A tone in the silence

On the fourth morning of a ten-day silent retreat in the hills of Kodagu, Sunita Hegde sits with her eyes closed. A small speaker behind her plays a tone so faint that she sometimes hears it and sometimes does not. She is fifty-four, teaches mathematics in Mysuru and has meditated for twenty years. When she hears the tone she presses a key. Afterwards, in a short interview, she says something the researchers had not expected.

"Between the tone and my knowing it there was a gap," she says. "Not long. But in the gap there was nothing, not even me. Then the knowing came, and I was there again."

Two cushions away sits Rohan Pinto, twenty-six, a software tester on his first retreat. On the third block of trials he reports hearing nothing at all. Yet on a forced-choice test after each trial, where he must guess whether a tone was played, he is right three times out of four.

The study is run by Dr Meera Iyer, a cognitive scientist. It also uses an AI assistant called Vani, which transcribes the interviews and summarises them. One evening a colleague, half joking, asks Vani whether it noticed anything when it processed the transcript of Sunita's gap. Vani replies, fluently, that it "registered a brief discontinuity before the content resolved". The sentence is almost the same as Sunita's.

Meera has to write the pre-registered analysis plan before the next cohort arrives. It must say which reports will count as evidence about experience and which will be recorded only as things people, or machines, said. Her philosophy collaborator, Dr Joseph Mathew, has a ready answer. Treat every report as data, he says, and none as authoritative. That is Daniel Dennett's method, heterophenomenology, and it is the standard answer in consciousness science.

Meera, Sunita, Rohan, Joseph, Vani and the retreat are invented. They make an old methodological question concrete and are never evidence for an answer. The question is real, and it is pressing. Every experiment on consciousness depends on reports, and nobody agrees how much to trust them.

This paper asks whether treating reports as data can keep experience in view without giving every introspective claim the last word. My answer is that it can, but only if the authority of a report is graded rather than granted or withheld wholesale. Reports about how things seem now, such as whether a tone was heard, can earn defeasible authority when they track an independent measure and resist suggestion. Reports about mechanism, such as why one noticed the tone, earn little. Reports from a system that has learned to talk about experience earn only as much as a test shows. Dennett's method, I shall argue, is a good start that hides a substantive choice inside a claim of neutrality.

What heterophenomenology asks of us

Dennett introduced the word in his 1991 book Consciousness Explained, and defended it at length in two later papers. The first, Who's on first? Heterophenomenology explained, appeared in the Journal of Consciousness Studies in 2003. The second, "Heterophenomenology reconsidered", appeared in Phenomenology and the Cognitive Sciences in 2007. The name contrasts with autophenomenology, the study of one's own experience from the inside. Heterophenomenology is the study of other people's experience from the outside, using what they say and do as the data.

The method has three steps. First, record what subjects say, and treat the words as sincere speech acts, not as noises. Second, interpret them, as an anthropologist interprets an informant, to build a picture of how things seem to the subject. Dennett calls this picture the subject's heterophenomenological world. Third, and only then, ask what is actually going on in the subject's brain that causes them to say these things. The world described in step two is the explanandum. The brain is where the explanation will be found.

The key move is a kind of bracketing. The researcher does not assume that the subject's world is accurate, and does not assume that it is false. It is treated, Dennett says, as a novelist's world is treated by a literary critic. Sherlock Holmes lived in Baker Street in the stories. The critic need not decide whether there was a real Holmes. In the same way, the researcher records that in Sunita's world there was a gap with nothing in it, without deciding whether there really was.

Dennett's claim is that this method is neutral. It does not prejudge any theory of consciousness. It takes the subject seriously, more seriously than behaviourism did, because it treats reports as reports and not as reflexes. And it does not take the subject too seriously, because it leaves open whether the subject is right.

What the method gets right

Before I criticise the method, it is worth saying plainly what it achieves, because the criticism depends on it. Heterophenomenology does three things well.

It refuses to throw reports away. Behaviourism treated reports as noises, and so had nothing to say about the very thing consciousness science is meant to explain. Dennett treats them as speech acts, with meanings, made sincerely by people trying to say something true. That alone restores experience to the data.

It refuses to treat reports as infallible. Introspective psychology before behaviourism often did, at least for trained observers, and paid the price at Würzburg. Dennett's method builds in the possibility that a sincere report is wrong. That is what makes room for neuroscience to discover something the subject does not know.

And it gives a clear account of the explanandum. What has to be explained is not some private inner glow that only the subject can see. It is the whole pattern of what subjects say, believe and do about their experience. A theory that explains that pattern completely, Dennett argues, has explained everything there is to explain.

I accept the first two achievements without reservation. The third is where the argument of this paper begins. If the explanandum is the pattern of reports, then which reports count, and with what weight, determines what any theory must explain. The method does not leave that open. It settles it, in a particular way, by treating every report of seeming alike.

Authority about seeming

There is one place where the method gives subjects real authority. Dennett allows that subjects are, in a sense, authoritative about how things seem to them. If Sunita says it seemed to her that there was a gap, the heterophenomenologist does not argue. In her heterophenomenological world, there was a gap. What the subject does not have, on Dennett's view, is authority about what is going on inside. If she says the gap was a moment of pure awareness without content, she may be wrong. Neuroscience decides that.

This is the division of labour Joseph has in mind. The subject is the author of the story. The scientist is the one who finds out what the story is a story about. It looks generous to subjects, and in one way it is. It also contains an assumption that does a great deal of work. It assumes that "how things seem" is a single, well-defined category, and that subjects have the same authority over everything in it.

That lesson is the first crack in the neutral picture. If Rohan's report is partly a decision, then it is not straightforwardly a description of his heterophenomenological world. His world in the first block contains "nothing"; in the second, "a faint tone". Nothing about his perception changed. What changed was his policy. Treating both reports as equally authoritative about how things seemed to him misdescribes the case.

A real controversy about introspection

The worry that reports are shaped by more than experience is as old as experimental psychology. Its most famous episode nearly ended the use of introspection in the science.

Neither side could produce a measure that settled the dispute. Each laboratory's observers reported what that laboratory's training led them to expect. To many psychologists the stalemate suggested that introspection was not a scientific method at all. Behaviourism, which arrived soon after, drew that conclusion and banned reports of experience from the data for half a century.

The story matters for Meera because it shows two things. Training changes reports. And without an independent measure, there is no way to tell whether training makes reports more accurate or merely more uniform. Dennett's method was, in part, a way to let reports back in without repeating Würzburg. It treats them as data to be explained, not as observations to be believed.

Neutral about what?

Change one feature of the running example. Joseph reads Meera's draft plan and objects to a single line. She has written that Sunita's reports of hearing a tone will count as evidence about her experience. "That is not neutral," he says. "You have decided to believe her. Heterophenomenology does not believe anyone. It records."

Meera's reply is the argument of this part. Heterophenomenology does believe people, about one thing. It grants them authority about how things seem. And that grant is not neutral, because it decides, ahead of any evidence, which reports go into the explanandum and which do not.

Two ways the grant goes wrong

The first way is that subjects can be wrong about how things seem. Eric Schwitzgebel made this case in 2008 in "The unreliability of naive introspection". His examples are deliberately ordinary. Ask people how wide the region of clear vision is when they fix their eyes on a point. Most say it is fairly wide. Careful testing shows it is tiny, a few degrees across, and that colour fades much sooner than people think. These are reports about how things seem, not about the brain. If people are badly wrong about them, then authority about seeming is not something everyone has by default.

Dennett has a reply. He says that, in such cases, people are not wrong about how it seems. They are wrong about how it is. It seems to them that clear vision is wide, and that seeming is a fact of their heterophenomenological world. The error lies only in taking the seeming at face value.

The reply works, but at a cost. It makes "how it seems" mean whatever the subject says. Then the authority granted is the authority of an author over a fiction. That is complete and empty. It is complete, because nothing the subject says about their world could be wrong. It is empty, because the scientist learns nothing about experience from it that is not already in the words.

The second way the grant goes wrong is the reverse. It gives too little. Some reports about how things seem turn out to be accurate and to carry information nothing else carries. The clearest case comes from work led by Francisco Varela, the founder of what he called neurophenomenology.

On a strict reading of heterophenomenology, the participants' categories were descriptions of their heterophenomenological worlds, to be explained by brain science. In practice, they were measurements. They were used to sort the brain data, and they found structure no one else could have found. They earned authority by doing work. That is not a role the neutral method has a name for.

What the neutral method assumes

Put the two failures together and a pattern appears. Heterophenomenology treats every report of seeming the same way, and treats every report about anything deeper in another, uniform way. Reports of seeming are true in the subject's world. Reports about the mechanism are hypotheses to be checked. There is no grading within each category, and no way for a report to move between them.

Yet the cases so far suggest that what matters is not the category alone. Rohan's report of nothing is a report of seeming, but it is shaped by a criterion. Schwitzgebel's subjects report seeming, and they are wrong. Varela's participants report seeming, and they are right, in a way that shows up in the brain. What distinguishes these cases is not what kind of claim they make. It is whether, for that kind of claim from that kind of reporter, the report tracks something an independent measure can see.

That is the alternative this paper proposes. Call it graded authority. The authority of a report is not granted by its category. It is earned, claim type by claim type and reporter by reporter, by evidence that the report tracks the state it is about. Dennett is right that no report is authoritative by default. He is wrong to make one category authoritative by definition.

Three analogies and their limits

Three analogies help to show what graded authority would be like. Each breaks in an instructive place.

The first is a witness in court. A witness is not believed or disbelieved wholesale. Their testimony about what they saw from the window is weighed by their eyesight, the lighting, their distance and whether their account fits the physical evidence. Their testimony about why the defendant acted carries much less weight. That maps graded authority well. It breaks because a court can visit the scene. The witness and the jury are looking at the same world. With experience, there is no scene to visit. The independent measure is always indirect.

The second is a thermometer. A thermometer is trusted when it has been calibrated against a standard. A new thermometer earns authority by matching the standard across a range of temperatures. That maps the calibration probe in the rubric below. It breaks because there is no standard thermometer for experience. Every independent measure, whether forced choice, brain signal or behaviour, is itself calibrated against reports at some point. The circle can be widened but not escaped.

The third is Dennett's own analogy, the author of a novel. An author has complete authority over what happens in the story. That maps the authority Dennett grants subjects. It breaks in the opposite direction from the others. Authors know they are making things up, and subjects do not. Sunita is not writing fiction. She is trying to describe something, and could fail. An analogy that removes the possibility of failure removes the very thing a science of reports has to measure.

A person who reported for a living

The best long study of how far careful reports can go is a dialogue between a believer and a sceptic.

The book does not settle the dispute, and its authors do not claim that it does. What it shows is that the disagreement can be followed claim by claim. A believer and a sceptic can look at the same moment and ask, of each thing Melanie says, how it could be checked and how it could have been shaped by the questions. That is the habit graded authority asks for. It turns a quarrel about introspection in general into a list of particular claims with particular risks.

Grading the report

Change one feature of the running example. Meera stops arguing with Joseph and designs a test. For each kind of claim her participants make, she will ask three questions of the reports, and score the answers. The result is the instrument of this paper.

The claims fall into kinds. Some are about seeming: was there a tone, was it loud or faint. Some are about structure: did the knowing come before or after the tone, was there a gap between them. Some are about cause: why did you notice that tone and not the last one. A fourth kind, about generality, such as "I always notice tones better in the morning", I leave aside, for reasons that will become clear at the failure boundary.

The rubric is built so that authority must be earned twice. A report that agrees with a measure but does not move when the state is changed may be tracking something else that happens to correlate. A report that moves with the state but also moves with suggestion is partly a reply to the experimenter. Only a report that tracks, responds and resists counts as evidence.

The minimal case

Start with the smallest case that tests anything: one reporter, one kind of claim, two hundred trials. Sunita reports whether she heard the tone. I wrote a short program that simulates the trials, with an independent measure that matches her actual experience nine times in ten.

The simulation needs an assumption about how often her reports follow her experience. I set it at 0.9 for Sunita on claims about seeming, with low suggestibility. The program does not discover that number; it shows what the rubric does with it. On two hundred trials, calibration came out at 0.772, intervention at 0.901 and contamination at 0.005. Authority was 0.768. Verdict: evidence.

Two hundred trials are few, and calibration is the probe that suffers most from small samples. With twenty thousand trials per probe, the same reporter scored 0.888, 0.904 and 0.017, for authority of 0.873. The minimal case shows the rubric working and shows that it needs enough trials to be trusted.

The realistic case

Now the whole study. Three reporters, three kinds of claim, twenty thousand simulated trials per probe. The assumed access values are where the realism lives, and I set them from the literature as best I can. Novices are fairly good at reporting whether they heard something, poor at structure, and close to useless about causes, as Nisbett and Wilson found. Trained meditators are better on all three, and less suggestible. For Vani, I took the figure from Jack Lindsey's 2025 work at Anthropic, where models detected concepts injected into their own activations about one time in five. I gave Vani high suggestibility, since language models are known to follow cues in prompts.

The program's output was as follows. For Sunita, authority was 0.873 on seeming, 0.716 on structure and 0.264 on cause. For Rohan, it was 0.609 on seeming, 0.290 on structure and 0.053 on cause. For Vani, it was 0.079 on seeming, 0.031 on structure and 0.012 on cause.

Read as a plan, those numbers tell Meera what to write. Sunita's reports of seeming and structure are evidence. Her claim that there was a gap between the tone and the knowing counts, provided her structure reports have been calibrated in this way. Rohan's reports of seeming are evidence, barely. His structural claims and everyone's causal claims are data only. Vani's reports are data only across the board.

The striking result is the shape, not any single number. Authority falls as claims move from seeming to cause, for every reporter. That is the graded picture my thesis proposes. Within each kind of claim, it varies by reporter, and training raises it. Heterophenomenology's uniform treatment of each category hides both gradients.

The result is also less impressive than it looks, and I want to say why. The shape is in the assumptions. I set access to fall from seeming to cause, so authority falls too. The program shows that the rubric preserves the gradient and combines the probes sensibly. It does not show that the gradient exists in real people. That is an empirical claim, supported by the studies I drew on and by nothing in my code.

Telling more than we can know

The low authority of causal reports is the best-established finding in this area, and it has a famous source.

In 1977, Richard Nisbett and Timothy Wilson published "Telling more than we can know" in Psychological Review. In one demonstration, shoppers were asked to choose the best of four identical pairs of stockings laid out in a row. They tended to favour the pair on the right. Asked why they chose it, they cited its quality: knit, sheerness, elasticity. None mentioned its position, and when asked, they denied that position had anything to do with it.

The finding has been argued over ever since, and later work suggests that careful interviewing can recover more of the process than Nisbett and Wilson allowed. Claire Petitmengin and colleagues reported in 2013 that participants given expert guidance in describing their choice detected a covert switch in 80 per cent of cases. In the original choice-blindness protocol they reproduced, participants missed the switch in about four cases out of five. But the basic point has stood. People explain their choices by plausible reasons, not by access to the causes. That is why the rubric's covert manipulations matter. Only a cause the reporter does not know about can test a report about causes.

The negative control

A rubric that passed everyone would be useless. The negative control is a reporter whose answers depend only on what the experimenter expects. I added one to the simulation: it has no access to the state at all and maximal suggestibility.

It scored 0.000 on calibration and 0.000 on intervention, with contamination of 0.901, on every kind of claim. Authority was 0. The control matters because its reports, in the expectation condition, can look excellent. When told a tone is likely and a tone is in fact often played, it says "tone" often, and is often right. Only the probes that separate state from suggestion expose it.

A counterexample: the reporter who reads its own behaviour

Here is a case the rubric gets wrong, or at least gets right for the wrong reason. Suppose a reporter has no inner access to the tone at all, but watches its own key presses and infers from them. When it notices that it pressed the key, it says it heard a tone. Its key presses track the tone fairly well, because they are driven by its perceptual system.

I simulated this reporter. Calibration came out at 0.708, intervention at 0.698 and contamination at 0.000. Authority was 0.698. Verdict: evidence. The rubric certifies it.

Is that a mistake? It depends on what authority is for. If authority means that the report is a good guide to the state, the verdict is right. The reporter's reports are a good guide. If authority means that the reporter has introspective access, the verdict is wrong. It is inferring, not introspecting. Nisbett and Wilson suggested that much human self-report works this way.

I accept the verdict and narrow the claim. The rubric measures whether a report is a reliable guide to a state, not the route by which the report was produced. For Meera's purposes, reliability is what matters: she needs to know which reports can be used as measures. For a theory of introspection, the route matters too, and the rubric cannot see it.

The sweep in the chart shows how the two inputs trade off. With low suggestibility, authority rose from 0.445 at an access of 0.5 to 0.716 at 0.75. With high suggestibility, the same steps gave 0.308 and 0.612. A suggestible reporter needs noticeably better access before its reports count as evidence. At the extremes, suggestion hardly matters: a reporter with no access scores zero, and one with perfect access scores one, whatever the experimenter says.

Running the rubric on a real model

Vani is invented, but the probes it would face are not. Three lines of recent research supply them. The first is concept injection. Lindsey and colleagues placed a known concept directly into a model's internal activations and asked whether the model noticed anything unusual. This is the intervention probe in its purest form. The state is changed directly, and the report is checked against the change. The models sometimes noticed, and named the injected concept, but far from always.

The second is self-prediction. In 2024, Felix Binder and colleagues reported in "Looking inward" that models trained to predict their own behaviour did so better than other models trained on the same behaviour. That is a form of calibration: the model's reports about itself agree with an independent record of what it does.

The third is unfaithful explanation. In 2023, Miles Turpin and colleagues showed that language models' step-by-step explanations often leave out the real influence on their answers, such as a hint planted in the prompt. That is the contamination probe, and Nisbett and Wilson's stockings in a new medium. The model cites reasons; the planted cue was the cause.

Put together, these findings suggest the same gradient for models as for people. A model's reports of whether something is present in its processing can be calibrated, a little. Its reports of why it answered as it did are often unfaithful. The rubric would give a real model low authority across the board today. It would also say exactly which experiments could raise it.

The failure boundary

The rubric needs an independent measure of the state a claim is about. For tones, there are many. For some of the claims contemplatives care about most, there are none.

Sunita's teacher, asked about the gap, says that in long practice it can widen into a state with no content at all: awareness without any object, sometimes called a pure consciousness event. Reports of such states are central to several contemplative traditions and to RTM. They are also, by their nature, reports of the absence of anything an independent measure could track. There is no tone to mask. A brain signal could correlate with the reports, but it would have been identified by the reports in the first place.

General claims fail for a related reason. "I always notice tones better in the morning" could be checked, but not by any single trial, and the reporter's own memory is the main source of the claim. Where there is no independent measure, the rubric does not lower authority; it cannot assign any. Such reports stay in the dataset as data only, not because they are false, but because nothing can yet test them.

Witnesses and the trained eye

Change one feature of the running example. Meera shows the rubric to the retreat's teacher, Gopal Rao, who has taught at the centre for thirty years. He reads it slowly and says that it reminds him of an old argument about when to believe what someone tells you. Indian philosophy, he points out, treated testimony as a source of knowledge in its own right, and spent a long time on the question of whose testimony counts.

He is right, and the parallel is closer than it first appears.

[AUTHOR NOTE: how a teacher in your own practice judged a student's report, and on what basis.]

The trustworthy speaker

The Nyāya school, one of the classical schools of Indian philosophy, recognised four means of valid knowledge, called pramāṇa: perception, inference, comparison and testimony. Its founding text, the Nyāyasūtra attributed to Gautama, defines testimony in a single short sūtra, numbered 1.1.7 in the standard arrangement: āptopadeśaḥ śabdaḥ, "testimony is the instruction of a trustworthy person". The word for trustworthy is āpta.

The commentator Vātsyāyana explained what makes a speaker āpta. In the translation at the same source, the trustworthy person is one who has direct and right knowledge of the thing, who wishes to make it known as they know it, and who is capable of saying it. Those three conditions are knowledge, sincerity and competence of expression.

The parallel with the rubric is plain once it is set out. Calibration asks whether the reporter knows: whether their reports track the state. Contamination asks whether the reporter is moved only by the wish to say what they know, or also by the wish to please. And the failure of reports such as Rohan's "nothing" is, in part, a failure of the third condition: the ability to put what one knows into a report without distortion. Nyāya did not treat testimony as authoritative by category. It treated it as authoritative when the speaker met the conditions.

There is a difference, and it matters. Nyāya's āpta testifies about the world, or about scripture. Introspective reports testify about the speaker. Nyāya's tests can check a witness's claims against the world. The rubric must check a reporter's claims against an independent measure of their own state, which is harder.

Self-awareness in Buddhist philosophy

A second Indian debate bears on authority about seeming. The Buddhist logician Dignāga held that every cognition is aware of itself, a doctrine called svasaṃvedana, reflexive awareness. On that view, when one sees blue, the seeing is also, in the same act, aware of itself as seeing. Other schools denied this, and held that a cognition is known only by a later cognition that takes it as an object.

If Dignāga is right, there is a kind of self-knowledge built into every experience, and reports of how things seem could draw on it directly. If his critics are right, every report of experience is a second act, about the first, and can get it wrong like any other judgement. The debate maps onto the dispute between Dennett and his critics with surprising precision. The rubric does not settle it, but it shows what would bear on it. Reflexive awareness predicts high authority for reports of current seeming, from everyone. The rival view predicts authority that varies with skill.

The Pratyabhijñā school, from which RTM takes its terms, sided with self-awareness in a strong form. For Utpaladeva and Abhinavagupta, consciousness is light, prakāśa, and also light that knows itself as light, vimarśa. Every appearance is grasped from within by the awareness in which it appears. On that view, the capacity to report is not a later act laid over experience. It is an expression of a reflexivity present in experience from the start.

That view makes a prediction the rubric can check, and it is a risky one. If reflexivity is present in every experience, then reports of current seeming should be well calibrated for everyone, trained or not, once criterion effects are removed. If training is needed to make them reliable, then reflexivity is at least partly a skill, and the strong view needs qualifying. The simulation cannot tell us which. The breath-counting results, which separated meditators from controls, lean toward skill.

Lutz and Varela's study and the breath-counting work point to the same conclusion. Training does not merely change reports, as Titchener's critics feared. It can make them more accurate, and that can be shown. The Würzburg stalemate happened because neither laboratory had an independent measure. Where one exists, training can be tested rather than trusted.

When the scan disagrees

The next case is the one I least like, because it is where graded authority could do harm.

What should Lakshmi's clinic do with the scan? Graded authority gives a definite answer. The scan may inform the doctor's judgement in the domain where it was validated, such as distinguishing painful heat from warmth in a research setting. It may not be used to lower the weight of her report of chronic back pain, because it was never calibrated against such reports. If the clinic wants a measure for chronic pain, it must build one, and it must build it on reports like hers. Until then, her report is the best evidence there is, and it should be treated as such.

There is a wider lesson in Lakshmi's case for heterophenomenology itself. The neutral method treats her report as a datum about her heterophenomenological world, and treats the scan as a third-person measure of her brain. It does not say what to do when they conflict. In a laboratory, that silence is harmless, because nothing hangs on the verdict. In a clinic, it is not harmless, because silence leaves the choice to whoever holds the more impressive instrument. A method for consciousness science that will be used in medicine needs a rule for conflicts. Graded authority supplies one: the instrument must earn its authority over reports on the same ground the reports must earn theirs, by calibration in the domain at issue.

The reverse case is just as instructive. In 2006, Adrian Owen and colleagues published a short paper in Science, "Detecting awareness in the vegetative state". A young woman who met the criteria for a vegetative state, after a road accident, was asked in a scanner to imagine playing tennis or walking through her house. Her brain activity matched that of healthy volunteers doing the same tasks. She could not speak or move, but she could, it seemed, follow instructions.

Here the measure is the report. It is a report made in brain activity rather than words, and it was calibrated against healthy volunteers who also reported in words. Graded authority handles both cases the same way. Authority flows along chains of calibration. It runs from reports to the measures built on them, and back to reports from people who cannot speak. It never flows from a measure to a report the measure was never calibrated against.

Reports and the layers of RTM

This series tests a hypothesis about consciousness, the Recognitive Triadic Model, which I treat as a claim under test and not a doctrine. RTM borrows its vocabulary from the Pratyabhijñā school of Kashmir Śaivism and describes five layers: L1, spanda, responsiveness; L2, prakāśa, manifestation; L3, vimarśa, reflexive grasp; L4, pratyabhijñā, recognition; and L5, cit, consciousness as ground. Two layers bear directly on reports. Vimarśa is the layer at which a state is grasped as what it is, and that grasp is what any report of experience draws on. Pratyabhijñā is the recognition, central to the school, that one's own awareness is consciousness itself.

RTM depends on reports more than most theories, and on reports of an unusual kind. Its evidence for L4 is largely the testimony of practitioners who say that they have recognised something. So the question of this paper is not a side issue for RTM. It is the question of whether its central evidence counts.

Change one feature of the running example. On the last day, Sunita tells Meera something she has not said before. "In the gap," she says, "I recognised that the knowing was what I am. Not that I have it. That I am it."

What the rubric says about L3 and L4

For L3, vimarśa, the news is mixed but hopeful. Reports of seeming and structure are the reports vimarśa would produce. They can earn authority, and training can raise it, as the breath-counting and neurophenomenology work show. If RTM is right that vimarśa is a distinct layer, it should be possible to find reports whose authority depends on it: reports that are better calibrated after training in exactly the kind of attention the tradition describes. That is a testable prediction, and the existing literature is at least consistent with it.

For L4, pratyabhijñā, the news is harder. Recognition reports sit at the failure boundary. There is no independent measure of recognising one's awareness as consciousness, and no way to intervene on it directly. They are also exposed to contamination by doctrine, which is the rival's point. Until a cross-tradition study of the kind just described is run, RTM's L4 reports are data only, by the standard of this paper.

My verdict is that RTM is refined. It must treat its own central evidence the way it would treat anyone else's. L3 can earn authority by calibration. L4 must first survive a contamination test against doctrine, and until it does, it must be stated as a hypothesis about what practitioners say, not about what they find.

What a test of L3 would look like

A study that could put RTM's L3 at risk is not hard to describe. Take novices and long-term meditators. Give them a task like Meera's, with tones masked on some trials and expectations manipulated on others. Measure the authority of their reports of seeming and of structure, using the rubric. Then train the novices for some weeks in a practice that targets reflexive attention, such as the breath count, and a control group in something else, such as a memory task.

RTM predicts that authority for structure reports will rise more in the reflexive-attention group than in the control group. It also predicts that the rise will be in calibration and intervention, not merely in lower suggestibility. A rise that came only from reduced suggestion would show that training made people steadier reporters, not that it gave them better access. That distinction is the one the rubric was built to make.

The strongest case against this paper

Here is the best objection I can build, in the voice of a careful defender of Dennett.

"The author has not left heterophenomenology. He has carried it out. Every step in his rubric treats reports as data and asks what they correlate with. Calibration against forced choice, brain signals and breath sensors is exactly the third-person science Dennett recommended. The author's graded authority is not a rival to the method. It is what the method yields when you do the work. Dennett never said all reports of seeming were equally reliable. He said that, before the work, we should not assume any of them are.

"His case against neutrality also fails. He says the method grants subjects authority over how things seem, and that this grant is either empty or prejudicial. But the grant is a definition of the explanandum, not a claim about accuracy. It says what is to be explained: the fact that it seems to Sunita there was a gap. That fact is beyond dispute; she said it sincerely. Whether there was a gap is the open question. Nothing prejudicial has happened.

"Finally, look at what his rubric ends up endorsing. The reporter who infers from its own key presses gets high authority. A report produced by reading one's own behaviour is exactly what Dennett said introspection often is. The author's best instrument cannot tell introspection from self-observation, which is Dennett's thesis about introspection."

I accept much of this. The rubric is third-person science, and Dennett would endorse most of its steps. I accept that the inferring reporter is a real problem, and I have narrowed my claim because of it. I do not accept the conclusion that nothing has changed. Two things have.

First, the method I propose lets reports move between categories by test. Reports of seeming can lose authority, and reports of structure, which the neutral method treats as hypotheses, can gain it. That changes what goes into the explanandum. If Sunita's structure reports are calibrated, then the gap is not merely a fact about what she said. It is a measured feature of her experience, and theories must explain it. Second, the neutral method gives no guidance about pain. Graded authority does: measures cannot overrule reports outside the domain on which they were calibrated. That is a substantive commitment, and it is not in Dennett.

There is a third difference, and it matters most for RTM. Dennett's method has no place for a practice that changes the reporter. It takes reporters as they come and studies what they say. Graded authority asks how a reporter came to be reliable, and allows that some trainings make people better instruments. That opens a route for contemplative traditions into the science that the neutral method closes. On the neutral method, a meditator's report is one more text to be interpreted. On the graded method, it may be a reading from a calibrated instrument. Whether any meditator's reports earn that status is an empirical question. The point is that the question can now be asked.

What the objection does show is that my disagreement with Dennett is narrower than the title of this paper suggests. We agree on the data, on the need for independent measures, and on the fallibility of sincere reports. We disagree on whether the explanandum should be fixed before the measurement or partly by it. I think that disagreement is real. I also think it is the kind that experiments can settle, which is the best thing one can say about a philosophical dispute.

What would change my mind

I would give up the central claim, that report authority should be graded by claim type and reporter and earned by test, under any of the following.

First, evidence that report authority does not vary by claim type in real people. If careful studies found that reports about causes were as well calibrated as reports about seeming, once the probes were run properly, the gradient I propose would be an artefact of my assumptions. This bears on RTM's L3, which predicts the gradient.

Second, evidence that training does not raise calibration but only uniformity, as Titchener's critics feared. If trained meditators' reports agreed more with each other but no better with independent measures, then the Würzburg worry would return in full and graded authority would reward conformity.

Third, a cross-tradition study in which reports of recognition separated cleanly by doctrine. That would leave RTM's L4 evidence as data only, perhaps permanently, and I would say so.

Fourth, a demonstration that the inferring reporter is the normal case. If human reports of seeming turn out, in general, to be inferred from behaviour, then the rubric's verdicts remain useful for measurement but say nothing about introspection. My thesis would be true but much smaller.

Each RTM layer is exposed differently. L1, spanda, is barely touched: nothing in this paper's evidence bears on whether experience has a pulsing, dynamic ground, and I would not change my view of L1 on any result described here. L2, prakāśa, is touched only indirectly. If reports of bare presence could never be calibrated, that would leave L2 without any first-person support, though it would not refute it. L3, vimarśa, is directly exposed by the gradient and by the training study I described. L4, pratyabhijñā, is exposed by the cross-tradition test. L5, cit, lies beyond anything a report rubric could reach, and I say so without embarrassment. A method that claimed to grade reports of the ground of all experience would be claiming too much.

Finally, I would change my mind about the practical proposal for pain if graded authority produced worse outcomes for patients. If clinics that weighted reports by tested reliability ended up dismissing more real pain, the rubric would have failed in the place where it most needed to succeed. A test of a philosophy of reports should include the people whose reports are at stake.

Two dated predictions put the argument at risk.

• By the end of 2028, at least one published study of language-model self-reports will score calibration separately for different kinds of claim about internal states, using interventions on activations, and will find higher calibration for presence claims than for causal claims. If none does, either the gradient fails for models or nobody has looked.
• By the end of 2028, no major consciousness-science collaboration will have published agreed criteria for weighting introspective reports by claim type. If one does, the field will have moved faster than I expect, and toward the view argued here.

Mahapatra, R. R. Grade the report: testing Dennett's heterophenomenology against trained meditators and a talking model. Mind Frontier. DOI to be assigned. https://rajeshrmahapatra.com/papers/mind-frontier/u1029-grade-the-report.html

@article{mahapatra-mf48,
  author = {Mahapatra, Rajesh Ranjan},
  title = {Grade the report: testing Dennett's heterophenomenology against trained meditators and a talking model},
  journal = {Mind Frontier},
  doi = {DOI to be assigned},
  url = {https://rajeshrmahapatra.com/papers/mind-frontier/u1029-grade-the-report.html}
}