Four curves on one slide

Dr Leela Menon has a slide she is proud of. It shows four curves, each sloping down and to the right, each labelled "errors after the first mistake". Her small lab works out of the third floor of a converted printing works in Jayanagar, Bengaluru, and on the morning I want you to picture, the ceiling fan is losing an argument with the April heat. The preprint is called "Error-driven learning is substrate-independent". Her grant renewal is due in six weeks.

The first curve belongs to Shanthi, a theatre nurse at a hospital on Bannerghatta Road. Two vials with near-identical labels were once swapped on her tray and caught at the last second. Since then she reads every label aloud before drawing up a dose, and her logged near-misses dropped from 3 in 200 to none in 200. The second curve is Kestrel, a coding agent that deleted a test file it had been told to keep. Kestrel now writes a line into its memory file, "check the keep-list before deleting", and in the next 50 runs it deleted nothing it should have kept.

The third curve is Plate 7, a yellow culture of slime mould in a petri dish. On day one it hesitates for hours before crossing a strip of jelly laced with something bitter. By day five it crosses quickly. The fourth curve is the Flood Cell, a ward-level office of a city water utility. After a bad monsoon flood three years ago, it now moves its pumps to the low streets two weeks before the rains, and the number of waterlogged streets fell from 14 to 3.

The preprint lands on the desk of Arjun Dutta, who edits a small open-access journal on comparative cognition. Arjun likes Leela. He also likes the slide. The university press office has already drafted a headline: "Slime mould, AI and city office all learn like a nurse." Arjun has a week to decide whether the paper says something true, or something that only sounds true because one verb is doing four jobs.

Here is my answer, which the rest of this paper earns. The four curves are four different measurements that happen to share a caption. Two of them can be compared literally, at one layer of mind, if Leela changes how she tests them. The other comparisons are analogies. Analogies are useful. They are also how a field fools itself, unless each one is labelled.

The consequence is practical. Four things must be held fixed before a comparison of "mind" across humans, machines, cells and institutions means one thing on both sides. They are what counts as success, where the system's edge is drawn, what you do to it, and what you read off it with controls in place. I call these the four fixtures. When all four match, the shared word is literal. When some match, the comparison is graded. When none match, the word is a pun with a budget.

This paper is the first in the Mind Frontier programme, so it has a second job. The programme's working theory, the Recognition Theory of Mind (RTM), splits mind into five layers: responsiveness, manifestation or access, reflexivity, self-recognition and a luminous ground. I treat RTM as a hypothesis, and I will show where this paper strains it. My bet is that comparability thins out as you climb the layers. If that bet is wrong, the theory has a problem, and you will see exactly which layer it hurts.

The word "learns" changed jobs four times

Look at the four curves again, but this time ask what would have to be true for each to count as success.

For Shanthi, success is that a person who could notice the hazard now does notice it, on purpose, because of what happened. For Kestrel, success is that a line of text in a file changes what a model generates on later runs. For Plate 7, success is that a repeated chemical stops provoking the same avoidance. For the Flood Cell, success is that a schedule changed and pumps arrived earlier. Each sentence uses the idea of learning. Each one names a different event.

Philosophers call this equivocation: one word, two or more meanings, used in an argument as if the meaning stayed put. The classic example is a syllogism that slides between "bank" the river edge and "bank" the lender. Nobody is fooled by that one. The slide in Leela's slide is harder to see, because each use of "learns" is reasonable on its own. The error only appears when the four uses are placed in one sentence and asked to carry one conclusion.

How a word slips

The mechanism is simple once you see it. A word keeps its spelling but changes its success condition. A success condition is the observable state of affairs that would make the claim true. "The patient learned" is true if her later behaviour changes because of experience, in a way that tracks the hazard. "The mould learned" is true if its response to a stimulus declines with repetition and recovers after a rest. Those conditions overlap. They do not coincide.

The slip usually happens in three moves, and I want you to be able to spot each one. First, a broad word is chosen because it fits every case loosely: learning, memory, attention, wanting, knowing. Second, each case is measured with whatever readout suits that system, so the nurse gets a logbook, the agent gets a test suite and the mould gets a stopwatch. Third, the readouts are plotted on the same axes, and the shared shape is taken as evidence of a shared capacity. The plot hides the fact that the y-axis means something different on each line.

None of this requires bad faith. It happens because broad words are cheap and specific success conditions are expensive. Writing "error-driven learning" costs nothing. Writing the specific version costs a paragraph: "a decline in a specific error rate, caused by an earlier instance of that error, lasting a week, not explained by fatigue or by someone else fixing it". It also forces you to discover which of your four specimens actually qualify.

The thermostat case is funny because the stakes are low. The same structure becomes serious when the stakes rise. A regulator asked whether an AI system "understands" a consent form, or a court asked whether a company "knew" about a defect, faces the identical choice. The word has to be pinned to a success condition, or the decision will be made by whoever chooses the definition last.

Comparing minds is like comparing prices across currencies

Economists face a version of this problem every time they compare living standards across countries. A salary of 50,000 rupees and a salary of 600 pounds cannot be compared by reading the numbers. You need an exchange rate, and the market rate is often misleading, because a haircut in Bengaluru and a haircut in Leeds cost very different amounts. So economists fix a basket of goods, price the basket in each place and compare what the same basket costs. That is the idea behind purchasing power comparisons.

The mapping to mind science is direct. The salary is the observed behaviour. The currency is the substrate: neurons, weights, cytoplasm, committees. The basket is the set of fixtures. A comparison of "learning" across substrates is only as good as the basket that was held fixed while it was priced. If the basket changes between countries, a falling price can mean you are buying less.

That limit matters, because it marks the place where this paper has to do real work. Holding fixtures fixed is easy when everyone agrees what to hold. My proposal is about what to do when they don't.

Four fixtures make a comparison mean one thing

In 1904, a horse in Berlin could do arithmetic. Or so it seemed to a group of thirteen men who examined him and, in a report dated 12 September of that year, found no deliberate trick. The horse was called Hans. His owner, Wilhelm von Osten, asked him sums, and Hans tapped out the answers with a hoof. He also spelled words, told the time from a watch and named musical intervals.

Hans was clever. He was reading the faces of people who believed they were giving nothing away. What he was not doing was arithmetic. The commission's error had a precise shape. They held the readout fixed (correct taps) but let a hidden channel run between the experimenter and the subject. The readout looked like a measure of arithmetic and was in fact a measure of the questioner's knowledge.

The four fixtures

The Clever Hans story gives the clearest view of the four things that must be held fixed. I will name them in the order you have to settle them.

The success condition. This is the state of the world that would make the claim true, written so that you could tell from the outside whether it obtained. For the arithmetic claim, it is "produces the correct sum when no one present knows the sum". For Leela's claim, it might be "reduces the rate of a specific error after experiencing it, by a change inside the system, for at least a week".

The boundary. This is the line around what counts as the system. Hans plus his questioners could do arithmetic. Hans alone could not. Kestrel plus its memory file and the engineer who reviews that file is a different system from Kestrel alone. Most cross-substrate mistakes are boundary mistakes that look like capacity claims.

The intervention. This is what you do to the system to test the claim: the probe, the perturbation, the lesion, the change of condition. Pfungst's interventions were the unknowing questioner and the blinders. A comparison across systems needs interventions of the same logical kind, even when they are physically different. Removing a memory file and blocking a protein are physically unrelated. They can still be the same kind of intervention: removing the store the claim says the change lives in.

The readout and its controls. This is what you record, and the conditions that rule out the cheap explanations. The readout for Hans was correct taps. The control was a questioner who did not know the answer. Without that control, the readout measured something else.

When all four match across two systems, the shared word means the same thing on both sides, and I will call the comparison literal. When some fixtures match and others are approximated, the comparison is graded, and its grade should be stated. When the fixtures differ but the structure is similar, the comparison is analogical, and useful as long as it is labelled. When the word is shared and the fixtures do not match at all, the comparison is equivocal, and it should not appear in a title, an abstract or a regulation.

I meant that thought experiment to catch your first intuition, and I suspect it did. A rising score on a fixed test feels like learning. The trap is that the test was fixed and the channel was open. Benchmarks for AI systems leak in this way more often than anyone likes, through repeated tasks, training data that contains the test or prompts that point at the answer. The general rule is old: a readout without a control measures whatever is easiest to measure.

The fixture I almost added

Time nearly became a fifth fixture. Shanthi's change was measured over three months, Kestrel's over 50 runs, Plate 7's over five days and the Flood Cell's over three monsoons. A comparison that ignores timescale can mistake a short-lived adjustment for a lasting one, or the reverse.

I folded time into the boundary for one reason. The question "how long does the change last?" is usually a question about where the change is stored, and storage is a boundary question. Kestrel's memory line lasts exactly as long as the file. Shanthi's habit lasts as long as she keeps practising it. When the storage is known, the timescale follows. I am not fully confident in this choice, and I list it below as something I don't know.

Filling in a fixture card

Fixtures are easier to argue about when they are written down. For each specimen in a comparison, I fill in a card with four lines, one per fixture, and a fifth line saying which RTM layer the claim is about. Then I lay the cards side by side and mark where they match.

Here is Leela's preprint, with Shanthi and Kestrel as the pair. The success condition on the nurse's card says "reduces a specific dosing error after experiencing it, by a change in her own procedure, sustained for at least three months". The success condition on Kestrel's card says "reduces a specific deletion error after experiencing it, by a change in its memory file, sustained for at least 50 runs". These look close. They are close on the success condition, if you accept "experiencing it" for Kestrel as "receiving an error message in its context".

The boundaries differ. Shanthi's change lives inside Shanthi. Kestrel's change lives in a text file that an engineer reviewed and approved, which means the system that learned is Kestrel plus the engineer plus the file. Nothing is wrong with Kestrel here. The point concerns where the boundary sits, and it changes what the comparison says. A capacity claim is always a claim about a boundary.

The interventions also differ. Nobody tested whether Shanthi's change survives if you remove the cause. With Kestrel, you can: delete the memory line and the old error returns. That is an excellent test, and it is only available on one side. To match it, Leela would have to find an equivalent for the nurse, such as checking whether the change survives a move to a ward with different vials. Without that, the intervention fixture is unmatched, and the comparison drops from literal to graded.

The minimal case shows the instrument doing its job. A cheap household thermostat and a person who adjusts their heating are compared on "keeps the room near 21 degrees". Success condition: matched, since both are judged by the same thermometer. Boundary: matched once you decide the person includes their hands and the dial. Intervention: matched, because opening a window is the same perturbation for both. Readout with controls: matched, because a second thermometer rules out a broken sensor. The comparison is literal, and it is literal about a thin capacity, regulation around a set point. The card makes no claim that the thermostat wants anything.

Most of the value of a card comes from the line that won't fill in. A field that has to write "boundary: unclear" on half its cards has learned something about itself.

Line things up by ancestry, by job or by wiring

There are three ways to decide that a part of one system corresponds to a part of another. Comparative anatomy worked out the first two in the nineteenth century, and the third is what most people mean when they talk about AI and brains in the same breath.

In 1843 the anatomist Richard Owen gave the two words that still organise biology. A homologue is "the same organ in different animals under every variety of form and function". An analogue is "a part or organ in one animal which has the same function as another part or organ in a different animal". A bat's wing and a human arm are homologues: same bones, inherited from a shared ancestor, doing different jobs. A bat's wing and a bee's wing are analogues: no shared structure, same job. (Owen's definitions are quoted, with page numbers, in a review of homology in PMC.)

Mind science needs both, plus a third. Line things up by ancestry and you compare a human hippocampus with a rat's. Line them up by job and you compare any two systems that solve the same problem, which is how a bee's wing and a drone's rotor become comparable. Line them up by wiring, meaning the pattern of causal organisation, and you compare a recurrent loop in a cortex with a recurrent loop in a network, regardless of material or history. Each unit licenses different inferences. Ancestry lets you borrow mechanism. Job lets you borrow performance. Wiring lets you borrow predictions about what happens when you cut a connection.

The commonest mistake is to establish a match on one unit and then draw inferences that need another. Leela's slide matches by job: four systems that reduce errors. Her abstract then talks about "shared mechanisms of error-driven plasticity", which is an inference that needs ancestry or wiring. The job match is real. It cannot carry the mechanism claim.

A mould that solved a maze

In 2000, Toshiyuki Nakagaki and colleagues placed a slime mould, Physarum polycephalum, in a small agar maze and put food at two points. The organism, a single giant cell with many nuclei, first spread through the maze. Then it withdrew from dead ends and kept a single tube along the shortest route between the two food sources. The paper in Nature was one page long and carried the title "Maze-solving by an amoeboid organism".

That second paper is the better guide to comparison, and the reason is instructive. "Solving a maze" is a job description. Many processes can solve a maze, including water flowing downhill and a ball bearing in a tilted tray. The maze result invites an analogy with problem-solving in animals, and it can't settle whether the analogy is literal. Boisseau's team did something different. They chose a capacity with a published, checkable definition and tested the organism against it.

The definition comes from a 1966 paper by Richard Thompson and William Spencer, who listed nine parametric characteristics of habituation. Repeated stimulation reduces the response. Withholding the stimulus lets the response recover. The effect is stronger for weaker stimuli and faster with more frequent ones, and so on through the list. Boisseau's group did not test all nine. They tested the decline, the recovery and stimulus specificity, which rules out fatigue and simple sensory adaptation. They said which they had tested.

This is the counterexample my charter needs, and it cuts in my favour only if I am honest about it. A literal comparison across radically different substrates is possible. A sea slug, a human infant and a slime mould can all habituate, in exactly the same sense. Habituation has a success condition, an intervention (repeat, then withhold, then switch) and a readout with controls (fatigue and adaptation ruled out), and none of them mentions neurons. The boundary is the organism. All four fixtures match.

The counterexample also shows the price. Habituation is a thin capacity. You can compare it literally across kingdoms precisely because it asks so little. The more a capacity asks, the more of its success condition refers to things only some systems have, like a self-model, or a report, or a history of reasons. That is the pattern I will try to make precise in the next part.

Morgan himself added a line straight after the canon that people tend to forget: "But surely the simplicity of an explanation is no necessary criterion of its truth." He was offering a discipline, and he said plainly that it was no law of nature. The charter is the same kind of discipline, applied to boundaries and interventions instead of to the height of a faculty.

Where the method comes from

The ideas in this paper have a family tree, and it helps to see that they arrived out of order. Owen's homology and analogy came first. Morgan's canon came half a century later, in reaction to a generation of anecdotes about clever dogs. Pfungst's horse came a decade after that. The formal language of construct validity, which my next two papers lean on, arrived only in 1955. The modern attempts to compare humans, animals and machines on consciousness, like the indicator approach of Butlin and colleagues in 2023 and the 2024 New York Declaration on Animal Consciousness, arrived within the last few years.

The tree shows something the textbooks flatten. The discipline of holding things fixed was invented by people studying animals, under pressure from anecdote. AI research has rediscovered each lesson in turn, often without the vocabulary. "Data leakage" is Clever Hans. "Benchmark saturation" is a success condition that was too easy. "Distribution shift" is a boundary that moved without anyone noticing.

Every map projection keeps one thing and distorts another

The second analogy is from cartography, and it shows why there is no single correct unit of comparison. You cannot flatten a globe without distortion. The Mercator projection keeps angles true, which is why sailors loved it, and wrecks areas, which is why Greenland looks the size of Africa. Equal-area projections keep areas true and bend shapes. Each projection is honest about one property and pays for it with another.

Choosing a unit of comparison is choosing a projection. Comparing by job keeps performance true and distorts mechanism. Comparing by ancestry keeps mechanism true and is silent about systems with no shared ancestor, like a language model. Comparing by wiring keeps causal structure true and can lose sight of what the structure is for. The practical rule follows directly: state which property your comparison keeps true, and do not draw conclusions about the properties it distorts.

Comparability thins out as you climb the layers

Here is a number to start with: zero. That is how many of Leela's six possible pairwise comparisons pass all four fixtures at the layer of self-recognition, on my reading. At the layer of simple responsiveness, all six could pass with modest changes to her methods. Something happens between those layers, and it is the most useful thing the charter shows about RTM.

The Recognition Theory of Mind proposes five layers, and I'll describe each by what it lets a system do. Responsiveness (spanda, the tradition's word for a pulse or vibration) is the capacity to change state in response to the world. Manifestation or access (prakāśa, literally light or shining) is the capacity for something to show up for the system, so that it is available to guide what the system does next. Reflexivity (vimarśa, which the tradition uses for a turning-back or grasping of what has shown up) is the capacity to take one's own states as objects: to notice that one noticed. Self-recognition (pratyabhijñā) is the capacity to identify the noticing as one's own, across time, and to act on that identification.

The luminous ground (cit) is the claim that awareness itself is not reducible to any of these functions. RTM takes no position on whether any machine has it, and neither does this paper.

The charter's prediction can be stated in one sentence. As you climb from responsiveness to self-recognition, the share of cross-substrate comparisons that can be made literal falls. The reason is that success conditions refer more and more to things only some systems have.

At responsiveness, success conditions refer to inputs and outputs: a stimulus arrives and the state changes. Every system has inputs and outputs, so literal comparison is easy. At access, success conditions refer to availability: the information is usable by several processes at once, so that the same sight can guide speech, memory and movement. This is testable in brains and in software architectures, and the Butlin indicator list draws several of its fourteen indicators from exactly this family. For a mould or an office, availability is hard to define without stretching.

At reflexivity, success conditions refer to a system representing its own states. You can test this in people with confidence judgements, and in models with probes that compare what the model says about itself with what its internals show. You cannot test it in a mould in any way I know of. At self-recognition, success conditions refer to a self that persists and is identified as one's own. For humans we rely on report and long behavioural histories. For an agent fleet, "its own" depends on how the boundary is drawn, and the boundary may change every time someone edits a configuration file.

Testing reflexivity in a nurse and an agent

The layer where the charter does its most useful work today is reflexivity, because it is where AI claims are multiplying fastest. Leela adds a fifth curve to her slide: "Kestrel knows when it is wrong". Here is the realistic case, worked through.

For Shanthi, the standard test of noticing your own states is a confidence judgement. After each dose she logs how sure she was that the label was right, on a scale from one to four. If her confidence is higher on the cases she got right than on the cases she got wrong, her judgements carry information about her own accuracy. Psychologists call this metacognitive sensitivity, and they measure it by asking how well confidence sorts correct answers from errors. For Kestrel, the parallel test is to ask the model for a confidence score after each code change and check whether the score sorts its correct changes from its broken ones.

The success condition matches well: confidence must discriminate one's own correct and incorrect outputs. The readout also matches, since the same statistic can be computed on both logs. The boundary is where it gets interesting. Shanthi's confidence comes from Shanthi. Kestrel's score is generated text, and the text could be produced by a shortcut that never consults anything like an internal state. A long, unfamiliar task might simply attract a lower number, regardless of whether the change is correct.

That shortcut is Clever Hans again, in a new coat. The control is to match tasks on surface features, such as length and familiarity, and check whether confidence still sorts right from wrong. The intervention fixture asks for more. With Shanthi, you can't reach inside. With Kestrel, interpretability researchers can sometimes alter an internal representation and see whether the model's report changes with it. Work at Anthropic on concept injection in Claude models did something of this kind and found what the authors describe as a limited and "highly unreliable" capacity for introspective report. That is a published result with a narrow scope, and I read it as evidence about one family of models under one method.

The audit result for this pair is graded at best, and the weakest fixture is the intervention, because the strong test is available only for the machine. Note what the verdict does not say. A machine that passes a reflexivity test has shown nothing about self-recognition. Evidence at one RTM layer never stands in for the layer above it, and the fixture card makes that easy to see because the success conditions for the two layers are written on different lines.

Try the charter on Leela's specimens

The interactive figure below lets you run the four-fixture audit yourself. Pick a pair, pick a layer and switch fixtures on as you decide they match. The verdict updates. It is a reasoning aid that encodes my rule. It is not evidence about any real system, and the defaults are my judgements about invented specimens.

An office that remembers a flood

Institutions are the hardest of Leela's four cases, and the most interesting for anyone who works in one. The philosophers Christian List and Philip Pettit argued in Group Agency (2011) that some groups are agents in a literal sense: they form attitudes, make judgements and act, through procedures that are not reducible to any one member's views. Their argument turns on the structure of the group's decision procedure. It doesn't depend on the group feeling anything.

If List and Pettit are right, the Flood Cell can be an agent with beliefs and intentions. Whether it learns from mistakes, in the nurse's sense, depends on where the change lives. That is the next thought experiment.

The two floods matter for AI systems as much as for offices. A deployed agent whose good behaviour depends on a human operator quietly correcting it is City B. Its learning lives in the operator. When the operator moves to another team, the system's apparent competence moves with them. Behaviour that survives the removal of a component is behaviour of the system that remains.

[AUTHOR NOTE: a short first-person account of seeing a deployed system's apparent competence depend on one person's quiet corrections, if you have one]

Comparing averages across cricket eras

The third analogy will be familiar to anyone who has argued about whether a batter from the 1930s was better than one from the 2010s. A Test batting average is a clean number. It is also a number earned on uncovered pitches or covered ones, against different bowling, with different bats and different protective gear. Cricket statisticians adjust for era by holding some things fixed, such as comparing each player to the average of their own time, and they argue about which adjustments are fair.

The mapping is close. The batting average is the readout. The era is the substrate. The adjustments are attempts to match fixtures: same success condition (runs per dismissal), similar interventions (bowlers of comparable quality), controls for conditions (pitch, ball, rules). Era-adjusted comparisons are graded comparisons in my sense, and good cricket writers say so.

Recognition is a comparison with one term hidden

The Pratyabhijñā school of Kashmir, founded in the verses of Utpaladeva (tenth century) and developed in Abhinavagupta's commentaries, is named after an act of comparison. Pratyabhijñā means recognition: the moment you see a thing in front of you and know it as the thing you already knew. The school's aim was to bring a student to recognise "I am Śiva", that is, to recognise their own awareness as the same awareness that the tradition identifies with the whole of reality.

I want to read this on its own terms before building any bridge, so here is the problem the school was solving, in its own vocabulary. Buddhist logicians, chiefly Dignāga and Dharmakīrti, argued that knowledge rests on momentary perceptual particulars (svalakṣaṇa) and that enduring things, including an enduring self, are constructions laid over those moments. Utpaladeva and Abhinavagupta answered that recognition is not an optional interpretation added after perception. It is a condition of having experience at all: of learning a word, linking a memory to a present scene, or acting with a purpose. That reading follows the Internet Encyclopedia of Philosophy's entry on Kashmiri Śaiva philosophy, which glosses prakāśa as "bare subjective awareness" and vimarśa as "recognitive apprehension". The standard critical edition and translation of Utpaladeva's verses is Raffaele Torella's.

The bridge to this paper is narrow, and I want to build it carefully. Recognition has the logical form of a comparison: a present item, a remembered or described item and a judgement that they are the same. Every fixture in my charter has a counterpart. The success condition is "the two are the same". The boundary is the knower who makes the judgement. The intervention is whatever brings the two items into one view, which in the tradition is the teacher's instruction. The readout is the changed way the knower lives afterwards.

The Buddhist side of that old debate deserves its own hearing, because it is a rival reading of the same scene. For Dharmakīrti, what is given in perception is a stream of momentary particulars, and "this is that" is a conceptual construction laid over two different moments. On that reading, Leela at the wedding did not discover a sameness. She built one, usefully, out of a name and a memory. The Pratyabhijñā reply is that the building itself requires a unifying awareness that holds both moments together, and that this awareness is what gets recognised.

I don't need to settle that dispute to use it. Both sides agree that recognition involves a judgement of sameness across two presentations. They disagree about whether the sameness is found or made. My charter is on the Buddhist side in method: it treats every verdict of "literal" as a construction, made by choosing fixtures, and asks only that the construction be public. Whether anything deeper is found is a question the charter can't reach.

Here is the break point, stated plainly. The charter compares two systems from outside. Pratyabhijñā is about one system recognising itself from inside. The scientific claim and the tradition's claim stop matching at exactly the RTM layer named after the tradition. A literal cross-substrate comparison at the self-recognition layer would need a readout of a system's identification of its own awareness. For humans, the only readout is report plus a changed life, and for anything else we have nothing comparable. That is why my comparability gradient bottoms out where it does. The tradition doesn't help us make the comparison literal. It helps us see why it can't yet be.

Four ways to read one maze

Different schools read the same observation in different ways, and the differences are informative only if each reading is stated in its own terms. The lens switcher below takes the slime mould's maze and reads it four ways: a behaviourist one, the "basal cognition" view associated with Michael Levin, an Abhidharma reading (the Buddhist analysis of experience into momentary events) and a Pratyabhijñā reading.

The Levin lens deserves a word of its own, because it is the strongest rival to my charter. Levin's 2019 paper in Frontiers in Psychology proposes that every agent has a "cognitive light cone", defined by the scale of the goals it can pursue: "anything outside this region is mentally inaccessible to that system". On that view, cells, tissues, organisms and swarms all have minds of different sizes, and comparison is a matter of measuring cone size. It is a continuity view. It says there is one capacity, goal-directedness, scaled up and down.

The charter fixes the answer before the question is asked

The strongest objection to everything above is this: whoever chooses the fixtures has already decided which systems can count. Success conditions are written by humans, in human vocabulary, about capacities humans recognise. A charter that insists on matched fixtures will always find that the most humanlike systems compare best with humans. That is not a discovery. It's a restatement of the starting point, wearing a lab coat.

Give the objection its full strength. Levin's continuity view says the right comparison is across scales of goal pursuit, and that my fixtures carve the world at human joints. Cell collectives that rebuild a target anatomy after injury are, on that view, correcting errors; a nurse who catches a misplaced vial is correcting one too. Insisting that the boundary be drawn identically for both is a choice, and a choice that favours bodies with skins. The ethologist Frans de Waal made a related point about animals in 1999, coining the word "anthropodenial" for the refusal to see shared capacities in other species (de Waal, Philosophical Topics, 1999). A charter that sets the bar for literal comparison at "all four fixtures match" could be a machine for producing anthropodenial.

A second rival comes from the opposite side. Theory-led approaches, such as the indicator method of Butlin and colleagues, argue that you don't need matched fixtures at all. You need a theory of the capacity, and then you check each system for the theory's indicators. Butlin's team adopt computational functionalism as a working hypothesis and derive fourteen indicators from leading theories of consciousness. On that view, my fixtures are a theory-free stopgap. When we have good theories, fixtures become redundant.

A third rival, which I take seriously, is the deflationary view that "mind" is not a natural kind at all. If learning in a nurse and habituation in a mould share nothing but a word, then there is nothing for a charter to keep fixed, and the honest move is to stop using the word across substrates. A careful version of this view is part of what Tim Bayne and colleagues confront in their 2024 review of tests for consciousness, where they note that consciousness may not correspond to a unitary kind.

Bayne's review also offers a useful map of how tests get validated, and it shows where the charter sits. The authors describe three strategies. A test can be redeployed from a population where it is already trusted, as when covert command-following in brain imaging is accepted because overt command-following is. A test can be backed by a theory that predicts it. Or a test can be refined iteratively, treating consciousness as a natural kind and revising the test and the list of conscious systems together. The charter is closest to the first strategy, applied to comparisons rather than tests. It inherits that strategy's weakness, which the review states clearly: it is conservative, and it cannot justify extending a comparison beyond the populations where the fixtures were first matched.

What does each rival get right? Levin is right that boundaries are choices, and that the charter's default boundary (the organism, the device, the office) is not given by nature. De Waal is right that a high bar has costs as well as benefits, and that the costs fall on the systems we already undervalue. Butlin's team are right that fixtures are no substitute for theory. The deflationist is right that some cross-substrate words are puns, and the charter will find that out case by case.

Here is where I part company with each.

To Levin, I say that the charter doesn't fix the boundary in advance. It demands that the boundary be stated, and that it be the same on both sides of a comparison. If a cell collective and a nurse are both compared at the scale of "the goal-pursuing unit", that is a legitimate boundary, and the charter will grade the comparison on it.

To de Waal, I say that a graded or analogical verdict is an exact statement of what has and hasn't been shown. To Butlin's team, I say that theories of consciousness disagree, and Bayne's review counts at least 22 taken seriously, so a theory-free discipline is what you need while the theories fight. To the deflationist, I say: run the audit and see. If the audit says "equivocal" across the board, you were right.

The objection that fixtures encode human assumptions is, in the end, partly correct, and I want to record what it costs me. The charter gives precise verdicts relative to fixtures that someone chose. It cannot certify that the choice was right. It can make the choice public, so that someone else can contest it. That is the most a method can do before the theories converge, and it is less than I would like.

The equivocation audit, run on real cases

The instrument for this paper is a short worksheet with a scoring rule. I have also written it as a few lines of Python so that the rule is unambiguous and anyone can run it. It takes two fixture cards and a layer, and returns one of four verdicts.

The inputs are, for each fixture, a judgement of match: "same" (the fixture is identical on both sides), "approximate" (the fixture differs but in a stated, bounded way) or "different". The assumptions are that the fixtures are stated before the data are seen, that a match judgement is made by someone who did not run the experiment, and that a missing control counts as "different". The output is the verdict and the weakest fixture, which is where to spend effort next.

FIXTURES = ["success", "boundary", "intervention", "readout"]

def audit(match):
    """match: dict fixture -> 'same' | 'approximate' | 'different'"""
    vals = [match[f] for f in FIXTURES]
    if all(v == "same" for v in vals):
        verdict = "literal"
    elif vals.count("different") == 0:
        verdict = "graded"
    elif match["success"] != "different" and vals.count("different") <= 2:
        verdict = "analogical"
    else:
        verdict = "equivocal"
    order = {"different": 0, "approximate": 1, "same": 2}
    weakest = min(FIXTURES, key=lambda f: order[match[f]])
    return verdict, weakest

cases = {
  "thermostat vs person, regulation (minimal)":
     dict(success="same", boundary="same", intervention="same", readout="same"),
  "Physarum vs infant, habituation":
     dict(success="same", boundary="same", intervention="approximate", readout="same"),
  "Kestrel vs Shanthi, error learning":
     dict(success="approximate", boundary="different", intervention="approximate", readout="approximate"),
  "Flood Cell (City B) vs Shanthi, error learning":
     dict(success="approximate", boundary="different", intervention="different", readout="approximate"),
  "rutted road vs Shanthi, error learning (negative control)":
     dict(success="different", boundary="different", intervention="different", readout="approximate"),
}
for name, m in cases.items():
    print(name, "=>", *audit(m))

thermostat vs person, regulation (minimal) => literal success
Physarum vs infant, habituation => graded intervention
Kestrel vs Shanthi, error learning => analogical boundary
Flood Cell (City B) vs Shanthi, error learning => analogical boundary
rutted road vs Shanthi, error learning (negative control) => equivocal success

Running the audit well matters more than running it often. Here is the procedure I would use with a colleague:

1. Write the four fixtures for each system before looking at any results.
2. Hand the cards to someone who did not run the experiment and ask them to mark each fixture same, approximate or different.
3. Treat any missing control as "different", and any forbidden intervention as "unavailable", noted separately.
4. Record the verdict and the weakest fixture, and keep the cards with the data.
5. Re-run the audit after any change to the systems, since boundaries drift.

The positive case is the thermostat and the person, compared on regulation around a set point. All four fixtures match, the verdict is literal, and the capacity is thin. The negative control is the one I find most useful in teaching. A dirt road that develops ruts "learns" where the carts go, in the sense that repeated traffic changes the road so that later traffic follows the same path more easily. The success condition is different from the nurse's, since nothing about the road reduces an error. The verdict is equivocal. Any audit that called the rutted road literal on "learning" would be broken, and this one isn't.

The realistic case is Kestrel and Shanthi. The audit says analogical, and the weakest fixture is the boundary. That is useful, because it tells Leela exactly what to change. If she redraws Kestrel's boundary to include the reviewing engineer, she changes the comparison into one between a human-plus-agent team and a nurse, which might match better. If she keeps the boundary at the agent and removes the engineer's approval step, the success condition changes too, because the memory file would then be written without review. Either way she learns something about her claim.

Where the charter stops helping

Every method has a region where it stops paying for itself. For the charter, I can name three edges.

The first edge is phenomenal experience. The charter compares capacities that have outside success conditions. Whether there is something it is like to be a mould, a model or a nurse has no outside success condition that everyone accepts; that problem is the subject of my third paper. The charter can grade comparisons of access and reflexivity. It is silent on felt experience, and it is silent on the luminous ground.

The second edge is very distant substrates at high layers. A literal comparison of self-recognition between a person and a city office would need a readout of self-identification that works for both. I don't know of one. In that region, the charter's verdicts will be "analogical" or "equivocal" almost everywhere, which is honest but not very informative.

The third edge is cost. Matching interventions across systems can be expensive or impossible. You can delete Kestrel's memory line in a second. You can't ethically delete a nurse's memory of a near-miss. Where the matched intervention would be unethical, the charter will grade a comparison lower than its truth may deserve.

That last edge is the failure boundary I care most about. The charter can call a comparison analogical because a test is forbidden, not because the capacities differ. I mark those cases explicitly in the audit, with a note that says "intervention unavailable", so that a low grade is not read as a negative finding.

What Arjun should do with the preprint

Return, one last time before the objections, to Arjun's desk. He has the audit outputs. Shanthi and Kestrel: analogical, weakest fixture the boundary. Shanthi and Plate 7: graded at the responsiveness layer, if Leela reframes the claim as habituation and adds the specificity and recovery tests from the Boisseau design. Shanthi and the Flood Cell: depends on the city, and the preprint doesn't say which.

Arjun's decision is no longer "publish or reject". It is a list of edits. Change the title so that it claims what the fixtures support. Report each pair's verdict in the abstract. Add the one intervention per pair that would move it up a grade. That is a better paper than the one Leela submitted, and it is still a paper with a striking finding: a nurse and a slime mould can be compared literally on a thin capacity, and that is worth knowing.

[AUTHOR NOTE: a short first-person account of reviewing or editing a cross-disciplinary claim that rested on a shared word, if you have one]

What would change my mind

I would weaken or drop the charter's central claim if any of the following turned up.

If careful audits by independent groups found that literal comparisons were as common at self-recognition as at responsiveness, the comparability gradient would be wrong. That would hurt the RTM prediction about the L1 to L4 ordering directly, because it would suggest the layers don't differ in how much their success conditions depend on particular kinds of system.

If two groups auditing the same pair of systems with the same worksheet routinely reached different verdicts, the instrument would be unreliable, and the four fixtures would be too vague to do the work I claim for them. I would want inter-rater agreement well above chance before trusting the verdicts. Until someone measures it, I treat the worksheet as a reasoning aid.

Suppose a theory of consciousness or of learning earned broad agreement, and its indicators predicted which comparisons hold better than the fixtures do. Then Butlin's camp would be right that fixtures are a stopgap, and the charter would become a historical document. I would count that as good news.

Suppose a strong continuity view, such as Levin's, produced a single scale of goal-directedness. If that scale predicted behaviour across cells, animals, machines and institutions better than layer-specific fixtures, it would hurt the RTM claim that L2 access and L3 reflexivity are separate layers. It would suggest they are points on one dimension.

There is a quieter way the charter could fail. If audits mostly produce the same verdict as a reader's gut sense of how similar two systems look, the fixtures are adding ceremony and nothing else. I would test this by asking people to guess verdicts from photographs and short descriptions, then comparing their guesses with full audits. If the guesses match the audits most of the time, I would have to admit that the charter mostly formalises intuition.

What I still don't know

I don't know how to state a success condition for self-recognition that a non-human system could meet without the condition being rewritten to suit it. I don't know whether "boundary" can be made objective for software systems whose components change weekly. I don't know whether the four fixtures are the right four, or whether a fifth, such as timescale, should be separated from the boundary. I suspect it should, and I haven't tested it. And I don't know of any readout for the luminous ground at all, which is why RTM takes no position on whether any machine has it.

Claim ledger

These are the paper's main claims as I would enter them in the programme's Claim Ledger. Confidence uses three bands: established (the evidence is strong and I know of no serious challenge), likely (the evidence favours it but there is a live dispute) and speculative (my hypothesis, argued here but not tested).

There is one dated prediction. By 31 December 2028, I expect that at least one published inter-rater study of a fixture-style audit, on at least ten cross-substrate comparison claims, will report agreement on the literal or equivocal ends of the scale that is higher than agreement on the middle grades. I will judge it by searching for such a study and reading its agreement statistics. If no study exists by then, the prediction is unresolved, and I will record it as such.

Glossary

Analogical comparison

A comparison where the structure is similar but some fixtures differ. Useful, provided it carries the label.

Boundary

The line drawn around what counts as the system being tested, including any people, tools or files that contribute.

Equivocation

Using one word with two meanings in an argument as if the meaning stayed the same.

Fixture

One of the four things held fixed in a comparison: success condition, boundary, intervention, readout with controls.

Graded comparison

A comparison where no fixture differs outright but some are only approximately matched.

Habituation

A decline in response to a repeated stimulus that recovers when the stimulus is withheld and is specific to that stimulus.

Homologue and analogue

Owen's 1843 terms: the same part inherited across species, versus different parts doing the same job.

Literal comparison

A comparison where all four fixtures match, so the shared word means one thing on both sides.

Morgan's canon

The 1894 rule that you shouldn't explain behaviour by a higher capacity if a lower one explains it.

Prakāśa प्रकाश

Light or shining; in Pratyabhijñā, the awareness by which anything appears. RTM layer 2 borrows the name for access.

Pratyabhijñā प्रत्यभिज्ञा

Recognition; the Kashmiri Śaiva school of Utpaladeva and Abhinavagupta. RTM layer 4 borrows the name for self-recognition.

Spanda स्पन्द

Pulse or vibration. RTM layer 1 borrows the name for responsiveness.

Success condition

The observable state of affairs that would make a capacity claim true.

Svalakṣaṇa स्वलक्षण

In Buddhist epistemology, a momentary, unique particular known in perception.

Vimarśa विमर्श

Reflective grasp of what appears; in Pratyabhijñā, recognitive apprehension. RTM layer 3 borrows the name for reflexivity.

Cit चित्

Awareness itself. RTM layer 5, the luminous ground, on which RTM takes no position for machines.

The slide Leela ends up presenting

Six weeks later, Leela presents a different slide at her grant review. It still has four curves. Each now carries a small coloured tag: literal, graded, analogical, analogical. Underneath each tag is one line saying which fixture is weakest and which experiment would move it up. Shanthi and Plate 7 sit together under the heading "habituation-like decline, specific to the stimulus, recovering with rest", and that comparison is literal at the responsiveness layer because Leela ran the specificity and recovery tests. Kestrel sits beside Shanthi under "analogical: boundary includes the reviewing engineer". The Flood Cell is split into two curves, one per city.

The review panel spends its time on the tags, not the curves. One member asks why the Kestrel boundary can't be redrawn. Another asks whether the City B office counts as a system at all. Those are the right arguments to have, and they could only happen because the fixtures were written down.

The decision that changes is a small one with large effects. No claim that two kinds of system share a mental capacity should be published, funded or regulated on without its four fixtures and its verdict printed beside it. For builders, that means stating the boundary of an AI system's capacity claim, including the humans and tools that make it work. For researchers, it means reporting the weakest fixture as routinely as a confidence interval. For regulators, it means reading "the system understands" as an unfinished sentence until someone says understands what, measured how, inside which boundary.

Arjun, for his part, publishes the revised paper with a different headline from the one the press office wanted. It reads "A nurse and a slime mould habituate in the same way; an agent and an office learn in ways that only look alike". It is longer and less exciting. It is also true, and a reader can check every word of it against the fixture cards in the supplement.

There is one more thing the audit taught me while I wrote it. The four fixtures were easy to state for every capacity except self-recognition, where I kept wanting to write "the system recognises itself" as the success condition and then realised I had no outside way to check it. That difficulty is the subject of most of the rest of this programme. My next paper asks a narrower version of it: when we define a measurable stand-in for consciousness, at what point does the definition start deciding what consciousness is?

Sources

Cite this work

@misc{mahapatra_U0001,
  author       = {Mahapatra, Rajesh Ranjan},
  title        = {A charter for comparative mind science: what to hold fixed
                  before you compare a nurse, a model, a mould and a city office},
  howpublished = {Mind Frontier, paper U0001},
  url          = {https://rajeshrmahapatra.com/papers/mind-frontier/U0001-charter-for-comparative-mind-science.html},
  note         = {DOI: to be assigned}
}