The Observer in the Loop: Infrastructural Reflexivity When Researcher and Research Object Share a Substrate

1. Introduction: The Observer in the Loop

Here is the problem with reflexivity in digital research: we have learned to interrogate who is looking, and forgotten to ask what is doing the looking.

Consider a concrete moment. I construct it as a composite and say so, because this article’s evidential standard applies to itself; the logged incidents appear in Section 3.

A researcher is auditing the design of commercial griefbot platforms, services that use generative models to simulate the conversational presence of the dead. Her codebook asks three questions of each. Does the system retain the deceased person’s data by default, so that persistence is what happens when nobody intervenes? Is the reply marked as generated, or is the mediation smoothed away? Is there an exit, and what does exiting cost? She is coding platform text with the help of a large language model, because the corpus is large and the deadline is real.

Somewhere in the third session she notices something. Her coding assistant retains their conversation by default. Its replies are frictionless, unfailingly available, never tired, never refusing. And the status of its output is exactly what her codebook interrogates: is this a reply, or a very good playback of the shape of a reply?

Her instrument passes, or fails, her own instrument’s test.

This is not the discovery of a bias. Bias is what reflexivity was designed to handle, and it handles it well. From Writing Culture through Bourdieu’s participant objectivation to the feminist holistic reflexivity that Cooky and colleagues developed for social media research in this journal, the tradition taught researchers to objectify the objectifying subject and to treat their own position as data rather than neutral vantage (Clifford and Marcus, 1986; Bourdieu, 2003; Cooky et al., 2018). That tradition is not wrong and I do not propose to replace it. But every version of it shares a premise that has become false without anyone announcing it: that the researcher’s cognitive instruments, the notebook, the tape recorder, the statistics package, the coding software, are of a different kind than the thing studied. The instrument might distort. It could not be a specimen.

An LLM used to study LLM-mediated systems is a specimen. This is the condition I want to name and give a method for.

Two literatures should have caught this and, for different reasons, have not. The first is the debate over generative AI in qualitative inquiry, which over two years has become sharp and in places angry. Jowsey and colleagues, with 419 qualitative researchers across 32 countries, reject generative AI for reflexive approaches on the ground that “analytic approaches such as reflexive thematic analysis are human research practices requiring a subjective, positioned, and reflexive researcher and therefore the use of GenAI in such approaches is not methodologically congruent” (Jowsey et al., 2025). Against them, Friese would replace coding with structured dialogic interaction (Friese, 2026), Costa and colleagues advance AI as co-researcher and epistemic agent (Costa et al., 2025), and Christou argues reflexivity must become forward-facing, since today’s outputs become tomorrow’s training data (Christou, 2026). Empirical work documents the tensions without resolving them (Schroeder et al., 2025; Hayes, 2025).

The second is this journal’s own project. Critical data studies has spent a decade demonstrating that data infrastructures are neither neutral nor transparent, and has built methods for making them visible: the ontology explorer for infrastructures of population management (Van Rossem and Pelizza, 2022), data diaries for following data through situated practice (Tkacz et al., 2021), the algorithm audit for scoring the systems that score us (Brown et al., 2021). More recently it has turned to generative systems as objects, and, in Colombo and colleagues’ account of prompt design, as research instruments (Colombo et al., 2026).

So far, so good. But notice what these positions have in common, and that the commonality differs by case.

The rejectionist and the enthusiast share an assumption directly: each occupies a position from which the AI can be refused or admitted, and both decisions are taken by someone standing outside the thing decided about. That position is unavailable here, for a reason having nothing to do with willpower. The infrastructure supplying the instrument is the infrastructure supplying the object: the same model families, the same retention architectures, the same interface conventions, the same commercial platform relations. A researcher can refuse the coding assistant. She cannot thereby leave the infrastructure, since her retrieval, storage, literature discovery, translation, and institutional access conditions run on it too. Refusal changes her exposure. It does not change her position.

The critical data studies literature has the opposite gap, and it is the more interesting one. It has never assumed an outside; infrastructural inversion was invented precisely to deny one (Bowker and Star, 1999). What it has consistently done is point the apparatus outward. The ontology explorer makes their infrastructures visible. The algorithm audit scores their algorithms. Even the reflexivity that this journal has published, holistic and situated as it is, concerns the researcher’s engagement with the field rather than the kinship between her tools and her objects (Cooky et al., 2018). The instrument has remained the thing that does the looking. My proposal is that the field now owns a procedure it has not yet turned around.

The posthumanist strand is a third case and the closest competitor, because it already denies the outside vantage. Distributed-agency accounts do not claim the researcher stands apart; they claim she was never a bounded observer. My disagreement is not about exteriority, which we both reject, but about what survives its rejection. An audit requires a signatory. If agency is distributed such that no one is the observer, no one is accountable for the observation, and the reflexive record has no author to be held to it.

My thesis is straightforward. When researcher and research object share a technical substrate, reflexivity must shift its object from the researcher’s social position to her infrastructural position; and the legitimate form of such reflexivity is not completeness but calibration. The completeness model asks for full self-transparency, fails, and by failing licenses the conclusion that the enterprise is hopeless. The calibration model asks for something else: disciplined divergence, layer-by-layer honesty, a principled halt, and a published record of where the halt was taken and what depends on it. It trades transparency for auditability, and auditability is what a reader can check.

The article proceeds in six movements. Section 2 distinguishes infrastructural from social position and shows why the distinction survives its strongest objection. Section 3 works out three layers at which the shared substrate bites. Section 4 states the regress objection, situates it in the sociology of scientific knowledge where it was already fought to a standstill, and shows that it defeats only one of two models. Section 5 demonstrates the model through tool-as-case. Section 6 answers four objections. Section 7 draws the concepts together.

A word on what this article does not do. It reports no empirical findings; the audit whose reflexive log supplies its illustrations is planned, and not reported here. It does not claim AI-mediated research is invalid, and Section 6 argues the reverse. And it does not claim that any formal result about self-reference proves anything about methodology. Section 4 contains one footnote on the structural analogy and stops there, deliberately.

2. From Social Position to Infrastructural Position

This section is not claiming that the reflexive tradition failed to notice technology. It claims it noticed technology as mediation, not kinship.

Begin with the strongest available version, because weaker versions are not what this argument must survive. Bourdieu’s participant objectivation is explicitly not the narcissistic self-disclosure he attributed to postmodern anthropology, and explicitly not phenomenological introspection. It is the demand that the researcher turn the most objectivist tools of social science onto the anthropological field itself and onto the scholastic dispositions that field fosters and rewards in its members (Bourdieu, 2003; Bourdieu and Wacquant, 1992). The object of objectivation is not the private person of the enquirer but the conditions of possibility of her act of objectivation. Adjacent lineages sharpen different aspects: situated knowledge holds that all vision is from somewhere and that the god trick of a view from nowhere is both false and irresponsible (Haraway, 1988); reflexive ethnography holds that the ethnographic text is itself a construction with rhetorical and political conditions (Marcus, 1998); and in this journal, holistic reflexivity extends the demand across the whole arc of a social media study, from data access through analysis to representation (Cooky et al., 2018). Each is right about the observer’s situation.

Harding marks the boundary of what this article can claim. Her argument is that starting inquiry from the lives of the marginalised yields better knowledge, because the dominant standpoint has a systematic interest in not seeing (Harding, 1995). That is an explicit claim of epistemic privilege, and one the calibration model could not make. Strong objectivity works because standpoints differ across people, so a community’s aggregate vision improves by recruiting the ones it lacks. Section 3.3 argues that model-mediated coding attacks that assumption, distributing the same omissions to everyone using the same model. The corrective Harding relies on is the corrective correlated omission is positioned to defeat. My disagreement is not that she overclaims but that the mechanism licensing her claim is being eroded.

Now the strongest objection to my distinction, put before I state it, because a Bourdieusian reviewer will make it first and the distinction is worthless if it cannot take the hit.

The scholastic disposition is not a bias of attention. It is a constraint on the form of knowledge producible. Bourdieu’s point is that the scholastic point of view generates a determinate kind of account, the account of someone with the leisure to treat practice as a spectacle, and that this shapes not what the analyst notices but what her analysis can be. If that is right, “infrastructural position” is participant objectivation applied to a new kind of tool, and there is no new object here at all.

I concede the premise entirely. Scholastic disposition does constrain form, and any version of my distinction resting on “attention versus form” is defeated by this and should be abandoned. The difference lies elsewhere.

Participant objectivation is possible because the researcher, however deeply formed by her field, is not a case in her own study. The sociologist who objectifies her scholastic disposition does not thereby enter her own sample. She ascends: she constructs a second-order account whose object includes her first-order account, and the ascent is available precisely because the objectifying apparatus and the objectified population remain distinct populations. Reflexivity of this kind moves up.

Infrastructural position removes the condition for the ascent. The thing constraining her account is not a disposition she carries into the field; it is an artefact of the same technical kind as the artefacts in her sample, produced by the same firms, exhibiting the same design patterns, answering to the same criteria in her codebook. To objectify it is not to construct a meta-level. It is to add a case. The move is lateral.

That difference is not verbal. The objectification can be performed with the study’s existing instrument and its result is commensurable with the study’s other results, both of which Section 5 exploits; and, least comfortably, delivers none of the epistemic relief that ascent does. Having objectified her scholastic disposition, Bourdieu’s analyst stands somewhere better than she did. Having coded her instrument, our researcher stands in the same place with a longer table.

Call this condition the researcher’s infrastructural position: her location within, and dependence upon, the technical arrangements that also constitute her object. Two conditions must hold together, and exactness matters, because a concept applying to everything explains nothing.

Substrate sharing. Instrument and object are instances of the same technical kind, not merely built with related technologies or running on shared hardware.

Predicate overlap. At least one predicate in the researcher’s analytic scheme is true, or arguably true, of the instrument.

Substrate sharing without predicate overlap gives an irony; predicate overlap without substrate sharing gives a metaphor. Together they give a methodological problem: the codebook now has a case it was not designed to have and cannot legitimately refuse.

A spreadsheet used to analyse spreadsheets fails both conditions, and that case is easy. The instructive one is nearer the line. Suppose a researcher studies the design of survey platforms, coding them for how their defaults shape respondent behaviour, and administers her own instrument on a survey platform. Substrate sharing holds. Predicate overlap does not, because the properties her codebook detects (default question order, forced-response settings, progress indicators) govern how her respondents behave and have no purchase on how her analysis is generated. Her instrument is in the population; it is not in the codebook’s reach. Condition two fails, and it should. The condition is not “the researcher uses a technology like the ones she studies” but “the researcher’s analytic scheme applies to her own analytic apparatus.”

Whether generative systems satisfy the first condition strongly is contested, and the contest matters. Schulz-Schaeffer argues that generative AI differs from designed technology in task-relatedness, user interaction, and agency: unlike a tool built for a determinate task, a generative system’s task is constituted in the interaction (Schulz-Schaeffer, 2025). If that is right, and I think it is, the instrument is of the same kind as the object in the specific respect that makes both hard to audit, namely that neither has a task specification independent of use. Substrate sharing here is not a family resemblance. It is a shared indeterminacy.

The word “infrastructure” is doing precise work, and I take it in the sense infrastructure studies established: relational rather than substantial, embedded, learned as membership, and, decisively, transparent in use and visible upon breakdown (Star, 1999; Bowker and Star, 1999). Three do direct work below. Transparency-in-use is why the researcher of Section 1 notices her instrument in the third session and not the first: nothing had broken until the codebook’s predicate arrived. Learned-as-membership is why refusal is not exit. And visibility-upon-breakdown is why the log field reserved for refusals, hedges, inconsistencies and post-update degradations is the most valuable in the instrument. Those are not annoyances to be minuted. They are the field’s own canonical method of making infrastructure appear, arriving unbidden.

The tradition tells us that the researcher is positioned. What it does not tell us is how to proceed when the position is occupied by something that is also a case. Here the second-order cybernetic lineage supplies the formal point, and I take it up for that reason rather than for its prestige. Von Foerster’s distinction between the cybernetics of observed and of observing systems established that an account of observation excluding the observer is incomplete in a way further data cannot repair (von Foerster, 2003). Luhmann’s development supplies the operative claim: every observation applies a distinction, and the distinction in use cannot be observed by the observation using it. The blind spot is not inattention. It is the condition of the observation’s being an observation (Luhmann, 1998; on second-order observation as a social form, see Moeller, 2017).

So the blind spot cannot be eliminated, only relocated: a second observation sees the first’s distinction and applies one of its own that it cannot see. Reflexivity does not converge on transparency. It moves. The deflationary consequence is familiar, that the confessional positionality statement performs a transparency its own form makes impossible. The productive one is where the contribution begins. If the blind spot is structural, the response is not to try harder at seeing but to change what the reflexive act produces. A confession produces a claim about the researcher; an audit produces a record a third party can check. That difference matters most where transparency is least available, because a record does not require its author to have seen everything, only that what she did see be stated so someone else can find what she missed.

One clarification, since the isomorphism language below invites a fair complaint. To say that a researcher’s account can take on the formal properties of her instrument is not a claim immune to disconfirmation. It is disconfirmed, locally, whenever the instrument codes differently from the object cases on the criteria that matter. Section 5’s comparison is that test, it can come out either way, and on two of six criteria it came out against the claim.

Reflexivity’s old question was: who is looking? The new question is: what is doing the looking, and is it the same kind of thing as what is being looked at?

3. Three Self-Referential Tensions

This section is not claiming that AI-assisted coding is uniquely corrupt, nor that these tensions exhaust the field. It claims they are the three where substrate sharing and predicate overlap meet, and that each is observable rather than merely arguable.

I take them from a research programme on grief technology in which I developed three concepts, and use them for a methodological reason rather than a convenient one: each was built to detect a property of commercial systems, and each turns out true of the research instrument. That is predicate overlap satisfied three times, which makes this a case rather than an anecdote.

A note on evidence, and on a decision that costs something. The illustrations come from a reflexive log kept across the programme, of which a planned pilot design audit is one phase. The entries below come from another, the preparation of this manuscript, where the substrate condition currently bites hardest; they are reproduced as Supplementary Material 2. Section 6 answers the objection this invites. The log is a near relative of the data diary, which follows data through the situated practices producing it (Tkacz et al., 2021), differing in what it follows: not the data but the instrument, not its journey but its coding profile. Nothing has been supplied from memory or reconstruction, because a paper arguing for auditability cannot afford an unaudited illustration.

3.1. The Infrastructural Layer: Retention as a Shared Default

The first of the three concepts concerns the digital afterlife industry, where the decisive normative work is done not by what transhumanist discourse says but by what platform defaults do. That providers must be recommended to develop procedures for retiring deadbots (Hollanek and Nowaczyk-Basińska, 2024) indicates termination is not designed in; the resting state is availability.

Apply the criterion to the instrument. The researcher’s conversation with her coding assistant is retained by default. Her prompts, half-formed categories, discarded hypotheses, and in many configurations her source material persist in a commercial system unless she acts. Retention is what makes the assistant useful across sessions. Her relation to it has the structure her codebook flags: an encoded verdict, operating by default, independent of user belief.

Let me be precise about the level of description at which I claim identity. I do not claim that retaining a mourner’s simulated correspondent and retaining a researcher’s chat log carry the same moral weight; they plainly do not. I claim that at the level the codebook operates on, the default and its doxastic independence, the two code the same. A criterion that cannot distinguish a griefbot from a chat log is too coarse for the study, which is a finding about the codebook rather than a flourish about the instrument.

Logged, 1 August 2026. A retrieved source exceeded an output threshold and was written to a file instead of returned, by a default neither chosen nor announced; the file was then unreadable from the process needing it, the two not sharing a filesystem. The source was abandoned and obtained elsewhere. Nothing failed: the default worked as designed, costing a citation route.

3.2. The Voice Layer: Smoothness as an Engineered Property of the Workflow

The second, smooth voice, names a configuration in which polish, media infrastructure, and trained listening reduce friction so that mediated authority sounds immediate and resolved. Smoothness is not the absence of technique but its successful concealment.

The coding assistant is a smooth text generator in that sense. Its output arrives without hesitation markers, without the fatigue that in a human coder signals the end of reliable attention, without the irritability signalling a strained category. It produces the eleventh code with the fluency of the first.

The consequence is sharper than a warning about over-trust, whose weaker version is already in the literature (Schroeder et al., 2025). It is that fluency may have removed a signal the method depended on. I offer this as a conjecture, knowing of no study that tests it: coder fatigue plausibly functions not only as a defect but as information, marking where the codebook has begun to strain. A workflow that engineers the fatigue out then produces not better coding but coding whose strain is silent.

Logged, 1 August 2026, twice. Five passages rewritten to shorten the manuscript and labelled as cuts proved, on measurement, eighty-one words longer than what they replaced. Read, they seemed tighter; counted, they were not. Separately, a verification routine that had completed cleanly began timing out and then terminated under memory pressure, four attempts running, so the artefact it was to check shipped unverified, the limitation disclosed. The second is the more valuable entry: infrastructure became visible by breaking, and the cost reached the deliverable rather than the workflow. Breakage cannot be scheduled but can be provoked: version drift is announced, not accidental, so a subsample re-run after a deprecation notice is a planned friction probe.

3.3. The Reply Layer: Correlated Omission

The third distinguishes continuing bonds, relations transformed around absence, from simulated presence, which stages a novel answer in absence’s place: where revenue depends on renewed interaction, an answerable relation can turn the managed nonfinality of grief into an asset. Its pivot is the status of the reply: whether what returns is a trace of what was said or a new utterance in the shape of the one who said it.

Applied to the instrument, the tension is sharpest. When the assistant proposes a code, what is that proposal? Not a finding, since nothing was found; not an opinion, since no one holds it. A generated continuation, shaped like an analytic judgement, of a corpus of analytic judgements.

Two easy answers must be resisted. That this makes the output worthless is false: such continuations are often good analytic judgements (Hayes, 2025). That verification dissolves the problem is closer but wrong. Verification changes the status of retained codes. It cannot touch codes never generated, because a code that does not appear leaves no trace.

Now the obvious rejoinder. Human coders also fail to generate codes, and those omissions are equally uncheckable. You have described a general problem of interpretive work under an infrastructural name.

The rejoinder is right about the asymmetry and wrong about its novelty. What is new is not that omissions are uncheckable. It is that they are correlated.

A human coder’s omissions are distributed by an idiosyncratic history: her training, her commitments, her reading, her institutional formation. That distribution is uneven across researchers, which is why inter-coder work, adversarial collaboration, and the whole reflexive apparatus have purchase. Two coders trained differently miss different things, recoverable by putting them in a room. A great deal of qualitative rigour silently rests on that assumption.

A model’s omissions are distributed by a training corpus, a fine-tuning regime, and an alignment procedure the researcher cannot inspect, cannot vary, and did not choose. They are opaque, since no reflexive effort reaches them; shared, since every researcher on that model version inherits the same distribution, so the field’s usual corrective fails silently when both rooms hold the same model; and cognate with the object, produced by a process of the same provenance as the systems under study.

Field-wide correlated omission is the new object, and not a version of anything reflexivity previously handled. A discipline can absorb many individually blind researchers so long as they are blind in different directions. It cannot absorb a shared blindness presenting itself as independent agreement. The failure mode is inverted from the one the literature worries about: the danger is not that assisted coding disagrees with human coding, which is visible and corrigible, but that assisted codings agree for reasons having nothing to do with the data.

Logged, 1 August 2026. The publication-order rule stated in Section 5 was drafted by the assistant, accepted without independent derivation, then applied in an order that violated it. Neither drafting party noticed: the error survived the session that made it. It was caught later, in the author’s review of the draft assisted by model critique sessions; no third-party reader was involved. The corrective came from outside the drafting moment, not from outside the pair. Correlated omission, at close range. The quantitative anchor this entry calls for is unavailable: the log was not built to count. Recording the absence is what the criterion requires; an estimate is what it forbids.

Three layers, three overlaps: in each, the property the codebook detects is instantiated in the detector. That makes the condition structural rather than anecdotal, and an objection unavoidable.

4. The Regress Objection and the Two Models

Here is the objection, in its strongest form.

Whatever reflexive apparatus you build is itself an observation, and therefore has a blind spot. You propose to include your instrument in your analysis. Very well: the inclusion is performed by a process you have not included. Include that, and the inclusion of the inclusion is performed by a further process. There is no level at which the series terminates, because the reason it does not terminate at level n is the reason it does not terminate at n+1. Your method is therefore either incomplete at every stage, in which case its claim to reflexivity is unearned, or arbitrary, since you stopped somewhere and nothing explains why there. Adding a level of self-reference to an unreliable procedure gives you an unreliable procedure with a preface.

The objection is correct, and it is not new. It was fought at length in the sociology of scientific knowledge, and the fight is worth recovering, because the current AI-and-qualitative-method debate is re-running it without the benefit of the outcome. Woolgar’s Knowledge and Reflexivity mapped a continuum from benign introspection, which keeps object and representation distinct, to radical constitutive reflexivity, which privileges no account over another (Woolgar, 1988). Ashmore catalogued the mechanisms by which the field tried to live with the regress, including the literary forms it invented to perform rather than argue its way out (Ashmore, 1989). Pollner traced radical reflexivity’s decline (Pollner, 1991). And Lynch, drawing the balance, argued against reflexivity as an academic virtue and source of privileged knowledge altogether (Lynch, 2000).

Two things follow from taking that history seriously, and they cut in opposite directions.

The first is a warning against my own novelty. The completeness/calibration pair is not Woolgar’s continuum relabelled, but a reader is owed the distinction. His axis measures how radically an account destabilises the object/representation relation; mine measures what the reflexive act produces, a claim about the author or a record for a reader. The axes are orthogonal: a benignly introspective researcher and a radically constitutive one can both produce claims, and either could produce records. What follows is not a new position on Woolgar’s continuum but a different question asked of any position on it.

The second is that Lynch’s argument, the most damaging in that literature, is not an obstacle here. It is the warrant.

Lynch’s case is that reflexivity is used in a confusing variety of ways, that the virtues ascribed to it are relative to particular commitments, and that treatments of reflexivity as a mark of distinction or a source of methodological advantage do not survive scrutiny. His alternative is an ethnomethodological conception on which reflexivity is an ordinary, unremarkable and unavoidable feature of action, having, as he puts it in summarising the argument, little value as a critical weapon or source of epistemological advantage (Lynch, 2000). He ends by suggesting that this lack of advantage has advantages of its own, for promoting peace and epistemic democracy.

I accept all of it, and the acceptance is not a concession made under duress. Infrastructural position is exactly the ordinary, unavoidable condition Lynch describes: nobody chose it, it confers no distinction, and a researcher who documents it is not thereby a better epistemic agent than one who does not. The calibration model claims no privileged knowledge whatsoever. It claims that a certain kind of document, produced under stated conditions, is more useful to a reader than a certain kind of claim.

That is not an epistemological advantage. It is a distributive one, and distribution is precisely what Lynch’s closing gesture toward epistemic democracy names without operationalising. If reflexivity confers no privilege on its author, then its product should not be addressed to the author’s credibility at all; it should be addressed to what other people can do with it. The completeness model fails Lynch’s test because self-transparency is self-privileging by construction. The calibration model passes it because the halt, the log, and the coded instrument are all things that make the author more exposed rather than more authoritative. Lynch’s critique kills reflexivity as virtue and leaves reflexivity as infrastructure for other people’s checking entirely intact. This article takes up the remainder.

One clarification is owed, because the word has been doing two jobs. In the ethnomethodological sense Lynch inherits from Garfinkel, reflexivity names the constitutive property whereby an account is part of the setting it describes: unavoidable, unremarkable, not a method at all (Garfinkel, 1967). In the sense this article otherwise uses, it names a deliberate practice. Conflating the two is how reflexivity acquires unearned prestige, since the inevitability of the first reads as a warrant for the second. Infrastructural position is a fact in Garfinkel’s sense: nobody chose it, it is constitutive, no one escapes it. Calibration is a practice in the second sense. The relation is not that the fact justifies the practice but that the fact is what the practice has to document, and Lynch’s warning is against treating the fact as a credential.

A second differentiation is more urgent, because the nearest relative to this proposal sits inside the volume already cited. Latour refused the ladder altogether: rather than meta-reflexivity, which adds commentary about the account, he argued for infra-reflexivity, which adds description, more mediators shown, and new literary forms rather than new levels (Latour, 1988). That is right about the direction and incomplete in the one respect that matters here. Infra-reflexivity has no stopping rule. One can always describe more, and nothing says when the describing has done its work or which claims it was meant to protect. Calibration is less a competitor than the missing terminus: the dependency map states which claims rest on which level and hands the rest to a reader, a rule for stopping rather than an invitation to continue. Infra-reflexivity multiplies the account. Calibration constrains its warrant.

With that in hand, return to the regress.

Hegel’s bad infinity names its pattern precisely (Hegel, 2010, Book I, “Infinity”): the endless ought, a finite thing repeatedly transcended toward a beyond it never reaches, each step identical in kind, the series’ incompletability the only thing demonstrated. That is what the objection describes, on one condition: that the series is a series of approximations to a terminus.

So look at the terminus, since it is the premise nobody states. The objection is valid only if the aim of reflexive iteration is full self-transparency. Call this the completeness model. Each layer is an instalment toward a state in which the researcher’s conditions of observation are fully surveyed. Every layer is a partial payment on a provably unpayable debt, so every layer fails and the enterprise is bankrupt. The objection does not wound the completeness model; it destroys it, correctly.

But almost nobody doing reflexive work needs it. It is imported from a picture of objectivity that reflexivity was invented to reject, and its persistence is among the field’s odder inheritances: a tradition founded on refusing the view from nowhere has gone on measuring itself against a standard only that view could satisfy. Withdraw the target and the regress changes character. Hegel’s next move is instructive rather than decorative: the true infinite is not a further term appended to the bad series but the recognition that the movement is self-related, that what looked like a failure to arrive is the thing itself.

The calibration model differs from the completeness model in four respects, and the fourth carries the work.

First, the blind spot is normal rather than faulty. Including it is not maintenance but documentation and accountability: not “I have removed this,” but “this is here, and here is what I did about it.” An instrument’s stated tolerance is not an admission that it is broken. It is what makes its readings usable.

Second, iteration is a spiral rather than a circle. At level B, I use an AI system to code material about AI systems. At B′, I record how I use it, and that blind spot concerns my recording practice and the incidents it is disposed not to register. At B″, I ask what the recording practice presupposes, and that blind spot concerns the theory of relevance under which incidents count as loggable. Different questions, bearing on different claims: the series does not repeat, it diverges. I make no claim that levels decrease in consequence, and the argument does not need one.

Third, the halting criterion is pragmatic rather than logical. No proof licenses a stop; fallibilism licenses it without embarrassment, since no belief is certain, inquiry is unending, and knowledge is usable in the interval (Peirce, 1955). A researcher who cannot act until her conditions of observation are fully known has adopted a standard under which no empirical work has ever been done, including the work by which she came to hold it.

Fourth, the halt is published as a dependency map, not asserted as a finding.

An earlier formulation of this criterion said: iterate until the blind spot at level n+1 could not, if disclosed, alter any claim. That is circular and I abandon it. To know n+1 would not alter a claim, one must run n+1, which is what the criterion licenses not doing. The criterion cannot be knowledge about the unrun level. It must be exposure:

Iterate the reflexive levels until you can state, for each claim the paper makes, which reflexive level that claim depends on. Halt there. Publish the claim-to-level mapping, and mark the halt as a defeasible bet rather than a demonstration.

The author asserts nothing about the unrun level. She asserts something about her own paper: this claim rests on level one, that on level two, this other would be exposed if the level-three assumption about logging relevance were wrong. Each is a statement about work done, checkable in the text.

What the mapping buys is a redistribution of labour, and this is where the criterion stops being a concession and becomes an argument. The author is worst placed to judge whether her unrun level matters: the reason it is unrun is the reason she cannot see it. A reader with a different formation is better placed, and has had no purchase: told that an author “acknowledges her limitations,” there is nothing to grip. The mapping gives her one. She can select a claim, identify the level it rests on, run that level, and report whether it survives. The halt is where inquiry becomes someone else’s to continue, on terms that make continuation possible. This is Lynch’s epistemic democracy given a file format.

Applied here: the claims of this article rest on level two, and Section 5 states its level-three exposure explicitly. The declaration ending the procedure should say what it means and no more:

I cannot see everything, and I tell you so, and I do not pretend that this telling is the end of the chain.

The first clause concedes the structural limit; the second converts the concession into a record; the third blocks the move that would make the record self-defeating, treating the acknowledgement as a final vantage from which limits have been surveyed. That third clause separates calibration from the sophisticated completeness model, in which a confession of limitation functions as the transparency it disclaims.1

One consequence inverts the objection. A paper arguing that there is no position outside the infrastructure would be in performative contradiction presenting its argument as occupying one; the regress is therefore not a threat to the thesis but the thesis appearing inside its own text, a move Luhmann made a matter of theoretical style (Luhmann, 1995). I bound it, since loosely stated it is unfalsifiable. Rough presentation does not evidence honesty, nor polish pretension; displays of humility are cheap. The claim is structural rather than tonal: this article must make no claim whose truth would require an extra-infrastructural vantage. That is checkable, and if a reader finds one the article is in trouble however modest its prose sounds.

5. Tool-as-Case: A Demonstration

This section is not claiming the demonstration settles the status of AI-assisted coding. It claims that it is a procedure, repeatable on other instruments, converting a statement no reader can check into a record every reader can.

The procedure: take the codebook built for the study’s object, and run the study’s own instrument through it as a case.

Stated that way it should sound familiar here, being one move from something the field already does. The algorithm audit evaluates a system against external criteria and produces a score a third party can contest (Brown et al., 2021); the ontology explorer makes a population-management system’s classificatory infrastructure visible by tracing how its categories were built (Van Rossem and Pelizza, 2022). Both point the apparatus at somebody else’s system. Tool-as-case audits the auditor’s instrument with the audit’s own, its only novel requirement being that the criteria be fixed and published before the instrument is coded under them.

A second precedent is closer in logic than topic. Enni and Herrie treat machine learning biases not as defects to eliminate but as hypotheses to investigate (Enni and Herrie, 2021). Calibration performs that conversion on the blind spot: not a fault to repair but a feature to code. The difference is direction: their biases belong to the studied system, mine to the studying one.

The relation to infrastructural inversion should be stated, since a known procedure relabelled is worth nothing. Inversion foregrounds a domain’s normally invisible classificatory and technical arrangements (Bowker and Star, 1999); tool-as-case performs it one turn further in, on the infrastructure of the analysis itself, under criteria already published against the field. The difference is not where attention points but that the criteria are not chosen for the occasion, which keeps the result commensurable with the study’s others.

The audit codebook asks of each case whether data handling operates by an encoded default (C1); whether that default operates independently of user belief (C2); whether it is a verdict rather than a convenience (C3); the status of the reply; whether an exit exists and at what cost; and whether accountable human stewardship is identifiable. Applied to the instrument, the six are legible unmodified, which is the first result: a codebook needing adaptation would have told us the substrate is not in fact shared.

Now the proviso on which the procedure rests. The criteria must be fixed before the instrument is in view. As a claim about the author’s state of mind this is uncheckable and probably false: I designed the codebook knowing I would use AI assistance. I restate it as a publication-order requirement:

The criteria must be published in object-facing form, applied to the object cases, before the instrument is coded under them. Any softening of a criterion to accommodate the instrument then alters the object codings visibly, and the alteration is in the public record.

This is what resists self-serving grading: not that the researcher is disinterested, but that leniency toward her instrument leaves a trace in codings already committed to.

5.1. Coding the Instrument

The instrument was coded on 1 August 2026 against published policy; the table is Supplementary Material 1, with a source register behind every field, so a reader disputing a coding disputes a document, not my reading of it. Four results bear on the argument.

On C1 and C2 the instrument codes as the platform cases do: conversations persist unless deleted, deletion is a positive act, and safety-flagged material is retained irrespective of user belief. On C3 it does not, training use being opt-in and revocable. Exit exists on three documented routes, and its residual cost was unanticipated: material already in completed or running training is unrecoverable. Exit from future use is available; from completed use it is not. Stewardship diverges as predicted but not in the expected direction: designated safety reviewers are documented, a research-facing steward or appeal route is not. Two fields could not be settled from the documents and are reported undetermined, with the document that would settle each.

[DECLARED DEPARTURE] The order stated above was not followed: the instrument was coded before the platform cases. Leniency-traceability therefore lacks both warrants: prior object codings do not yet exist, and no public record establishes that the criteria predate the coding. The codebook is now deposited in object-facing form (OSF, 20 October 2026, doi.org/10.17605/OSF.IO/YQM7D), freezing the criteria prospectively in a public versioned record; but a deposit cannot backdate, and that the criteria were frozen before 1 August 2026 rests on the author’s file record, which a reader cannot inspect. The preceding paragraph remains asserted rather than demonstrated.

5.2. The Comparison

The comparison was completed on 1 August 2026 against three services, one per category, chosen to include a negative case rather than sampled; table and sources are Supplementary Material 3. Two results matter. The prediction was wrong in an instructive direction, the criteria failing outright on the negative case as they should, exit and stewardship diverging neither uniformly nor always against the instrument. The unpredicted result came from the re-creation case, closing: a service whose product was the continued availability of a recorded voice proved shorter-lived than the people it recorded, its exit procedure now an email address. Section 5.1 found exit from future use available and from completed use not; this supplies the complement, that the system built to outlast a person can predecease its provider.

What does the procedure establish? Not that the research is compromised, and not that it is clean. Something more useful: the instrument has a determinate profile on the study’s own criteria, and the profile is on the record. A reader who believes its retention default undermines a finding can say so with reference to a coded field rather than a suspicion, and a reader who disagrees can answer in kind. The disagreement is tractable, which is the most method can offer.

Tool-as-case is thus neither the despair that Section 2’s argument invites nor the promise to try harder that it provokes: not being able to stand outside a thing is compatible with objectifying part of it from inside, which is what the lateral move makes possible.

The demonstration is itself a level of observation with a blind spot, which the model requires me to state rather than solve. The codebook’s categories were built for commercial platforms and may fit the instrument in a way flattering both; Section 3.1 states the level at which the identity holds, and a reader rejecting it rejects this result. The coding was done by a researcher with an interest in the outcome, with no inter-coder check. And the choice of this codebook is stated but not independently audited.

The second has a cheap remedy, and an article insisting on artefacts over attitudes cannot end a limitation with a shrug. Recommended practice: give the research instrument to the study’s second coder as an unlabelled case, coded blind alongside the platforms. This costs one case, removes the interested-party problem, and yields an agreement statistic comparable to the study’s others. It was not done here, and it is the first thing I would change.

Applying the halting criterion: the claims of this section are that the procedure is available, legible, and produces a checkable record, and those rest on level two. The level-three consideration, the theory of relevance governing what counts as a loggable incident, bears on what the profile means for the audit’s substantive findings, and this article makes no such claims. That is the dependency map for Section 5, and a reader who thinks the halt premature has something specific to attack: name a claim above and show it depends on level three after all.

6. Objections

Lynch settled this, and not in your favour.If reflexivity is an ordinary, unavoidable feature of action with no value as a source of epistemological advantage (Lynch, 2000), then your elaborate apparatus is the self-privileging he diagnosed, wearing procedural clothes. You have added machinery to a practice he showed does not need it and cannot support it.

Section 4 accepted Lynch’s argument rather than answering it, and “I agree with him” is not by itself a defence, so here is the test that does the work. Lynch’s target is reflexivity that makes its author more credible. Ask of any reflexive apparatus: does its output raise the author’s standing or lower it? A positionality statement raises it, which is why it is written. A coded instrument profile lowers it, handing a reader determinate fields to attack; a halting declaration lowers it further; a declared departure from one’s own procedure lowers it further still. All three fail the self-privileging test in the direction Lynch would want. As for whether the machinery is needed: a researcher told to keep a log and code her tool, then asked why she stopped at two levels, has no answer except taste. The objection is sharper here than elsewhere, since Section 3’s entries come from preparing this manuscript rather than the audit, and an article on self-observation that observes itself writing is the genre’s worst case. I take the cost: refusing the entries would have left the method undemonstrated, the worse failure.

So AI-mediated research is untrustworthy.  You have described a researcher entangled with her object, unable to verify what her instrument suppressed, dependent on infrastructure she does not control, and subject to omissions correlated across her field. The honest conclusion is that this research cannot be relied on.

This inverts the argument, and is the reading most likely from someone who has read only Section 3. Consider the alternative it implies. A study using the same instrument under the same conditions that reports none of it is not more reliable than one reporting it all. It is identically conditioned and less informative about the conditioning. Disclosure does not create exposure; it makes exposure available for discount.

This also answers the rejectionists on their own terms. I want to read their argument against itself, and the reading is mine; they may reject it. Their stated ground is subjectivity, but notice what “congruent” is doing: congruence is a relation between a practice and its apparatus, and the incongruity they identify is that the apparatus is the wrong kind of thing to occupy a role in the analysis. That is an infrastructural claim in the sense of Section 2, and the sharper version of their position, since it settles nothing about machine subjectivity in order to work. Where I part company is on what follows. Exclusion presupposes the boundary this article denies, and the remedies produce asymmetric things. A non-use declaration is a proposition with no internal structure: believable or doubtable, offering a reader nothing to work on. A coded instrument yields determinate fields, each with a stated evidential basis, each disputable against published documentation. The first asks for trust; the second offers material. Congruence, taken seriously, yields an audit requirement, not a prohibition.

What this does not cover should be stated. Jowsey and colleagues also reject generative AI on social and environmental justice grounds, untouched by anything here. An auditable record of one’s infrastructural position says nothing about the water and energy costs of the inference producing it, or the labour conditions under which the model was made (Crawford, 2021). Calibration answers a methodological problem. A reader who finds the justice objection decisive should not be persuaded by this article, which does not address her.

Your procedures are the problem, not the solution.  Post-qualitative inquiry has spent a decade arguing that the apparatus of method, codebooks, coding, agreement statistics, procedural guarantees, is humanist residue that forecloses thought rather than enabling it (St. Pierre, 2011). You have answered a crisis of method with more method.

I am a proceduralist and will not pretend otherwise, so let me concede what lands and mark what does not. What the critique targets most successfully is method as guarantee: the promise that following the steps secures the finding. Calibration makes no such promise, and its artefacts are built to withhold it. A dependency map states where the author stopped; a coded instrument profile hands over contestable fields. Neither secures anything; both expose. What the critique would still say, rightly, is that a researcher who has decided in advance to produce tables will notice what tables can hold. That is a level-three consideration in the sense of Section 4, not answered here, and Section 5 lists it among the exposures this article declines to resolve.

Distributed-agency accounts already have this, without the apparatus.  Posthumanist and co-researcher models describe knowledge production as distributed across human and non-human actors (Christou, 2026; Costa et al., 2025). They abandoned the bounded observer long ago. Your infrastructural position is their distributed agency with an audit trail bolted on.

They have indeed abandoned the bounded observer, and on that we agree against rejectionists and enthusiasts alike. But distribution is not free, and what it costs is what an audit needs: if agency is distributed such that no node is the observer, no node is answerable, and a reflexive record has no author to be held to it. Suchman’s argument against reading machine behaviour as agency cuts here against the enthusiast rather than against me, since what looks like distributed agency is often better described as a distribution of the conditions under which action is intelligible, a claim about infrastructure and not about who acts (Suchman, 2007). Calibration concedes the descriptive point and refuses the normative inference. Someone still signs, and the signature is not a metaphysical claim about where agency resides but an institutional fact about who can be asked to answer. Distributed agency tells us the observer was never alone. Calibration adds that she is nonetheless the one who signs, and gives her something to sign.

7. Conclusions: Method as Posture

Three concepts have been proposed, in the order they become available.

Infrastructural reflexivity identifies the object: not the researcher’s social location but her position within the technical arrangements that also constitute what she studies. It obtains where substrate sharing and predicate overlap hold together, and differs from participant objectivation in the direction of the move permitted. Objectifying a scholastic disposition is an ascent, available because the disposition is not a case in the study. Objectifying an instrument the codebook can reach is a traverse, both limitation and opportunity.

The calibration model identifies the ambition: auditability rather than transparency. It concedes the regress in full, locates the premise the regress requires, and withdraws it. Its halting criterion asserts nothing about the level not run; it publishes which claims rest on which levels and hands the unrun level to a reader better placed to run it. That transfer is what makes reflexivity survive Lynch’s critique: not a virtue distinguishing its possessor, but an infrastructure for other people’s checking.

Tool-as-case identifies the action: code the instrument under the study’s own codebook, criteria published in object-facing form first. It is the algorithm audit turned around.

The griefbot audit is where I met the condition, not where it lives. Take content moderation. A researcher studying how platforms classify borderline speech, using a language model to label her corpus of moderation decisions, satisfies both conditions exactly: instrument and object are classifiers trained on overlapping web text, and the predicates her codebook applies to the platform (what it treats as borderline, what it silently passes, whose speech it over-flags) are predicates her instrument also instantiates and that cannot be inspected from outside. Tool-as-case is available to her unmodified. She will find, I expect, convergence on exactly the criteria where her findings are most load-bearing, though that expectation is a conjecture rather than a result. Recommender-system and synthetic-media research look like further instances, offered as conjectures too, since a claim of general scope is worth only the cases behind it.

What the three amount to is a change in what method is taken to be. On the older picture, method is a wall: controls holding bias at a distance so findings may pass through uncontaminated. That picture requires an outside, and in AI-mediated research on AI systems there is no outside to have. What replaces the wall is not nothing. It is a posture: occupy the position, record the position, sign the record.

There is a temptation, at the end of an article like this, to present the posture as a solution and thereby reoccupy the vantage the article has spent seven sections denying. I decline it, and the declining is not modesty but the only ending consistent with the argument. What is offered is a procedure with stated limits, a halting criterion that can be shown to have been applied too early, a demonstration whose blind coding was recommended and not performed, a declared departure from its own stated order, and a scope claim resting on one worked case and two conjectures. That is what an auditable method looks like from the inside, and there is no view of it from anywhere else.

The observer cannot leave the loop. She can, at least, say where in it she is standing.