Your Watch Bears Witness - The Doctrine Of Epistemic Mediation And Machine-Generated Evidence In Indian Criminal Trial
Author: LawSphereX
Category: Criminal Law
Published:
Updated:
YOUR WATCH BEARS WITNESS: THE DOCTRINE OF
EPISTEMIC MEDIATION AND MACHINE-GENERATED EVIDENCE IN INDIAN CRIMINAL TRIAL Devyani Rastogi & Gaurish Sharma, O.P. Jindal Global University, Jindal Global Law School
ABSTRACT
Evidence law has traditionally been built to test human testimony, that is, a witness's perception, memory, sincerity, and narration. But an increasing share of criminal evidence today is no longer generated by a human declarant at all. It comes from wearable devices, smart home sensors, and forensic algorithms that sense, process, and interpret physical reality before any human being ever perceives it. Under Indian law, Section 63 of the Bharatiya Sakshya Adhiniyam,
2023 treats this kind of evidence as an ordinary electronic record, verified through
a dual certificate and a hash value that establishes the authenticity and integrity, but it says nothing about whether the interpretive process that produced the recorded conclusion was itself reliable. This article argues that authenticity paradigm is conceptually inadequate to what courts are actually being asked to rely on. It also develops a Doctrine of Epistemic Mediation under which machine-generated evidence is treated as the product of a sequential chain of technological transformations, namely sensing, measurement, signal processing, software processing, algorithmic interpretation, behavioural classification, and legal characterisation, each of which introduces its own, independently assessable risk of error.
Using a hypothetical Indian prosecution modelled on the Connecticut "Fitbit murder" case, the article develops four doctrinal instruments to operationalise this theory: the Principle of Inferential Distance, the Machine Credibility Test, Algorithmic Cross-Examination, and a Digital Forensic Due Process Framework. It builds them within, rather than against, the existing scholarship of Andrea Roth, Edward Imwinkelried, and Sabine Gless. The article also tests the theory against comparative reasoning from the United States and the European Union, against the constitutional structure of Articles 20(3) and 21 of the Indian Constitution, and against the empirical literature on wearable-sensor accuracy, before candidly examining the limits, objections, and institutional risks of the framework it proposes.
I. Introduction
On 23 December 2015, Connie Dabate was shot dead inside her Connecticut home. Her husba. nd, Richard Dabate, t. old police a masked intruder had broken in. He alleged t. hat t. he intruder tied him up, t. hen assaulted him a. nd shot Connie. T. here were no ot. her eye witnesses of t. his incident. T. he Police did recover a firearm, but t. he forensic traces around t. he scene were murky at best. A. nd Richard's account of how t. he intruder moved t. hrough t. he house didn't hold up well against t. he layout of t. he home itself. Taken on its own, none of t. hat physical evidence was enough t. o prove much of a. nyt. hing. What actually broke t. he case was not a person or a document in t. he usual sense, but a small electronic device worn on t. he deceased's wrist. Connie Dabate's Fitbit fitness tracker had logged her movement t. hrough t. he house for roughly a.
n hour after her husba. nd said a. n intruder had already shot her. At trial, t. he device's data was admitted as scientifically reliable after a foundational hearing. A researcher familiar wit. h t. he same model of tracker testified t. o its general accuracy, but ca. ndidly admitted t. hat he did not know how t. he underlying algorit. hm converted raw sensor data int. o t. he step counts t. he jury was asked t. o rely on. Richard Dabate was convicted of murder and is currently serving a sixty five year sentence. 1 . T. he case matters here not because it is unusual but because it is ordinary in every respect except one: t. he decisive witness against t. he accused was a machine, a. nd t. he court's inquiry int. o t. hat witness's trustwort. hiness st. opped at t. he wrist, wit. hout ever going int. o t. he chip itself. T. he expert vouched for t.
he device's accuracy in general, but no one examined t. he specific algorit. hm t. hat turned wrist acceleration a. nd optical pulse readings int. o what t. he courtroom treated as “she was alive a. nd walking. ” T. his gap, between aut. henticating a. n output a. nd actually understa. nding t. he process t. hat produced it, is what t. his article is about. Suppose t. he same case arose before a. n India. n sessions court. T. he prosecution would tender t. he deceased's wearable data as a. n electronic record under section 63 of t. he Bharatiya Sakshya Adhiniyam, 2023 (“BSA”), 2along with a certificate under sub-section (4), one part signed by the person in operational charge of t. he releva. nt computer system a. nd t. he ot. her by a. n expert, a. nd a crypt. ographic hash value showing t. hat t. he file produced in court is identical t. o t. he file extracted from t. he device.
T.his certification architecture is inherited from Section 65B of t.he India.n Evidence Act, 1872, a.nd is now codified in substantially similar form.3 It answers a narrow question: is this the same record, unaltered, t.hat t.he device actually generated? T.hat question is wort.h asking, but it isn't t.he same as asking whet.her t.he process behind t.he record was reliable, whet.her t.he way a device converted . . . . . . a physical event int.o a behavioural label, such as "walking," "running," or "under stress," was itself trustwort.hy. Section 63 was never designed t.o a.nswer t.hat second question.
T. his article's central claim is t. hat t. he second question is a. nyt. hing but a minor footnote t. o t. he law of electronic evidence. It is, t. his article argues, t. he central epistemological problem t. hat machine-generated proof poses for criminal adjudication. Existing doctrine, in India a. nd elsewhere, simply has no settled vocabulary for asking it. Wearable devices serve as t. he main case study here for two reasons. T. hey are already before India. n courts, in matrimonial disputes, insura. nce claims, a. nd increasingly in criminal cases. A. nd t. heir sensing-t. o-inference architecture is well documented in t. he biomedical engineering literature, which mea. ns t. he t. heoretical claims made below ca. n be tested against real data instead of guesswork. But t. he argument itself isn't really about wearables. It reaches t.
he wider category of machine-generated evidence: smarthome systems, aut. onomous-vehicle logs, medical impla. nts, drone a. nd CCTV a. nalytics, biometric a. nd predictive forensic software. Wearables simply offer t. he most tractable, beststudied example of t. hat category. T. his article proceeds in fourteen parts. Part II locates t. he conceptual gap in existing doctrine. Part III examines, rat. her t. ha. n assumes, t. he increasingly common label "machine witness. " Part IV asks what a legal fact actually is when t. he fact in question is a. n algorit. hm's inference. Part V develops t. he philosophical architecture of t. he argument, a seven-stage Knowledge Production Chain grounded in t. he philosophy of technology a. nd evidence epistemology, a. nd tests it against t. he empirical literature on wearable-sensor error. Part VI sets out t. he Doctrine of Epistemic Mediation a.
nd its four instruments. Part VII separates out seven concepts of judicial trust t. hat existing doctrine tends t. o conflate. Part VIII proposes a taxonomy of machine-generated proof. Part IX addresses subsidiary doctrinal questions of hearsay, trade secrecy, burden allocation, a. nd differential sta. ndards of proof. Part X undertakes a comparative a. nalysis of judicial reasoning in t. he United States, t. he Europea. n Union, a. nd India. Part XI tests t. he t. heory against t. he existing literature a. nd states plainly what is, a. nd is not, original about it. Part XII examines t. he t. heory's limits a. nd likely objections. Part XIII proposes a. n implementation pat. hway for India. n courts, a. nd Part XIV concludes. II. T. he Problem: From Testimony t. o Technological Mediation A. What t. he Traditional Approach Tests T. he architecture of t.
he law of evidence, in India as elsewhere, was built around one recurring problem: a. n aggrieved person reports somet. hing t. o t. he court, a. nd t. he court has t. o decide whet. her t. here exists enough evidence t. o support t. heir claims. T. he four traditional testimonial infirmities, defects of perception, memory, narration, a. nd sincerity, are all properties of a huma. n declara. nt Cross-examination exists t. o test exactly t. hese properties: did t. he witness actually see what she claims t. o have seen, has her memory of t. he event decayed or been contaminated, is her account of t. he event accurate, a. nd is she telling t. he trut. h as she believes it or lying. Hearsay is excluded, subject t. o numerous exceptions, mainly because a. n out-of-court declara. nt who is not before t. he court ca. nnot be tested on a. ny of t. hese four dimensions. None of t.
his machinery is well suited t. o a wearable device like a smart watch, a smart home hub, or a forensic classification algorit. hm. A device does not perceive in t. he huma. n sense, it tra. nsduces a physical signal. It does not remember, it st. ores. It does not narrate, it simply outputs a value or a label according t. o a fixed or trained procedure. A. nd it ca. nnot lie, because it has no beliefs t. o misrepresent in t. he first place. Yet a wearable’s classification of a wearer’s cardiovascular state as “stressed,” or a smart lock’s log of a door opening, still functions in t. he courtroom exactly as huma. n testimony once did: as a. n assertion offered for t. he trut. h of what it asserts, which t. he fact-finder is invited t. o rely on. B. T. he Aut. henticity Paradigm a. nd Its Limits Section 63 of t. he BSA, like its predecessor section 65B of t. he India.
n Evidence Act, addresses t. his problem by simply re-describing t. he machine’s output as a document. A “computer output” is deemed t. o be a document, a. nd is t. herefore admissible wit. hout furt. her proof of what it states, if it satisfies four conditions: regular use of t. he computer, t. he ordinary course of t. he activities carried on, t. he proper operation of t. he computer during t. he releva. nt period, a. nd t. he accuracy of t. he information fed int. o it. T. he Supreme Court of India has treated t. his certification requirement as ma. ndat. ory rat. her t. ha. n merely direct. ory. In t. he case of A. nvar P. V. v. P. K. Basheer, t. he Court held t. hat Section 65B lays down a complete code for admitting electronic records4. Without the statutory certificate, secondary evidence of an electronic record simply could not be led. For a while, courts chipped away at t.
hat rule. One line of cases said t. he certificate requirement could be relaxed if getting t. he certificate was genuinely impossible. T. hat relaxation didn't survive. A larger bench in t. he case of Arjun Pa. nditrao Khotkar v. Kailash Kusha. nrao Gora. ntyal stepped in, reaffirmed t. he holding given in t. he A. nvar case, a. nd made clear t. hat t. he certificate is a condition precedent t. o admissibility wit. h only a narrow exception carved out. 5 . . . . T. his body of law is rigorous a. nd it does a. n importa. nt job: it guards against tampering, misattribution, a. nd t. he silent substitution of one record for a. not. her. But every one of its tests is directed at t. he question of aut. henticity, mea. ning is t. his t. he record t. he device actually produced, a. nd at t. he question of proper operation at a general, systemic level, mea. ning was t.
he computer functioning normally during t. he releva. nt period. Neit. her t. he statut. ory text nor t. he case law asks whet. her t. he specific interpretive process by which t. he device converted raw sensor readings int. o t. he label now presented as evidence was itself a sound process. A hash value proves t. hat a file has not been altered since it left t. he device. However, it ca. nnot prove t. hat t. he algorit. hm which produced t. he file’s contents was accurate, well calibrated, validated against a representative population, or free from t. he kind of context-dependent error t. hat t. he biomedical engineering literature, discussed in Part V, shows t. o be common in exactly t. his class of device. T. he conceptual deficiency, in ot. her words, is a category mistake. Existing doctrine treats t. he reliability of a machine-generated conclusion as if it were reducible t. o t.
he integrity of a file. Integrity a. nd reliability are related but different properties, a. nd a legal test built for t. he first ca. nnot, on its own, discharge t. he second. T. he rest of t. his article develops a doctrine addressed specifically t. o t. he second problem, not t. o displace t. he aut. hentication framework of section 63, which remains necessary, but t. o supply t. he inferential-reliability inquiry t. hat framework does not, a. nd structurally ca. nnot, perform. III. Is “Machine Witness” t. he Right Concept? A Critical Examination T. he phrase "machine witness" has begun t. o appear in judicial a. nd academic writing as short. ha. nd for evidence t. hat originates from a device rat. her t. ha. n a person. It is a. n appealing label, largely because it renders somet. hing unfamiliar in terms we already understa. nd but convenience is not t. he same as accuracy, a.
nd a label adopted for convenience risks smuggling in assumptions t. hat do not hold up under scrutiny. Before t. his article makes use of t. he term, four ca. ndidate framings need t. o be tested against what a device actually does.
T.he first option is t.he witness a.nalogy itself. A.ndrea Rot.h, its strongest defender, points out t.hat ma.ny machines now generate statements no programmer or operat.or could have independently verified at the time, some of them with a testimonial character to them.6 Courts, faced with this tend t.o improvise. One might treat t.he output as hearsay. A.not.her might treat it as real evidence. A t.hird might treat it as t.he raw data behind a huma.n expert's opinion.
T.hat doctrinal improvisation is real, a.nd Rot.h rightly points out t.hat some machine outputs are credibility dependent in a way t.hat resembles testimony. But t.he a.nalogy breaks down at t.he point t.hat matters most here: a witness's trustwort.hiness depends on sincerity, on t.he absence of a.ny motive t.o lie. A device has no sincerity t.o assess, no motive t.o interrogate. T.he witness a.nalogy is useful for one reason: it reminds a court t.hat trustwort.hiness needs t.o be assessed before a.n assertion ca.n be relied on. But it obscures somet.hing just as importa.nt: what t.hat assessment actually consists of is not t.he same t.hing at all. . . . . . . .
T.he second option is t.he scientific instrument. A wrist-worn optical sensor is, at its first stage, genuinely a.n instrument. It converts a physical qua.ntity, specifically light absorption in blood vessels, int.o a.n electrical signal. A t.hermometer does somet.hing similar, converting t.hermal expa.nsion int.o a needle position. T.he instrument a.nalogy holds for t.hat first stage. It breaks down for everyt.hing after it. A t.hermometer measures. It doesn't interpret.
However, a wearable's downstream software does interpret. It classifies a pattern of processed signal as "running" or as "stressed." A simple measuring instrument has no equivalent step. A.nd t.hat step introduces a different category of error entirely, one discussed in Part V.
T.he t.hird option is t.o call it a document, or a.n electronic record. T.his is actually t.he label India.n law has gone wit.h. It works well for capturing t.hings like st.orage a.nd reproducibility, which is why Section 63's certification rules hold up reasonably well for t.hat purpose. T.he problem is t.hat t.his framing treats t.he machine's output like a static piece of text, when it isn't one. T.hat's t.he same mistake flagged back in Part II. A document just holds a proposition someone already made. A classification algorit.hm is different. It's t.he one generating t.he proposition in t.he first place.
T.he fourt.h option is t.o treat t.he machine as somet.hing new entirely, a.n "aut.onomous observer," its own category of evidence. Of t.he four options, t.his one is t.he most honest about what's actually happening. But it comes wit.h a real risk: unless it's built on t.op of existing rules for admissibility, disclosure, a.nd appeals, it could end up t.oo vague t.o actually use in court.
Here's where t. his article la. nds. "Machine witness" is a useful phrase, but only as a short. ha. nd for one idea: t. hat trusting a machine's output requires t. he same kind of upfront trust judgment t. hat trusting a huma. n witness does. It shouldn't be taken t. o mea. n t. he device is literally testifying, or t. hat it has a. nyt. hing like a witness's capacities. For t. hat reason, t. he term used t. hroughout t. he rest of t. his article is machine source instead. "Witness" is set aside for a narrower purpose, t. he credibility-testing framework laid out in Part VI. None of t. his conflicts wit. h t. he America. n case law holding t. hat a machine ca. n't be a hearsay declara. nt, since hearsay requires a person making a. n assertion. If a. nyt. hing, t. his article's argument depends on t. hat same idea. In t. he case of United States v. Lizarraga-Tirado, t. he Nint.
h Circuit held t. hat computer generated map coordinates were not hearsay precisely because no person made the relevant assertion7. That reasoning has been applied t. o ot. her categories of machine-generated data as well. T. his is correct on its own terms, a machine is not a declara. nt, but it is incomplete, because it removes machine output from t. he one doctrinal mecha. nism, hearsay, a. nd t. he cross-examination it presupposes, designed t. o test a source’s credibility, wit. hout putting a. nyt. hing in its place. T. he “no declara. nt” holding closes a door. It does not open t. he one t. his article argues is needed. . . .
IV. Jurisprudential Foundation
T.he central argument of t.his article is built on a distinction t.hat evidence scholars tend t.o acknowledge only in passing, rat.her t.ha.n examine directly: t.he distinction between observing somet.hing a.nd inferring it. In most contexts, t.his difference doesn’t carry much practical weight. Here, it does because t.he “observer” in question is a machine a.nd machines do not merely record what happens, t.hey also draw conclusions, often wit.hin t.he same step as t.he observation itself, a.nd t.hat shift from seeing t.o inferring tends t.o pass unnoticed. A. What Is a Legal Fact?
For adjudicative purposes, a fact is usually underst.ood as somet.hing direct evidence could prove, somet.hing a person could have seen or heard a.nd t.hen reported. A raw sensor voltage, a GPS coordinate, a timestamp: t.hese are all close t.o t.hat idea of observation. T.hey're roughly what a witness would describe seeing or hearing, just picked up by a machine instead of a person.
B. Does Machine Generated Evidence Discover Facts, or Construct T.hem?
A smartwatch’s conclusion t. hat its wearer was “running,” or was in a state of “high stress,” is not a. n observation in t. his sense at all. It is a classification, t. he product of a trained or ha. nd-coded model applied t. o processed sensor data, choosing among a fixed vocabulary of behavioural labels t. he one judged statistically most probable. No instrument actually observes “stress. ” What is measured is heart-rate variability, electrodermal activity, or a similar physiological correlate. Similarly, “Stress” is a label t. hat t. he model assigns t. o a pattern in t. hat measurement. It is built int. o t. he model, carrying contestable assumptions about what stress actually is, a. nd about how it's supposed t. o show up in t. he body. In Susa. n Haack's terms, t. he output is really a knowledge claim, a probabilistic, model relative proposition, not t. he kind of fact t.
he law of evidence has traditionally assumed it t. o be.
Haack makes a broader point in her critique of America.n reliability jurisprudence: courts tend to confuse being "scientific" with being reliable8. The label itself gets treated as proof, when it isn't. Somet.hing similar happens wit.h machines. Courts often treat a machine's output as selfaut.henticating just because t.he device looks technically sophisticated. But a number isn't more trustwort.hy because a microprocessor produced it instead of a person. It's only as trustwort.hy as t.he process behind it t.hat hasn't cha.nged, a.nd it isn't going t.o. . . .
C. T.he Doctrinal Consequence: Distinguishing Observation from Interpretation
T. he consequence for evidence law is t. hat it should distinguish, far more sharply t. ha. n it currently does, between raw observations a. nd interpreted conclusions, a. nd should treat t. he latter as carrying a. n accumulated inferential risk t. hat t. he former does not. T. his is not just a sema. ntic point. It determines how far a court’s inquiry has t. o travel before it ca. n responsibly rely on a machine-generated proposition: t. he furt. her a. n output sits from direct tra. nsduction of a physical qua. ntity, t. he more interpretive judgment has been built int. o it, a. nd t. he more its presentation as a settled “fact” is a legal fiction t. hat needs justification rat. her t. ha. n a starting assumption. Part V develops t. he architecture t. hat operationalises t. his distinction, a. nd Part VI builds t. he doctrine on t. op of it. V. T. he Philosophy of Proof: T.
he Knowledge Production Chain A. Technological Mediation is Never Neutral T. his article's t. heory begins wit. h t. he philosopher Don Ihde, who studied how instruments shape t . he way huma. n beings perceive t. he world. His central claim is t. hat no instrument presents reality directly. Every instrument reveals certain features while concealing ot. hers, a. nd t. he two effects ca. nnot be separated. For example, a pair of eyeglasses magnifies vision, yet narrows what ca. n be seen at t. he periphery. A. ny instrument t. hat brings one feature int. o focus does so by pushing a. not. her out of focus. Ihde furt. her observes t. hat instrument readings are not self-expla. nat. ory. T. hey require interpretation, a. nd only a reader wit. h t. he releva. nt training ca. n interpret t. hem correctly. T. his has real consequences for courts. When a court treats a wearable’s print.
out as a direct, unmediated view of what t. he wearer’s body was doing, it makes a mistake. T. he device reveals a. nd conceals at t. he same time, a. nd reading it correctly requires expertise t. he court doesn’t have unless someone supplies it. T. he print. out isn’t a record of what happened. It’s a record of what a chain of sensors, filters, a. nd models concluded had happened, shaped by design choices t. hat were never placed before t. he court. B. Reliability is Not Reducible t. o a Single Judgment T. he second foundation comes from t. he philosopher Susa. n Haack. She argues t. hat trustwort. hy knowledge doesn’t rest on a single foundational fact. It’s built more like a crossword puzzle: ma. ny . . small pieces t. hat check a. nd support one a. not. her, rat. her t. ha. n a t. ower bala. nced on one base.
10That image suits this article’s purpose, since what’s being judged here isn’t a single output but a. n entire chain of steps. No individual step in a machine’s process needs t. o be perfect on its own. What matters is how t. he steps work t. oget. her, checking a. nd reinforcing each ot. her. A weakness in one step ca. n quietly propagate t. hrough t. he rest of t. he chain wit. hout surfacing at t. he end, which is usually t. he only part a court actually sees. C. T. he Seven Stage Knowledge Production Chain T. hese two ideas, from Ihde a. nd Haack, set up t. he rest of t. his Part. Ihde’s argument: every instrument shows you somet. hing by hiding somet. hing else, so a device’s output is never a neutral window ont. o reality. Haack’s argument: trustwort. hy knowledge comes from a chain of steps checking each ot. her, not from one perfect step. T. oget. her, t. hey point t. o t.
he same conclusion. If a device’s output is shaped at every stage, a. nd if trustwort. hiness depends on t. he whole chain rat. her t. ha. n a. ny single link, t. hen machine-generated evidence should be evaluated stage by stage, not treated as one single, all or not. hing fact. T. his article proposes a seven stage architecture for doing exactly t. hat, using a wrist worn wearable (smart watch) as t. he running example t. hroughout.
1. Sensing / Tra.nsduction
A physical qua.ntity, say, light absorption in blood vessels beneat.h t.he skin, or t.he acceleration of a wrist t.hrough space, gets converted int.o a.n electrical signal by a physical sensor. Errors creep in early here: sensor placement, contact quality, ambient light interference. Skin t.one matters t.oo because t.he green light phot.oplet.hysmographic sensors responds differently depending on mela.nin content, which raises bot.h a.n accuracy problem a.nd a fairness problem, taken up again in Part XII.11
2. Measurement / Qua.ntisation
T.he device ca.nnot record a continuous signal forever, so it takes snapshots of it at regular intervals a.nd converts each one int.o a digital number. T.hat process, called qua.ntisation, introduces its own errors. If t.he snapshots are taken t.oo infrequently, or if t.he digital conversion isn't precise enough, or if t.he signal exceeds what t.he sensor ca.n actually capture, t.he resulting numbers ca.n already be off before a.ny furt.her processing even begins.
3. Signal Processing / Filtering
Once digitised, t.he signal is still messy. It gets smoot.hed a.nd clea.ned up t.o strip out motion artefacts a.nd isolate t.he physiological signal underneat.h. T.his step is harder t.ha.n it sounds, because t.he software has t.o assume, in adva.nce, what counts as "noise" a.nd what counts as "signal," a.nd t.hat assumption isn't always right. A filter tuned for someone sitting still ca.n misfire once t.hat person starts moving. It may suppress a signal t.hat's actually real, or invent one t.hat isn't t.here.
T.his isn't hypot.hetical. Wrist worn optical sensors have a well known failure called "signal crossover."12 Limb motion, arm swing during walking, for example, produces its own strong, repetitive rhyt.hm. T.he sensor ca.n mistake t.hat rhyt.hm for t.he heartbeat it's mea.nt t.o be measuring. When t.hat happens, t.he heartrate reading ca.n be wrong at exactly t.he moment of activity most likely t.o matter in court.
4. Software Processing
Sensor data does not reach t.he classifier unt.ouched. T.he device's firmware processes it first by smoot.hing t.he stream, removing readings t.hat fall outside expected ra.nges, converting t.he values int.o sta.ndard units. T.hese are ordinary operations, but ordinary does not mea.n reliable. A defect in t.he processing code ca.n alter t.he output, a.nd so ca.n a firmware update, which may install aut.omatically a.nd wit.hout a.ny record of what it cha.nged. Should t.he device later become evidence, t.he situation is stark: neit.her party may be aware t.hat t.he software was ever modified, a.nd neit.her is positioned t.o determine whet.her t.hat modification affected t.he result.
5. Algorit.hmic / Statistical Interpretation
After t.he signal is processed, a.n algorit.hm, often based on machine learning converts it int.o a measurable output, such as heart rate in beats per minute. Its accuracy depends on how t.he model was designed a.nd whet.her it was trained or validated on a population similar t.o t.he user. Performa.nce also varies wit.h activity. One study found a mea.n absolute percentage error of only 1.2% during running, but 16–17% during badminton and football, where rapid arm movements interfere with sensor readings13. A.not.her study reported t.hat wrist worn devices were least accurate during sudden movement or rapid changes in heart rate, compared to periods of steady activity.14 . . . . .
6. Behavioural Classification
T.he processed data is t.hen assigned a behavioural label like "running," "asleep," "highly stressed" by applying predefined t.hresholds or a classification algorit.hm. T.he reliability of t.hat label turns on several t.hings: where t.he t.hresholds were set, how finely t.he categories distinguish one state from a.not.her, a.nd how t.he underlying model was developed. A classifier well suited t.o everyday fitness tracking may nonet.heless produce t.oo ma.ny false positives t.o be safe in a criminal trial.
7. Legal Characterisation
Presented in court, t.his output is often treated as evidence of a legal fact: t.hat t.he deceased was alive a.nd moving at a given moment, for insta.nce. Yet t.he uncertainty built up across t.he earlier stages has, by now, dropped out of sight. What reaches t.he court looks like a settled factual conclusion, when in trut.h it is t.he outcome of several layered probabilistic judgments.
D. What t.he Chain Shows
T.his evidentiary chain exposes two weaknesses in how machine generated evidence is currently assessed.
T.he first concerns reliability, which does not hold consta.nt across t.he process. A wearable may record physiological data accurately while t.he software t.hat processes a.nd classifies t.hat data introduces uncertainty of its own. T.o ask only whet.her t.he device is "accurate" is t.o miss t.he stages where t.he errors actually arise.
T.he second weakness runs deeper. As t.he empirical studies show, wearables perform least accurately during rapid movement, physical exertion, a.nd sudden physiological cha.nge at precisely t.he moments most likely t.o matter when a criminal event is reconstructed, be it a struggle, a flight, or a collapse. Section 63 of t.he BSA nevert.heless directs its attention t.o aut.henticating t.he final output, not t.o t.he reliability of t.he stages t.hat produced it. Part VI accordingly proposes a stage specific approach t.hat t.he courts must assess t.he reliability of each step in t.he chain rat.her t.ha.n treat t.he final output as trustwort.hy in itself.
VI. T.he Doctrine of Epistemic Mediation
A. Statement of t. he Core Doctrine T. his article proposes t. he following principle for evaluating machine generated evidence in criminal proceedings: Doctrine of Epistemic Mediation: Machine-generated evidence does not directly represent reality. It is t. he end product of a sequence of technological processes t. hrough which physical events are converted int. o digital information a. nd, ultimately, int. o legal proof. Courts should t. herefore assess t. he reliability of each stage of t. hat process, rat. her t. ha. n treating t. he final output as reliable simply because it appears aut. hentic. T. he central idea is straightforward. Reliability ca. nnot be determined by examining only t. he final output. Every stage involved in producing t. hat output has t. he potential t. o introduce error, a. nd each should t. herefore form part of t. he court’s assessment. T. his is not a.
n entirely new insight. As Part XI explains, bot. h Rot. h’s account of machine credibility a. nd Imwinkelried’s work on foundational reliability recognise t. hat machine evidence requires more than simple authentication. 15 . . . T. his article builds on t. hose ideas in a different way. It develops a seven stage framework based on engineering research on error propagation from data collection t. o algorit. hmic inference. Unlike Rot. h’s broader typology, or t. he lifecycle models used in digital forensics, t. he framework focuses on how t. he mea. ning of machine-generated evidence is produced a. nd where reliability may be lost before t. he evidence reaches t. he courtroom. T. he remainder of t. his Part explains four principles t. hrough which t. he doctrine ca. n be applied in practice. B. T. he Principle of Inferential Dista. nce T. he first principle tra. nslates t.
he Knowledge Production Chain int. o a guide for evaluating evidentiary weight. Principle of Inferential Dista. nce: T. he furt. her a machine-generated conclusion is removed from t. he original sensor reading, t. he less evidentiary weight it should receive unless it is supported by independent evidence. Not every machine-generated output deserves t. he same degree of confidence. A raw timestamp or a. n unprocessed GPS coordinate has undergone relatively little interpretation a. nd may t. herefore carry considerable evidentiary weight once aut. henticated. By contrast, a conclusion such as “highly stressed” represents t. he end of a much longer chain of processing, modelling, a. nd classification. Wit. hout independent support such as medical records, witness testimony, or ot. her forensic evidence, its evidentiary value should be treated more cautiously. T. his is not intended t.
o create a new rule of admissibility. Rat. her, it offers a structured way for courts t. o assess t. he weight of machine-generated evidence. Just as t. he law distinguishes between direct a. nd indirect testimony, it should also recognise t. hat different machine-generated conclusions involve different degrees of interpretation before t. hey reach t. he courtroom.
C. T.he Machine Credibility Test
T.he doctrine also requires a structured met.hod for assessing whet.her a particular chain of technological mediation is reliable enough t.o support t.he inference for which it is offered.
Machine Credibility Test: In assessing t.he reliability of machine-generated evidence, courts should consider, so far as practicable: (i) t.he accuracy of t.he releva.nt sensor a.nd its documented error rates for t.he activity in question; (ii) t.he device's calibration hist.ory; (iii) t.he integrity of its firmware a.nd software, including version a.nd update hist.ory; (iv) its cybersecurity safeguards a.nd vulnerability t.o tampering; (v) t.he tra.nsparency of t.he algorit.hm a.nd t.he extent t.o which its classification process ca.n be explained; (vi) t.he existence a.nd quality of independent, peerreviewed validation studies; a.nd (vii) t.he integrity of a.ny cloud synchronisation or data tra.nsmission pat.hway.
T.he fact.ors are drawn from established scholarship on machine-generated evidence but are reorga.nised around t.he seven-stage framework developed in t.his article. T.hey also reflect t.he distinction made by t.he President's Council of Advisors on Science a.nd Technology between foundational validity, which asks whet.her a met.hod is capable of producing reliable results in principle, a.nd validity as applied, which examines whet.her t.hat met.hod was used reliably in t.he particular case. 16 Sensor accuracy, calibration, and independent validation primarily address foundational validity. Firmware integrity, cybersecurity, a.nd data tra.nsmission are concerned wit.h validity as applied, while algorit.hmic tra.nsparency is releva.nt t.o bot.h.
D. Algorit.hmic Cross-Examination
. Reliability ca.nnot be assessed unless t.he defence has a mea.ningful opportunity t.o examine how t.he machine reached its conclusion. Unlike a huma.n witness, however, a machine ca.nnot be questioned in court. T.he scrutiny must instead focus on t.he technological process t.hat produced t.he evidence.
Algorit.hmic Cross-Examination: T.he procedural mecha.nism t.hrough which a party challenges machine-generated evidence by obtaining pre-trial disclosure of t.he releva.nt device or system's design documentation, validation studies, a.nd information necessary for independent testing, including source code or model parameters where appropriate, t.oget.her wit.h t.he opportunity t.o call expert evidence on t.he stages of t.he mediation chain placed in issue.
T.his proposal builds on Christia.n Chessma.n's argument t.hat criminal defenda.nts should be able t .o examine t.he source code of software used t.o generate evidence against t.hem . It also reflects Rebecca Wexler's work showing t.hat trade-secret claims should not operate as a.n evidentiary privilege that prevents meaningful scrutiny18. The expression "algorithmic cross-examination" is t.herefore short.ha.nd for a process of disclosure, independent testing, a.nd expert evaluation. It does not suggest t.hat a.n algorit.hm ca.n be questioned in t.he same way as a witness.
E. T.he Digital Forensic Due Process Framework
. . . T. he doctrine also requires procedural safeguards at every stage of t. he evidentiary process, from t. he seizure of t. he device t. o t. he court’s final assessment of its evidentiary value. T. hese safeguards include aut. hentication under section 63, forensic preservation techniques t. hat avoid altering t. he device, disclosure of software a. nd model versions, independent validation t. hrough t. he Machine Credibility Test, a. nd a. n unbroken chain of cust. ody. Courts should also provide reasoned findings on a. ny disputed stage of t. he mediation chain instead of merely noting t. hat a section 63 certificate has been produced. T. his framework is distinct from t. he lifecycle models used in digital forensics. While t. hose models are designed t. o preserve t. he integrity a. nd cust. ody of digital evidence, t. hey do not examine whet. her t. he processes t. hat tra.
nsformed raw sensor data int. o t. he final evidentiary output were t. hemselves reliable.
VII. A Taxonomy of Machine-Generated Proof
Treating all electronic records as a single category overlooks t.he different risks associated wit.h different forms of machine-generated evidence. T.he reliability concerns raised by a pacemaker, for example, are not t.he same as t.hose raised by a smartwatch or a facial-recognition system. T.o reflect t.hese differences, t.his article groups machine-generated evidence int.o five broad categories based on t.he nature of t.he technology a.nd t.he kinds of errors t.o which it is most susceptible. T.he categories are not mutually exclusive, a.nd a single device may fall wit.hin more t.ha.n one of t.hem.
1. Passive Systems
Medical devices such as pacemakers, continuous glucose monit.ors, a.nd impla.ntable cardiac defibrillat.ors generally have a strong evidentiary foundation because t.hey are subject t.o extensive clinical testing a.nd regulat.ory oversight. T.heir principal legal concern is t.herefore not measurement accuracy but t.he use of sensitive medical data in criminal proceedings. Questions of privacy, consent, a.nd t.he privilege against self-incrimination become particularly significa.nt when information collected for treatment is later relied upon as evidence. T.hese issues are examined furt.her in Part IX.
2. Environmental Systems
Smart home hubs, voice assista.nts, a.nd networked security cameras record information about t.he surrounding environment rat.her t.ha.n about a specific individual. T.hat distinction creates different evidentiary concerns. A smart lock ca.n establish t.hat a door was opened, but not necessarily who opened it. Cloud st.orage a.nd continuous synchronisation also extend t.he chain of cust.ody, increasing t.he possibility t.hat data may be altered, lost, or retained selectively.
3. Mobile Systems
Wearables a.nd smartphones form t.he central focus of t.his article. Because t.hese devices are usually associated wit.h a single user, t.hey ca.n provide highly individualised evidence. At t.he same time, as discussed in Part V, t.hey are especially vulnerable t.o errors during movement, physical exertion, a.nd ot.her tra.nsient physiological states. T.heir evidentiary value t.herefore depends heavily on t.he context in which t.he data was generated.
4. Predictive Systems
Risk assessment t.ools, facial-recognition software, a.nd probabilistic forensic matching systems occupy t.he furt.hest end of t.he inferential-dista.nce spectrum. T.heir outputs are shaped by multiple layers of processing a.nd statistical modelling, making t.hem bot.h highly interpretive a.nd difficult t.o scrutinise. As a result, concerns about tra.nsparency a.nd mea.ningful challenge are particularly acute. T.hese issues are considered in greater detail in Part VIII.
VII. Doctrinal Implications
A. Machine-Generated Evidence a. nd t. he Hearsay Rule Under t. he America. n evidence law, machine-generated output is generally treated as non-hearsay because t. he hearsay rule applies only t. o statements made by a huma. n declara. nt. Alt. hough t. his position is doctrinally sound, it leaves a. n importa. nt gap. By placing machine-generated evidence outside t. he hearsay rule, t. he law also excludes it from t. he framework traditionally used t. o examine t. he credibility of t. he source. No comparable mecha. nism has developed t. o assess t. he reliability of t. he machine or t. he process t. hat produced its output. T. his article does not argue t. hat machine-generated evidence should be brought wit. hin t. he hearsay rule. Instead, it argues t. hat courts need a separate met. hod for testing its reliability. T. he Machine Credibility Test a. nd Algorit.
hmic Cross-Examination are proposed t. o perform t. hat role by directing judicial scrutiny t. owards t. he process t. hrough which t. he evidence was generated, rat. her t. ha. n treating t. he final output as sufficient in itself.
B. Trade Secrets a.nd t.he Right t.o a Fair Trial
One of t. he main barriers t. o Algorit. hmic Cross-Examination is t. he use of trade-secret claims t. o restrict access t. o source code a. nd model parameters. T. his issue has received considerable attention in t. he United States, particularly in litigation involving probabilistic genotyping software. Rebecca Wexler argues t. hat intellectual property rights should not prevent a defenda. nt from examining t. he process that produced the evidence relied upon by the prosecution. 19 The same principle should apply in India. T. he guara. ntee of a fair trial under Article 21 supports disclosure of material necessary t. o test t. he reliability of t. he evidence, while allowing courts t. o protect genuinely confidential information through appropriate protective orders. 20 C. Understa. nding Algorit. hmic Cross-Examination . . . . . T. he expression "algorit.
hmic cross-examination" should not be underst. ood literally. A. n algorit. hm ca. nnot be questioned in t. he same way as a witness. Instead, t. he term describes a procedural mecha. nism t. hrough which a. n opposing party ca. n test t. he reliability of machinegenerated evidence. T. his includes access t. o releva. nt technical material, independent testing where necessary, a. nd expert evidence directed at t. he stages of t. he mediation chain t. hat are genuinely in dispute. T. he objective is not t. o question t. he machine itself but t. o examine t. he process t. hat produced its output.
D. T.he Burden of Establishing Reliability
Where t.he reliability of machine-generated evidence is challenged, t.he burden should fall on t.he party seeking t.o rely upon it, which in criminal proceedings will ordinarily be t.he prosecution. T.his approach is consistent wit.h t.he prosecution's obligation t.o establish guilt beyond reasonable doubt. It also follows naturally from section 63, which already requires t.he party tendering t.he electronic record t.o establish its aut.henticity. Extending t.hat obligation t.o include t.he reliability of t.he process t.hat generated t.he evidence represents a modest but importa.nt development.
E. Different Categories, Different Sta.ndards
T. he taxonomy proposed in t. his article recognises t. hat not all machine-generated evidence raises t. he same concerns. T. he level of scrutiny should t. herefore reflect bot. h t. he category of evidence a. nd its inferential dista. nce. Passive systems, environmental systems, a. nd low-inference outputs from mobile devices, such as raw timestamps or unprocessed location data, will often be capable of sta. nding on aut. hentication alone. Predictive systems, hybrid systems, a. nd high-inference outputs from mobile devices, including behavioural a. nd physiological classifications, require greater caution because t. hey are furt. her removed from t. he underlying sensor data. As a general rule, t. hey should be supported by independent corroborative evidence before t. hey are relied upon t. o establish guilt. T. hat expectation may be displaced only where t.
he Machine Credibility Test demonstrates t. hat t. he evidence is sufficiently reliable.
IX. Comparative A.nalysis: Judicial Approaches Across Jurisdictions
Different legal systems have responded t.o machine-generated evidence in different ways. T.heir approaches reflect different priorities. Some focus on scientific reliability, ot.hers on regulat.ory oversight, while India.n law has largely concentrated on t.he aut.henticity of electronic records. Examining t.hese approaches t.oget.her helps identify bot.h t.he strengt.hs a.nd t.he limits of t.he current India.n framework.
A. T.he United States of America
In t.he United States, questions about machine-generated evidence are addressed t.hrough two distinct bodies of law. Scientific reliability is assessed under Daubert v. Merrell Dow Pharmaceuticals.21 The constitutional right to confront adverse evidence, on the other hand, is shaped by Crawford v. Washington 22 and later decisions such as Melendez-Diaz v. Massachusetts.23 . . . . . . .
T.he distinction becomes importa.nt when evidence is generated by opaque or “black-box” software. A defenda.nt may challenge t.he scientific reliability of t.he technology under t.he Daubert reasoning. Yet t.he Confrontation Clause provides no equivalent opportunity t.o question t.he process t.hat produced t.he result because a.n algorit.hm is not a witness. America.n law t.herefore offers a well-developed framework for evaluating scientific reliability but provides far less guida.nce on how t.he underlying technological process should be scrutinised. T.he model of Algorit.hmic Cross-Examination proposed in t.his article seeks t.o address t.hat gap t.hrough disclosure, independent testing, a.nd expert evaluation.
B. Europea.n Union
T.he Europea.n Union has taken a different approach. Rat.her t.ha.n leaving questions of reliability t.o be resolved during litigation, t.he Artificial Intelligence Act regulates certain AI systems before they are deployed.24 AI systems used by law enforcement to assess the reliability of evidence are classified as high-risk a.nd must satisfy detailed requirements relating t.o risk ma.nagement, documentation, tra.nsparency, a.nd huma.n oversight.
T.his shifts much of t.he inquiry away from t.he courtroom because by requiring extensive documentation before deployment, t.he AI Act promotes greater consistency across systems. Even so, it says comparatively little about t.he weight a judge should attach t.o a particular output in a.n individual case. T.he framework proposed in t.his article complements t.hat model. Documentation prepared t.o satisfy t.he AI Act ca.n also assist courts when applying t.he Machine Credibility Test.
C. India
. . India. n law has developed a rigorous framework for establishing t. he aut. henticity of electronic records. Section 63 of t. he Bharatiya Sakshya Adhiniyam, t. oget. her wit. h its predecessor section 65B of t. he India. n Evidence Act, has been interpreted in t. he cases of A. nvar a. nd Arjun Pa. nditrao as making t. he statut. ory certificate a ma. ndat. ory requirement for admissibility. T. he Bharatiya Nagarik Suraksha Sa. nhita, 2023 has also expa. nded investigative powers relating t. o electronic records. 25 Aut. henticity, however, is only part of t. he inquiry. A section 63 certificate may establish t. hat a. n electronic record accurately reproduces t. he data st. ored on a device. It does not establish t. hat t. he conclusions drawn from t. hat data are correct. A wearable device might classify a person’s activity as “walking” when t.
he individual was in fact being dragged. T. he record may be aut. hentic even t. hough t. he inference is wrong. No reported India. n appellate decision has yet examined t. his issue in t. he context of behavioural inferences generated from wearable devices. T. he opening illustration in t. his article is intended t. o a. nticipate t. hat question before it emerges in litigation.
T. he constitutional framework reflects a similar distinction. Article 20(3) of t. he India. n Constitution protects t. he accused from forced self incrimination, t. hough t. he scope of t. hat protection has never been self-evident from the text, it is judicial interpretation that has done most of the work26. Selvi v State of Karnataka is the clearest example27. There, the Supreme Court held that the guarantee could not be confined t. o conventional confessions, a. nd extended it t. o investigative techniques such as narco-a. nalysis, polygraph testing, a. nd Brain Electrical Activation Profile (BEAP) testing, on t. he reasoning t. hat each of t. hese techniques extracted communicative information wit. hout t. he subject’s voluntary participation. T. he Court st. opped short of collapsing a. n older distinction, t. hough. State of Bombay v Kat.
hi Kalu Oghad had already separated testimonial compulsion from . . . . . . . . . . he compelled production of physical or identifying material, a. nd Selvi left t. hat line intact . So, Article 20(3), as it sta. nds, protects compelled communication specifically, not every form of evidence a court might obtain from a. n accused person.
Privacy presents a.not.her, separate challenge. Since t.he K.S. Puttaswamy case, Article 21 has been understood to protect personal privacy, subject to the requirement of proportionality.29 Wearable devices often record continuous streams of information extending well beyond t.he period releva.nt t.o a criminal investigation. T.he question is t.herefore not only whet.her t.he data is reliable, but also whet.her t.he scope of forensic extraction is proportionate. T.he Digital Forensic Due Process Framework proposed in t.his article is intended t.o operate wit.hin t.hose constitutional limits.
X. Limits, Objections, a.nd Unintended Consequences
t . . . . No doctrinal framework is wit.hout limitations. T.he proposal adva.nced in t.his article is no exception. Several objections require careful consideration. T.his is because t.hey expose t.he practical limits of t.he framework a.nd also identify issues t.hat a.ny future implementation would need t.o address.
A. T.he Risk of Over-Inclusiveness
If every item of machine-generated evidence were subjected t.o stage-by-stage scrutiny, t.he framework could become unnecessarily burdensome. Routine evidence, such as a server timestamp, would rarely justify t.hat level of examination, a.nd applying t.he full framework in every case would place additional pressure on already congested trial courts. T.he Principle of Inferential Dista.nce is intended t.o avoid t.hat result. Outputs involving only limited interpretation should ordinarily require not.hing more t.ha.n aut.hentication under section 63. T.he more dema.nding Machine Credibility Test is reserved for evidence t.hat has passed t.hrough multiple interpretive stages.
B. Institutional Capacity
T.he framework also assumes a level of technical expertise t.hat ma.ny trial courts may not possess. Questions about firmware integrity, calibration records, or model validation often require specialised knowledge. T.here is also a risk t.hat t.he proposed framework could become a.not.her procedural formality, much like t.he section 65B certificate sometimes did in practice. Existing institutions offer a partial a.nswer. T.he Examiner of Electronic Evidence, recognised under section 79A of the Information Technology Act, 2000, can provide technical assistance where necessary.30 Courts should also explain t.heir conclusions on each disputed stage of t.he mediation chain instead of treating t.he exercise as a routine procedural requirement.
C. T.he Importa.nce of Corroboration
. . . A furt.her concern is t.hat a preference for corroboration may lead courts t.o discount reliable evidence simply because no independent source is available. T.hat concern is legitimate, but it does not require aba.ndoning t.he framework. Corroboration is proposed as a general expectation rat.her t.ha.n a.n inflexible rule. Where t.he Machine Credibility Test demonstrates a high level of reliability, a court should remain free t.o rely on a single source of machine-generated evidence.
D. Unequal Access t.o Information
T.he proposed framework ca.nnot eliminate every practical obstacle. Technology compa.nies may continue t.o rely on trade-secret claims t.o restrict access t.o source code a.nd ot.her technical information. Unless t.he law recognises t.hat a defenda.nt’s right t.o a fair trial takes priority in appropriate cases, Algorit.hmic Cross-Examination may exist in t.heory while remaining difficult t.o exercise in practice. A.ny implementation of t.he framework should t.herefore address disclosure directly instead of leaving t.he issue t.o be resolved case by case.
E. Aut.omation Bias
Even a well-designed legal framework may not prevent judges or juries from placing excessive trust in machine-generated evidence, research on aut.omation bias shows t.hat decision-makers often give greater weight to automated outputs than to conflicting evidence.31 This is exactly where procedural safeguards become importa.nt. Ma.ndat.ory disclosure is one part of it. So is placing t.he burden on t.he party relying on t.he evidence t.o establish its reliability. A.nd judicial findings t.hat . . . . . . are actually reasoned, rat.her t.ha.n assumed, push t.oward closer scrutiny instead of unquestioning accepta.nce.
F. Alternative Approaches
T.he framework proposed in t.his article is not t.he only possible solution. One alternative would be t.o incorporate Daubert style reliability review t.hrough a purposive interpretation of t.he Bharatiya Sakshya Adhiniyam. A.not.her would be t.o adopt Rot.h’s model of testimonial safeguards. A t.hird would require t.he use of interpretable or “glass-box” systems, following t.he approach suggested by Garrett and Rudin32. These approaches need not compete with the Doctrine of Epistemic Mediation. Instead, t.hey address different aspects of t.he same problem. T.he doctrine proposed here provides a broader conceptual framework by distinguishing between observation a.nd inference a.nd by recognising t.hat different machine-generated outputs involve different levels of interpretation before t.hey reach t.he courtroom.
XI. T.he Proposed Framework
. . . . T.he principles developed in t.his article ca.n be incorporated int.o existing evidentiary practice wit.hout waiting for legislative reform. Much of t.he proposed framework ca.n be implemented t.hrough a purposive interpretation of t.he Bharatiya Sakshya Adhiniyam, particularly its provisions governing electronic records a.nd expert evidence.
T.he starting point remains uncha.nged. Courts should continue t.o aut.henticate electronic records under section 63 by requiring t.he statut.ory certificate, t.he applicable hash values, a.nd complia.nce wit.h t.he existing statut.ory requirements. T.he framework proposed in t.his article supplements t.hat process; it does not replace it.
T.he next question is t.he nature of t.he evidence before t.he court. Using t.he taxonomy developed in Part VII, judges should identify t.he category int.o which t.he machine-generated evidence falls a.nd determine its inferential dista.nce. Evidence t.hat involves little interpretation, such as raw timestamps or unprocessed location data, will usually require no furt.her inquiry unless its reliability is specifically challenged.
Where t. he evidence is furt. her removed from t. he original sensor data, additional scrutiny becomes necessary. Courts should t. hen apply t. he Machine Credibility Test, wit. h t. he burden resting on t. he party seeking t. o rely on t. he evidence. A. ny available regulat. ory documentation, including material comparable t. o t. hat required under t. he Europea. n Union's AI Act, should also be considered where releva. nt a. nd a mea. ningful opportunity t. o test t. he evidence should follow. T. his requires disclosure of technical material a. nd, where appropriate, independent expert examination. Legitimate claims of confidentiality ca. n be addressed t. hrough protective orders, but t. hey should not prevent a party from examining t. he basis of t. he evidence relied upon against t. hem. T. he court should t. hen determine t. he weight t. o be attached t. o t.
he evidence by applying t. he Principle of Inferential Dista. nce. High-inference outputs will ordinarily require independent corroboration before t. hey are relied upon t. o establish guilt. T. hat expectation may be displaced where t. he Machine Credibility Test demonstrates t. hat t. he evidence is sufficiently reliable.
Finally, t.he court should explain its conclusions on every disputed stage of t.he mediation chain. Clear reasons not only improve t.he quality of decision-making but also allow mea.ningful appellate review. T.hey provide a.n importa.nt safeguard against bot.h institutional limitations a.nd t.he tendency t.o place excessive confidence in aut.omated systems.
T.he framework is intended t.o be introduced gradually. T.he first stages ca.n be adopted wit.hin t.he existing statut.ory framework, while later stages would benefit from judicial practice directions, delegated legislation under t.he Bharatiya Sakshya Adhiniyam, or future legislative amendment. One possibility would be t.o supplement t.he existing aut.henticity certificate wit.h a separate certificate addressing inferential reliability.
XII. Conclusion
T.he smartwatch worn by Connie Dabate did not witness her deat.h in t.he way a person could. It recorded physical signals, processed t.hem t.hrough software, a.nd generated a conclusion t.hat t.he prosecution relied upon t.o argue t.hat she was alive a.nd walking. Each step involved interpretation, a.nd each carried t.he possibility of error.
Yet t.he trial focused largely on t.he overall reliability of t.he device rat.her t.ha.n on t.he specific processes t.hat produced t.he disputed output.
India.n evidence law is well equipped t.o establish t.he aut.henticity of electronic records t.hrough section 63 of t.he Bharatiya Sakshya Adhiniyam. Aut.henticity alone, however, ca.nnot a.nswer every question t.hat machine-generated evidence presents. A record may be genuine even if t.he inference drawn from it is mistake.
T.he Doctrine of Epistemic Mediation, t.oget.her wit.h t.he Principle of Inferential Dista.nce, t.he Machine Credibility Test, Algorit.hmic Cross-Examination, a.nd t.he Digital Forensic Due Process Framework, is intended t.o address t.hat una.nswered question by providing a structured met.hod for evaluating inferential reliability.
Wearable devices have provided t.he principal case study because t.heir limitations are well documented a.nd t.heir accuracy often depends on t.he circumsta.nces in which t.hey are used. T.he underlying argument, however, extends beyond wearable technology. Smart homes, aut.onomous vehicles, medical impla.nts, a.nd predictive forensic software are becoming increasingly importa.nt sources of evidence in criminal proceedings. As courts rely more frequently on t.hese technologies, t.hey will need t.o ask more t.ha.n whet.her a.n electronic record is aut.hentic. T.hey will also need t.o consider whet.her t.he process t.hat tra.nsformed physical events int.o legal evidence was reliable. Meeting t.hat challenge is essential if machine-generated evidence is t.o support findings of guilt beyond reasonable doubt.