence of two near-simultaneous actions?	Potentially inferable within a margin of temporal resolution; often underdetermined
D — Counterfactual events	What would have happened had the person not acted?	Not observable by definition; addressable only through causal modelling assumptions
E — Meaning beyond observation	What did the message mean to the sender, given their private context?	Conditionally testable only with independent corroborating context

The Epistemic Boundary Framework for Forensic Science (A proposed conceptual framework developed for analytical discussion from the literature reviewed, not an established international forensic standard.)

Eight dimensions determine how far a given forensic question sits from what evidence can directly settle:

Direct observability

Can the proposition be measured, or only inferred from something else that is measured?

Trace survival

Does relevant evidence still exist to be found, or has it decayed, been overwritten, or never formed?

Attribution

Can the trace be reliably connected to the specific entity or person in question, rather than merely to "someone" or "something"?

Inferential distance

How many additional assumptions separate the raw observation from the stated conclusion?

Alternative explanations

How many other hypotheses remain compatible with the same evidence?

Ground truth accessibility

Can the conclusion ever be checked against an independently known correct answer?

Replicability

Can the relevant claim, or an analogous one, be tested experimentally?

Meaning dependence

Does answering the question require access to the actor's own subjective interpretation, which no instrument reaches directly?

Figure 6. The Eight Dimensions of the Epistemic Boundary Framework. A question that scores strongly across all eight axes sits near the top of the evidentiary gradient; one that scores weakly on most sits near the bottom, regardless of data volume.

A question that scores strongly observable, well-attributed, short in inferential distance, with few alternative explanations, checkable ground truth, and low meaning-dependence — the identity of a controlled substance, for instance — sits near the top of the gradient. A question that scores weakly on most of these dimensions — what a specific gesture meant to the person who made it — sits near the bottom, regardless of how much data surrounds the case.

Cross-Disciplinary Comparison

Domain	What evidence can directly establish	What requires inference	Major epistemic limitation
DNA profiling	Genetic profile composition; statistical rarity within a reference population	Source attribution under competing transfer hypotheses; timing of deposit	Cannot establish how or when contact occurred, only that it did (Aitken et al., 2021)
Digital evidence	File existence, hash values, raw metadata fields	Who operated the device; intent behind an action; corrected chronology	Metadata reliability and fragmentation (Casey, 2002; Greeshma, 2023)
Toxicology	Presence and concentration of a compound	Route, timing, and voluntariness of ingestion	Concentration alone rarely fixes a single behavioural narrative
Fingerprint examination	Minutiae configuration in a latent and reference print	Common source to the exclusion of all others in the population	Individualisation claims outpace validated statistical foundations (Cole, 2001; PCAST, 2016)
Trace evidence (fibres, GSR)	Physical or chemical match class	Transfer mechanism and event association	Class evidence is compatible with many innocent transfer histories
CCTV / imagery	Recorded visual events at known camera coordinates and timestamps	Identity of a partially obscured individual; intent behind an observed action	Resolution and camera clock drift bound what can be attributed
Forensic pathology	Cause of death; wound morphology	Manner of death (accident, suicide, homicide) in ambiguous cases	Contextual information demonstrably shifts manner-of-death rulings on identical medical findings (Dror et al., 2021)
AI-assisted evidence	Pattern detection across large datasets at machine speed	Causal or intentional narrative generated from the pattern	Fluent, confident output can be ungrounded in the actual evidentiary record

Case-Based Analytical Scenarios

Scenario 1 — Presence Does Not Equal Intention

Location data places a device at a residence for eleven minutes. This establishes presence with high confidence. It does not, by itself, distinguish a planned entry from an accidental one, a delivery, or a mistaken address — that distinction requires context the location data alone does not carry.

Scenario 2 — Device Activity Does Not Equal Human Purpose

A smartphone logs a search query and an app launch in close succession. The log establishes that the device performed these actions. Whether a human deliberately typed the query, an automated update triggered it, or another person briefly used the device are three different hypotheses the log alone cannot rank.

Scenario 3 — Injury Pattern Does Not Automatically Establish Motive

A wound pattern can support a mechanism (blunt force, sharp force, a specific implement class) with strong confidence. It rarely supports, on its own, why the injury was inflicted — self-defence, premeditation, and sudden provocation can all be physically compatible with the same wound morphology.

Scenario 4 — Multiple Reconstructions From the Same Evidence

Three independent traces — a transferred fiber, a partial shoe impression, and a timestamped access log — are each compatible with two different sequences of entry and exit. Combined, they narrow the compatible set from four sequences to two, but do not, without further evidence, collapse it to one.

Scenario 5 — The Missing Trace Problem

No trace evidence links a named individual to a scene. This is consistent with that individual's absence, but is equally consistent with contact that left no recoverable residue given the surface, cleaning, and time elapsed — the absence of a trace narrows probability without settling the question of absence.

Major Research Gaps

1. Formal methods for quantifying inferential distance between an observation and a stated conclusion.
2. Validated protocols for testing whether a proposed reconstruction is the unique account compatible with the evidence, or one of several.
3. Standardised language for communicating likelihood-ratio evidence without collapsing it into categorical certainty in court.
4. Cross-disciplinary error-rate studies covering pattern-comparison methods still lacking the black-box validation PCAST identified as absent (PCAST, 2016).
5. Empirical models of how contextual bias propagates specifically through digital-forensic metadata interpretation, extending Dror's contextual-bias framework beyond pattern-comparison disciplines.
6. Formal criteria for when absence of a digital or physical trace should be treated as informative versus uninformative, extending the logic of Altman and Bland (1995) into forensic practice.
7. Frameworks for AI systems that can recognise and flag when a question exceeds what the supplied evidence can support, rather than generating a fluent answer regardless.
8. Comparative research on narrative-bias mitigation techniques (such as scenario-based falsification) across different investigative and judicial systems.
9. Research on how motive-related evidence (communications, financial records) should be statistically weighted relative to physical trace evidence in combined assessments.
10. Validation research specific to the Indian evidentiary and forensic infrastructure context, given the expanded "any other field" language for expert opinion introduced by the Bharatiya Sakshya Adhiniyam, 2023.
11. Studies on how examiners' confidence language correlates with actual accuracy across forensic sub-disciplines, mirroring assertiveness-calibration research now emerging for AI systems.
12. Formal decision-theoretic frameworks for when a forensic scientist should abstain from offering a conclusion rather than offer a weak one.
13. Longitudinal research on how digital-evidence volume affects investigator confidence independent of actual evidentiary strength.
14. Comparative studies of underdetermination across forensic sub-disciplines to identify which produce the widest sets of compatible reconstructions.
15. Research connecting Bayesian network modelling of forensic evidence (Taroni et al., 2014) to practical courtroom communication training.
16. Studies examining whether forensic reconstruction training measurably reduces narrative-completion error under time pressure.
17. Frameworks for evaluating when large-language-model assistance in casework should be disclosed, and how its outputs should be independently verified against source evidence.
18. Research into the specific epistemic risks of AI-assisted digital-forensic knowledge-graph construction, extending early findings on metadata-driven misclassification.
19. Cross-jurisdictional comparison of how "expert opinion" statutes handle emerging forensic methods lacking established error rates.
20. Development of shared vocabulary distinguishing technological, methodological, data, logical, and epistemic limitation across forensic sub-disciplines, to prevent the collapse of distinct limitation types into a single undifferentiated category of "uncertainty."

Future Research Agenda

1. Measuring inferential distance. Rationale: no shared metric currently exists for how many assumptions separate an observation from a courtroom conclusion. Difficulty: requires formal collaboration between statisticians and working examiners. Value: would allow direct comparison of claim strength across disciplines.
2. Validating reconstruction uniqueness. Rationale: current practice rarely states explicitly whether a reconstruction is the only compatible account. Difficulty: requires systematic enumeration of alternative hypotheses, which is resource-intensive. Value: would prevent single-narrative overconfidence in complex cases.
3. Formal models of alternative explanations. Rationale: alternative-hypothesis testing is inconsistently applied across sub-disciplines. Difficulty: requires discipline-specific reference data. Value: strengthens the evidentiary basis PCAST and NRC found lacking.
4. Standards for communicating epistemic limits. Rationale: likelihood-ratio science exists but courtroom language often defeats it. Difficulty: requires legal and scientific communities to align on shared phrasing. Value: reduces the gap between probabilistic evidence and categorical testimony.
5. AI systems that recognise unanswerable questions. Rationale: current generative systems tend to answer regardless of evidentiary sufficiency. Difficulty: requires grounding and abstention mechanisms not yet standard in deployed systems. Value: prevents fluent but ungrounded forensic narratives.
6. Narrative bias detection. Rationale: tunnel vision is well documented but rarely measured in real time during active investigations. Difficulty: requires observational access to live casework. Value: enables earlier correction before charging decisions are made.
7. Forensic causal inference. Rationale: association-to-causation conflation remains common in reconstruction. Difficulty: applying formal causal-hierarchy methods (Pearl, 2009) to forensic casework is still nascent. Value: clarifies when causal language is actually earned.
8. Evidence survival and missingness research. Rationale: absence-of-trace reasoning is applied inconsistently across sub-disciplines. Difficulty: requires controlled studies of trace decay across surfaces and time. Value: prevents unwarranted inferences from negative findings.
9. Formal distinction between observation and explanation. Rationale: reports frequently blend the two without flagging the shift. Difficulty: requires reporting-template reform across laboratories. Value: improves transparency for courts and juries.
10. Epistemic abstention frameworks. Rationale: examiners currently have limited formal support for declining to answer a question the evidence cannot settle. Difficulty: requires cultural and institutional change around what counts as a useful report. Value: reduces speculative testimony at its source.

Conclusion: What Questions Can Evidence Never Answer?

The honest answer is neither "none" nor "most." Forensic science produces highly reliable knowledge about observable traces and scientifically testable propositions — the identity of a compound, the class of a mark, the existence of a digital artefact — and modern statistical methods let it state that reliability with genuine rigor (Aitken et al., 2021; Taroni et al., 2014). Evidence can strongly support specific reconstructions when the compatible-hypothesis set has been narrowed by converging traces. It can contribute meaningfully to inferences about intention or motive when combined with context, without ever directly measuring either. Technological improvement genuinely expands what is observable — better sequencing, better metadata recovery, better imaging. Some current limitations are simply evidence not yet collected or methods not yet validated, and those will shrink with investment (NRC, 2009; PCAST, 2016). Others are inferential: multiple accounts remain compatible with the same trace, and no further measurement of the same kind resolves the tie. And some are more fundamental still: questions about subjective meaning, private intention, and counterfactual history depend on access to a person's inner life that no physical or digital residue directly encodes, however completely that residue is measured.

The scientific strength of a forensic conclusion, in other words, depends not only on how much evidence exists, but on whether the question being asked is one the evidence is capable of answering at all. A department with unlimited resources and perfect instruments would still face this boundary, because the boundary is not about instrument quality — it is about what a physical trace, by its nature, can and cannot carry across the gap between a person's action and a person's reasons.

Perhaps the more durable measure of a mature forensic science, then, is not only how much it can discover, but how precisely it can say what it cannot. The next advance in this field may owe as much to the discipline of knowing when to stop claiming as to any new instrument capable of detecting more. The line worth drawing is not between what has been found and what remains to be found, but between what a trace can actually reveal and what an investigation simply wishes it would.

Reference

Aitken, C. G. G., Nordgaard, A., Taroni, F., & Biedermann, A. (2018). Commentary: Likelihood ratio as weight of forensic evidence: A closer look. Frontiers in Genetics, 9, 224. https://doi.org/10.3389/fgene.2018.00224

Aitken, C. G. G., Roberts, P., & Jackson, G. (2010). Fundamentals of probability and statistical evidence in criminal proceedings: Guidance for judges, lawyers, forensic scientists and expert witnesses. Royal Statistical Society.

Aitken, C. G. G., Taroni, F., & Bozza, S. (2021). Statistics and the evaluation of evidence for forensic scientists (3rd ed.). Wiley.

Altman, D. G., & Bland, J. M. (1995). Statistics notes: Absence of evidence is not evidence of absence. BMJ, 311(7003), 485. https://doi.org/10.1136/bmj.311.7003.485

Ask, K., & Granhag, P. A. (2005). Motivational sources of confirmation bias in criminal investigations: The need for cognitive closure. Journal of Investigative Psychology and Offender Profiling, 2(1), 43–63.

Bharatiya Sakshya Adhiniyam, 2023, Act No. 47 of 2023 (India). https://www.indiacode.nic.in/bitstream/123456789/20063/1/aa202347.pdf

Biedermann, A., Bozza, S., Garbolino, P., & Taroni, F. (2022). Bayes factors for forensic decision analyses with R. Springer.

Casey, E. (2002). Error, uncertainty, and loss in digital evidence. International Journal of Digital Evidence, 1(2).

Chiam, S. L., Dror, I. E., & Higgins, D. (2021). The biasing impact of irrelevant contextual information on forensic odontology radiograph matching decisions. Forensic Science International, 327, 110957.

Cole, S. A. (2001). Suspect identities: A history of fingerprinting and criminal identification. Harvard University Press.

Dror, I. E. (2020). Cognitive and human factors in expert decision making: Six fallacies and the eight sources of bias. Analytical Chemistry, 92(12), 7998–8004.

Dror, I. E., Kassin, S. M., & Kukucka, J. (2013). New application of psychology to law: Improving forensic evidence and expert witness contributions. Journal of Applied Research in Memory and Cognition, 2(2), 78–81.

Dror, I. E., Melinek, J., Arden, J. L., Kukucka, J., Hawkins, S., Carter, J., & Atherton, D. S. (2021). Cognitive bias in forensic pathology decisions. Journal of Forensic Sciences, 66(5), 1751–1757.

Dror, I. E., & Wertheim, K. (2011). The impact of human–technology cooperation and distributed cognition in forensic science: Biasing effects of AFIS contextual information on human experts. Journal of Forensic Sciences, 57(2), 343–352.

Findley, K. A., & Scott, M. S. (2006). The multiple dimensions of tunnel vision in criminal cases. Wisconsin Law Review, 2006(2), 291–397.

Giannelli, P. C. (2010). The 2009 NAS forensic science report: A literature review. SSRN. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2039024

Greeshma, K. V. (2023). Epistemic uncertainty in digital forensics: Exploring the boundaries of knowledge. NFSU Journal of Forensic Justice, 2(2), 25–38. https://doi.org/10.62995/jfj1720250617

Haack, S. (2004). Epistemology legalized: Or, truth, justice, and the American way. American Journal of Jurisprudence, 49(1), 43–61.

Haack, S. (2009). Evidence and inquiry: A pragmatist reconstruction of epistemology (2nd ed.). Prometheus Books.

Haack, S. (2014). Evidence matters: Science, proof, and truth in the law. Cambridge University Press.

Inman, K., & Rudin, N. (2000). Principles and practice of criminalistics: The profession of forensic science. CRC Press.

Inman, K., & Rudin, N. (2002). The origin of evidence. Forensic Science International, 126(1), 11–16. https://doi.org/10.1016/S0379-0738(02)00031-2

Kassin, S. M., Dror, I. E., & Kukucka, J. (2013). The forensic confirmation bias: Problems, perspectives, and proposed solutions. Journal of Applied Research in Memory and Cognition, 2(1), 42–52. https://doi.org/10.1016/j.jarmac.2013.01.001

Kukucka, J., Kassin, S. M., Zapf, P. A., & Dror, I. E. (2017). Cognitive bias and blindness: A global survey of forensic science examiners. Journal of Applied Research in Memory and Cognition, 6(4), 452–459.

Kunda, Z. (1990). The case for motivated reasoning. Psychological Bulletin, 108(3), 480–498.

Loftus, E. F. (2005). Planting misinformation in the human mind: A 30-year investigation of the malleability of memory. Learning & Memory, 12(4), 361–366.

Loftus, E. F., & Palmer, J. C. (1974). Reconstruction of automobile destruction: An example of the interaction between language and memory. Journal of Verbal Learning and Verbal Behavior, 13(5), 585–589.

Lynch, M. (2013). Science, truth, and forensic cultures: The exceptional legal status of DNA evidence. Studies in History and Philosophy of Biological and Biomedical Sciences, 44(1), 60–70.

National Research Council. (2009). Strengthening forensic science in the United States: A path forward. National Academies Press.

Nishith Desai Associates. (2023). Navigating criminal law reforms: Part III – Bharatiya Sakshya Adhiniyam 2023. https://www.nishithdesai.com/hotline.aspx/navigating-criminal-law-reforms-part-iii-bharatiya-sakshya-adhiniyam-2023-14926

Pearl, J. (2009). Causality: Models, reasoning, and inference (2nd ed.). Cambridge University Press.

President's Council of Advisors on Science and Technology. (2016). Forensic science in criminal courts: Ensuring scientific validity of feature-comparison methods. Executive Office of the President.

President's Council of Advisors on Science and Technology. (2017). An addendum to the PCAST report on forensic science in criminal courts.

Taroni, F., Aitken, C., Garbolino, P., & Biedermann, A. (2006). Bayesian networks and probabilistic inference in forensic science. Wiley.

Taroni, F., Biedermann, A., Bozza, S., Garbolino, P., & Aitken, C. G. G. (2014). Bayesian networks for probabilistic inference and decision analysis in forensic science (2nd ed.). Wiley.

Taroni, F., Biedermann, A., & Aitken, C. (2023). Statistical interpretation of evidence: Bayesian analysis. In M. M. Houck (Ed.), Encyclopedia of forensic sciences (3rd ed.). Elsevier.