The order that said no peanuts

At half past nine on a Friday night, an order arrives at a dumpling kitchen in Indiranagar. It is written in Chinese, sent through a messaging app by a customer from a tour group staying nearby. Twelve pork and chive dumplings, one bowl of dan dan noodles, and a line at the bottom: 花生严重过敏. Severe peanut allergy.

The kitchen belongs to Lin Wei, who moved to Bengaluru from Chengdu eleven years ago. Six months ago she installed a robot line cook her staff call Bao. Bao is two arms on a rail, four cameras, a set of scales and a language model trained to read orders, plan cooking steps and talk to staff. It also keeps a running model of the kitchen. It knows which jar is where, what has been used, what is cooking and what each dish contains. Tonight, the dan dan noodles are the problem. The house recipe finishes them with crushed peanuts.

Lin is not in the kitchen. She is at her daughter's school concert, watching her phone. Bao has already flagged the order. Its message to the night manager, Prakash, reads: "Customer reports a severe peanut allergy. Dan dan noodles normally contain peanut. I will use the sesame-only finish, change gloves, and use the clean wok. Please confirm." Prakash grew up in Mysuru and reads no Chinese at all. He confirms, the dish goes out without peanuts, and the customer is fine.

The next morning, Lin has to decide whether Bao can be trusted with allergy orders when she is not there. Her insurer wants an answer. Her cousin, who studied philosophy, sends her a link to a 1980 paper by John Searle and a message: "Your robot does not understand Chinese. It just shuffles symbols. Do not let it near allergies."

The kitchen, Lin, Prakash and Bao are invented. They are a way to make the argument concrete, never evidence for it. The paper Lin's cousin sent is real, and so is the question it raises.

Searle's argument, in its famous form, puts a person in a room with a rulebook. The person does not read Chinese. Questions in Chinese come in through a slot, and the rulebook tells him which Chinese characters to send back. If the rules are good enough, the answers fool native speakers. Searle's point is that the man understands nothing. Running a program is just following rules for shuffling symbols. So, he concludes, no computer understands anything merely by running a program.

The question this paper asks is sharper than the one Searle answered. Bao is not a man with a book. Its rules are learned, not written. They run inside a system that sees the kitchen, acts on it and keeps a model of it. Put the rulebook inside an embodied system with a causal model of its world. Does it still fail Searle's test for intentionality, the property of being about something?

Here is my answer. Yes, it still fails Searle's test, but only because Searle's test is built so that nothing except a brain-like causal system can pass. That makes the test unanswerable by any evidence about Bao. The useful part of Searle's argument is a different and smaller claim. Symbol manipulation alone does not fix what the symbols are about. That claim is right. But it turns the real question into one we can measure: are a system's symbols coupled to the world in the right counterfactual way? Bao's word for peanut either changes when the peanuts change, drives what Bao does with peanuts, and gets corrected when it is wrong, or it does not. Those are tests.

The consequence for Lin is concrete. She should not ask whether Bao understands Chinese. She should ask whether Bao's peanut symbol is coupled to peanuts, and she should test it by swapping the jars.

What Searle argued, and what he did not

Start with something everyone has done: getting by in a country whose language you do not speak, using a phrasebook. You point at the phrase for "where is the station?" and say it badly. Someone points left. You have produced a correct question and received a correct answer, and you still do not know the language. If someone asks you a follow-up question, you are lost.

The Chinese Room is the phrasebook taken to its limit. Searle published it in 1980 in Behavioral and Brain Sciences, a journal that prints a target article followed by many short replies from other researchers. The paper was aimed at a specific claim made by some researchers in artificial intelligence at the time. Searle called it strong AI: the claim that an appropriately programmed computer literally has a mind, rather than simulating one, and that its programs explain human understanding.

His target was a story-understanding program by Roger Schank's group at Yale. Given a short story about a man who orders a hamburger, gets it burnt and storms out without paying, the program could answer questions such as "did the man eat the hamburger?" Some researchers said the program understood the story. Searle's reply was to imagine himself executing the same program by hand, in English, on Chinese symbols. He would produce the same answers and understand no Chinese. So the program, he concluded, does not supply understanding either.

Ten years later, in a Scientific American article titled "Is the brain's mind a computer program?", Searle put the argument as a short proof. Programs are formal, which is to say syntactic: they are defined by the shapes of symbols and rules for moving them. Minds have mental contents: thoughts are about things. And syntax by itself is neither constitutive of nor sufficient for semantics. Conclusion: programs are neither constitutive of nor sufficient for minds.

Notice how narrow the conclusion is, because Searle did not say machines cannot think. He said repeatedly that the brain is a machine and it thinks. He did not say that a computer could not be part of a thinking system. He said that running a program is never, by itself, enough. A thinking machine would need what he called the causal powers of the brain, whatever those turn out to be. This positive view is biological naturalism: mental states are caused by neurobiological processes and are features of the brain, as liquidity is a feature of water.

The replies, and Searle's answers to them

The 1980 paper printed Searle's answers to replies he had heard when presenting the argument. Each one tries to locate understanding somewhere he had missed. Two of them matter most for Bao.

The systems reply says that the man does not understand Chinese, but the whole system does: man plus rulebook plus scratch paper plus room. The man is like a single neuron, and no neuron understands English either. Searle's answer was to shrink the system into the man. Let him memorise the rulebook and do all the calculations in his head. Now there is nothing in the system that is not in him. He still understands no Chinese, so the system does not either.

The robot reply says that the trouble is the room's isolation. Put the program inside a robot with cameras and arms, so that the symbols are connected to perception and action. Then the symbols are about the things the robot sees and handles. Searle's answer was to put the man inside the robot's head. Some of the incoming symbols now come from cameras, and some outgoing symbols move motors. The man still sees only symbols. He does not know that one string means "the robot is looking at a hamburger". Adding sensors adds more symbols, not meaning.

What the argument gets right

I want to state Searle's strongest point fairly before I pull at it. A string of symbols has no meaning on its own. The marks 花生 mean peanut to a Chinese reader because of a vast background of use, perception and teaching that connects those marks to peanuts. A program that manipulates the marks according to their shapes has none of that background. It can be run on marks that mean something entirely different, or nothing. The same program, run on the same marks, could be relabelled as a chess engine if the marks happened to line up. Meaning is not in the shapes.

That point is right, and it has a respectable modern descendant. In 2020 Emily Bender and Alexander Koller argued, in "Climbing towards NLU", that a system trained only on the form of language cannot learn its meaning. Their thought experiment is an octopus that taps into an undersea cable between two islanders. It learns to predict their messages perfectly. Then one islander is attacked by a bear and asks for help building a weapon from sticks. The octopus, which has never seen a bear or a stick, fails. Their claim is that meaning requires a connection between form and something outside form: the world, or the speaker's intentions.

So the shared claim between Searle and Bender and Koller is this. Symbol manipulation alone does not fix what the symbols are about. I think that claim is established, in the sense that no one has given a good argument against it. What is contested is what follows. Searle concluded that no program could ever be enough. Bender and Koller concluded that text-only training is not enough. Those are different conclusions, and the gap between them is where Bao lives.

Grounding is the problem Searle found

In 1990 Stevan Harnad gave the problem its modern name in a paper called "The symbol grounding problem"90087-6). His image was a dictionary-go-round. Suppose you try to learn Chinese using only a Chinese-Chinese dictionary. Every definition is made of other Chinese words, and you go around in circles forever. Some symbols, Harnad argued, have to be grounded in something other than symbols, such as the categories a system forms from what it senses. Only then can the rest be built on them.

Harnad accepted Searle's argument against the pure room and rejected his conclusion about robots. A robot that could pass a full Turing test in the world, recognising, handling and naming things as people do, would have grounded symbols. Searle's reply, putting the man in the robot's head, misses the point. The grounding is not in what the man sees. It is in the causal link between the robot's categories and the things they pick out.

That is the move this paper builds on. Grounding is a causal and counterfactual relation, not a feeling of understanding. A symbol is grounded when it would change if the world changed in the relevant way, when it guides action on the thing it is about, and when errors about that thing get corrected. Each of these can be checked.

ELIZA is the room with the walls removed. Its rulebook was tiny and public, and it still produced a feeling of being understood. That is the first lesson the history teaches, and it supports Searle's caution. Fluent output is cheap evidence of understanding. People supply the meaning, and then credit the machine with it.

The second lesson is quieter. ELIZA's words for mother and family were coupled to nothing. Change a user's family and nothing in the program would change. In the language of this paper, ELIZA fails Track for every word it uses. It is the clean negative case for grounding, and it was clean long before anyone built a world model.

A rulebook that grew a world model

Return to the kitchen and change one feature. Bao's rules were not written by anyone. They were learned. What happens to the Chinese Room when the rulebook is a trained network with an internal model of the kitchen?

The first thing to see is that the learned version is in some ways more room-like than Searle's. Searle imagined rules a person could, in principle, read. A trained network's rules are billions of numbers. No one can read them as rules at all. If the original room was a phrasebook, this is a phrasebook written in a code nobody knows, by a process nobody supervised line by line.

The second thing to see is that the learned version can contain something Searle's room did not: a model of the world it is about. That claim used to be speculation, and it is now a lab result.

That edit is the kind of experiment the Chinese Room debate never had. It is an intervention on the internal symbol, followed by a check that behaviour changes in the way the symbol's content predicts. The board representation is about the board in a precise sense. Change it, and the system acts as if the board were different.

Othello is a toy world, with 64 squares and no peanuts. The claim that larger models trained on human text build world models of comparable quality is contested. Wes Gurnee and Max Tegmark reported in 2023 that language models represent space and time linearly: city coordinates and historical dates can be read from activations. Critics reply that reading a quantity out of activations is not the same as the model using it. Still, the Othello result breaks one assumption in the old debate. A system trained only on symbols can build an internal model whose content is checkable by intervention.

What the world model does to Searle's argument

Here is where the old replies and the new evidence meet. Searle's answer to the robot reply was that sensors only add more symbols. That is true of the man in the robot's head. It is less plainly true of the system, and the Othello experiment shows why. The internal board is more than extra symbols. It is a structure whose parts stand in the same relations as parts of the board. When the board changes, the structure changes, and the structure's changes drive moves.

Philosophers call this structural resemblance an isomorphism, and resemblance alone is cheap. Many structures resemble many others. Resemblance alone does not make the Othello board about the board. It is that the resemblance is maintained by the causal process that links moves to board states. A structure that resembled an Othello board by accident would not update when a move was played.

Dimitri Coelho Mollo and Raphaël Millière put this sharply in a 2023 paper, "The vector grounding problem". They separate several things people mean by grounding. A system might have sensory grounding, from senses. It might have referential grounding, which ties its states to things in the world through the right causal history. Their claim is that referential grounding is what matters for meaning, and that it does not strictly require senses. A system can acquire it through a history of interaction and feedback. Training with human feedback, for example, can tie outputs to whether they are true. If they are right, a language model can have some grounded states even without eyes.

I think they are partly right, because the history matters. But I want a test that does not depend on reconstructing a model's training history, which is often secret. That is what the third part of this paper offers.

The two kitchens

The most useful thought experiment for Lin keeps the dish fixed and changes what happens inside.

This is the experiment I would put in front of Lin's cousin. Searle's argument cannot tell Bao from Dim, because both run programs and neither has a brain. A grounding test can tell them apart, and the difference is the one the customer's life depends on.

Three analogies and their limits

Three analogies run through this debate. Each maps something real and each breaks somewhere specific.

The phrasebook tourist maps the room well. The book is the program, the tourist is the processor, the correct exchanges are the outputs. It breaks because real tourists learn. After a week of pointing at the phrase for "station" and walking where people point, the tourist starts to understand. The feedback from the world fills the phrase with meaning. Searle's room has no such feedback, and Bao does.

The thermostat is the analogy Searle's critics used to ridicule him, and the one his defenders use back. A thermostat's bimetallic strip bends with temperature and switches the heating. In a thin sense, the strip is about the room temperature: it tracks it, and its state drives action on it. That maps grounding as coupling. It breaks because the thermostat has one variable, no capacity to be wrong in an interesting way and no way to correct itself except by bending back. Grounding in a system like Bao involves many symbols with structured relations, and errors that can be noticed.

The simulated rainstorm is Searle's own analogy. A computer simulation of a rainstorm does not make anything wet, so why would a simulation of understanding understand? It maps the difference between simulating a process and having it. It breaks for processes defined by what they do with information rather than by their physical medium. A simulated calculator really does arithmetic. If understanding is more like arithmetic than like wetness, the analogy fails. Whether it is depends on what understanding is, which is what is in dispute. So the analogy restates the conclusion rather than supporting it.

Swap the jars: a test for grounding

Lin cannot wait for philosophers to agree about understanding. She needs to know, by Monday, whether Bao's handling of the word for peanut is safe. So change one feature of the running example: Lin stops asking what Bao is and starts asking what Bao's peanut symbol does.

What follows is the instrument of this paper. I call it the Counterfactual Coupling Test, and it asks three questions about one symbol at a time. It is not a test for understanding. It is a test for one necessary ingredient of understanding that Searle's argument shows a program cannot supply by itself: a link between a symbol and the thing it is supposed to be about.

The three probes map onto three things people mean when they say a symbol is about something. Track is the old idea that a symbol carries information about what causes it. Use is the idea that a symbol's content is fixed partly by what the system does with it. Correct is the idea, which goes back at least to the philosopher Ruth Millikan's work on biological function, that content requires the possibility of error and something that counts as getting it right.

Each probe alone can be faked. A thermostat passes Track and Use for temperature, and fails Correct. A label-reading system passes Use, if you edit its lookup table, and fails Track. A system that re-checks constantly but never acts on what it finds would pass Track and Correct and fail Use. Only the combination says that the symbol is doing work for the system, about the thing, in a way that responds to the thing.

The minimal case: one sensor, one symbol

Start with the smallest system that could pass. A peanut sensor sits under one jar. It reads 1 when the jar contains peanuts and 0 otherwise, with 90 per cent accuracy. The sensor's output is copied into a single memory cell, S, and S directly decides which jar the arm avoids. Nothing ever re-checks S after it has been set.

I ran the test on this system using a short program that enumerates every case exactly, with no random sampling. The porter swaps the jars on 30 per cent of nights. The results:

The minimal system is coupled for Track and Use and not coupled for Correct. It is, in this sense, a thermostat for peanuts. Its symbol carries information about peanuts and drives action on peanuts, but it cannot be wrong in a way it could notice. Hazard is low only because the sensor is good, not because the system checks its own work.

The point of the minimal case is to show that the test is not mysterious. Any engineer can run it. And it is to show that passing two of three is not enough to say a symbol is grounded in the full sense. Searle would add, rightly, that nobody thinks this sensor understands peanuts. That is where the test's limit begins, and I come back to it.

The realistic case: Bao and Dim

Now the real kitchen. Bao tracks jars by what goes into them, using cameras, and keeps a shelf model that updates as things move. Its arms consult the shelf model, not the labels. Before each allergy order, Bao looks again at the jar it intends to use, which is a re-check. Dim, the Koramangala robot, picks jars by label and has no shelf model.

I set Bao's parameters from Lin's maintenance logs, which are invented like the rest of the kitchen. Bao's shelf model is updated from cameras on 98 per cent of jar movements, and its jar recognition is 97 per cent accurate. Its arms follow the shelf model on 95 per cent of moves, and it re-checks before 90 per cent of allergy orders. The program's output was:

Bao passes all three probes above the working threshold. Dim passes none. On the nights the porter swaps the jars, Dim always serves peanuts to an allergy order, because its symbol tracks labels and the labels did not move. Across all nights, that is a 30 per cent hazard, against Bao's 1.7 per cent.

The two robots write the same messages, and their peanut symbols are about different things. Dim's symbol is about the label. Bao's is about the peanuts. That is not a metaphor. It is a measured difference in what the symbol responds to, what it controls and what corrects it.

The negative control

Every test needs a case where it should fail, built so that the only thing removed is the thing being tested. For the Counterfactual Coupling Test, the negative control is Bao with its cameras covered. Everything else is identical: the same network, the same shelf model, the same arms and the same habit of re-checking. Only the causal link from jars to the shelf model is cut.

The output: Track 0.00, Use 0.95, Correct 0.00, hazard 0.30. With the cameras covered, Bao is Dim with better arms. Its symbol still drives action, but it no longer tracks or corrects. The hazard rises to exactly Dim's level.

This matters for Searle's argument in a specific way. The covered Bao runs the same program as the working Bao. If understanding, or grounding, were a property of the program, covering the cameras would change nothing. It changes everything the test measures. So whatever grounding is, it is not a property of the program alone. On this point the test agrees with Searle. Grounding belongs to the program in its world, not to the program.

Turning the dial

The program can also sweep one parameter while holding the rest fixed. I held jar recognition at 0.97, use at 0.95 and re-checking at 0.9, and turned the coupling from cameras to shelf model from zero to one. The results were smooth. At coupling 0.25, Track was 0.24 and hazard 0.18. At 0.5, Track was 0.49 and hazard 0.10. At 0.75, Track was 0.73 and hazard 0.04. At full coupling, Track was 0.97 and hazard 0.016.

Two things follow from the sweep. First, grounding comes in degrees. There is no point on the dial where Bao's symbol jumps from meaningless to meaningful. Searle's argument is all or nothing: either a system has intrinsic intentionality or it does not. The test says that the part of meaning we can measure is graded. Second, Correct rises more slowly than Track, because it depends on Track: you cannot repair an error with a sensor you do not trust. A system with weak coupling is worse at correcting itself than at noticing changes.

A counterexample: the jar on the left

Here is a case where the test, run carelessly, gives the wrong answer. Lin's kitchen keeps the peanut jar on the left of the shelf, always. Suppose a third robot, Momo, has a symbol that tracks "the jar on the left", not "the jar with peanuts". When the porter swaps the contents of the jars, he also swaps their positions, because he puts each jar back where its contents used to be. Momo's symbol follows the swap, so it passes Track. Edit the symbol and Momo's arm moves to the other side. It passes Use. Set it wrong and Momo looks again at the left of the shelf and repairs it. It passes Correct.

Momo's symbol is about position, and it passes a test for being about peanuts. The fault is in the interventions, not the scores. Every intervention we ran kept peanuts and position together, so the test could not tell them apart. This is an old problem in philosophy. Willard Van Orman Quine's example was a linguist who hears a speaker say "gavagai" when a rabbit runs by, and cannot tell whether the word means rabbit, rabbit part or a stage in the life of a rabbit. Every pointing at a rabbit is also a pointing at its parts.

The repair is part of the instrument. For every rival content, run an intervention that separates it from the intended one. Move the peanut jar to the right without opening it. Bao's symbol follows the peanuts; Momo's stays on the left. One extra intervention exposes the difference. The general rule is that the Counterfactual Coupling Test does not tell you what a symbol is about. It tells you which of the contents you thought of it is not about.

The failure boundary

The test needs something to intervene on. For peanuts, that is easy: you can swap jars, move them and empty them. For much of what an allergy order says, it is not.

Take the word for "severe" in the customer's message. What would an intervention on severity be? You cannot make the customer's allergy more severe to see whether Bao's symbol follows. Take the customer's intention, the thing Bender and Koller said meaning must connect to. You cannot reach into the customer's head and change what she meant. For abstract words, social words and words about minds, there is no jar to swap.

That is where the test stops working, and it marks a real limit on what grounding by coupling can explain. A system could be well grounded for every object in the kitchen and still have no coupled symbol for what the customer wants. Its handling of that would rest on patterns in text, which is the borrowed grounding Dim has for peanuts. Coupling grounds the nouns of the kitchen first and the meaning of the order last. Whether the rest can be built on the nouns, as Harnad hoped, is open.

What the test does not do

I want to be exact about the boundary. The Counterfactual Coupling Test measures grounding in a causal sense. It does not measure intrinsic intentionality, which is what Searle said minds have and programs lack. By intrinsic intentionality he meant aboutness that belongs to a state in itself, rather than being assigned by an observer. A map is about a city because people use it that way. A thought about a city, Searle held, is about the city whoever is or is not looking.

The test assigns aboutness from the outside, by intervening and watching. A Searlean will say that is exactly derived intentionality: the engineer decides what counts as the right response. That objection is fair. It is the reason the test cannot settle the original question. But it also shows why the original question is hard to make empirical. Any test of intentionality that a scientist can run will intervene from outside and watch. If intrinsic intentionality is defined as whatever such tests leave out, no evidence can ever bear on it.

Pipes, magnets and a reciter

A test that only confirms what its author hoped is not worth much. This part runs the Counterfactual Coupling Test against four cases built to break it, or to break the conclusion I drew from it. Each changes one feature of the kitchen.

The first comes from Searle himself, and it is the one I find hardest.

I accept the water-pipe case's verdict about the test. I do not accept its verdict about understanding, and the reason is the next case.

The luminous room is aimed at Searle, but it has a second target. It is also the strongest case a rival could make against the hypothesis I test in this series, the Recognitive Triadic Model, RTM. RTM says understanding needs more than coupling. It needs a layer it calls vimarśa, a reflexive grasp in which a state is known as meaning something. If intuitions about the room are unreliable, the Churchlands' camp will say, so are intuitions that something called vimarśa is present in us and missing in Bao. RTM's case for vimarśa rests partly on introspection, and introspection is the evidence the luminous room calls into doubt. I return to this in Part 5.

The supervisor who also cannot read Chinese

The third case is the funny one, and it turns on the systems reply.

The reciter

The fourth case comes from India, and it is older than any of the others by about two and a half thousand years.

So the problem Searle posed in 1980 has an ancestor: a practice that preserves form perfectly, and a debate about whether form preserved without understanding carries meaning. The tradition's grammarians went further than Yāska. The fifth-century philosopher of language Bhartṛhari, in his Vākyapadīya, argued that the meaning of a sentence is not assembled word by word. It arises as an undivided flash of understanding, which he called pratibhā. On his view, the sounds are only the occasion for meaning. The meaning itself is a whole, grasped at once.

[AUTHOR NOTE: a short account of chanting a text you did not understand, and what changed once you learned its meaning.]

The uncomfortable case

The last case is the one I find hardest to write, because it is about people.

Five layers, two questions

This series tests a hypothesis about consciousness, the Recognitive Triadic Model, which I treat as a claim under test and not as a doctrine. RTM draws its vocabulary from the Pratyabhijñā school of Kashmir Śaivism, the tradition of Utpaladeva and Abhinavagupta. It describes five layers. L1, spanda, is pulsation or responsiveness: a system that changes with what affects it. L2, prakāśa, is manifestation: something being present or lit up. L3, vimarśa, is reflexive grasp: the manifest being known as such and as meaning something. L4, pratyabhijñā, is recognition: the self recognising itself in what it knows. L5, cit, is consciousness as ground. For this paper only the first three matter.

Change one feature of the running example. Lin's cousin has read about RTM and asks a sharper question. Not "does Bao understand Chinese?" but "which layer would Bao need, and which can be tested?"

Mapping the test onto the layers

The Counterfactual Coupling Test's Track probe tests L1 directly. A symbol that changes with the world is responsive in exactly RTM's sense. Bao has it; Dim does not, for peanuts; the room as Searle described it does not, because it sees nothing but the slot.

Use and Correct test something RTM does not have a separate layer for: the integration of a responsive state into the system's control of its own action, and its revision when it conflicts with the world. Correct in particular has a reflexive shape. The system checks its own representation against the thing it represents. That is a thin, functional version of what RTM calls vimarśa.

Is it the real thing? In Utpaladeva's account, vimarśa is not merely a check. It is the very nature of consciousness as self-aware, and Utpaladeva calls it the supreme speech, parā vāk, which arises of its own accord. That is a direct link to Bhartṛhari, whose idea that all cognition is shot through with language the Pratyabhijñā thinkers took over and transformed. On this reading, understanding a sentence is a case of vimarśa: consciousness grasping a manifestation as meaningful.

So RTM's verdict on the room is subtle. The man in the room has full vimarśa, about the rulebook. He has none for the Chinese characters, which never become meaningful to him. Bao has strong L1 coupling and a functional loop that resembles L3. Whether that loop is accompanied by manifestation, L2, and by a grasp that is manifest to Bao, RTM cannot test.

Where RTM has to commit

That leaves RTM exposed in the same place as Searle. If vimarśa is defined as whatever the Correct probe leaves out, RTM inherits Searle's problem. Its key notion becomes untestable from outside. The luminous room then applies to it: our sense that vimarśa is something extra may be an intuition produced by our position, not a discovery.

RTM can avoid this only by saying what vimarśa adds that a test could detect. Here is my attempt, which I offer as a hypothesis. A system with vimarśa for a symbol should be able to take the symbol itself as an object: report on it, doubt it and revise its confidence in it, separately from revising the symbol's content. Correct tests whether the symbol gets fixed. A vimarśa test would ask whether the system knows it was fixed, and why. That is testable in principle, with model introspection methods like those discussed in the previous paper in this series. It is a higher bar than Bao has been shown to meet.

The finding for RTM is that it is refined, not supported. The Chinese Room shows something RTM needs: a gap between responsiveness and grasp. But RTM's account of the grasp must become testable, or the same luminous-room objection that weakens Searle weakens it.

A first attempt at the probe

Here is what the probe might look like in Lin's kitchen. Change one feature: after the taped-camera test, Lin asks Bao a question in English. "On Tuesday you changed your mind about which jar held peanuts. Why?"

There are three kinds of answer Bao might give, and they come from different machinery. The first is a fluent explanation written after the fact: "I noticed that the jar contents had changed." That sentence could be produced whether or not anything inside Bao represents its own revision. It is a report, and the previous paper in this series showed how weak reports are as evidence of what produced them.

The second kind of answer is checkable. Bao's logs hold a record of the moment its shelf model changed, with the camera frames that caused it. If Bao's explanation names the right moment and the right cause, and would have named a different moment if the logs had been different, then its report tracks its own revision. That is the Track probe run one level up, on the symbol's history instead of the jar.

The third kind is the one RTM needs. Suppose an engineer edits Bao's internal record of its own revision, so that Bao's memory says the change happened for a different reason, while the shelf model is left alone. Does Bao's account change to match? And does its confidence in the peanut symbol change too, because it now has a worse reason for holding it? If the account changes and the confidence moves with it, Bao is treating its own symbol as an object, with grounds. That is a functional version of what Utpaladeva described: awareness taking up its own content as its own.

I call this the second-order version of the test, and I want to be honest about it. It has not been run on any robot cook. Something close to it has been run on language models: methods that plant a concept in a model's activations and ask whether the model notices. The results so far, discussed in the previous paper, show that models sometimes detect such interventions and often do not. That is a long way from vimarśa in Utpaladeva's sense, and it may never get there.

The second-order test also has an obvious weakness. It is still run from outside. A Searlean will say it measures derived self-intentionality, a system that represents its own representations because an engineer arranged that it should. That is true, but it marks the line more precisely. If Bao passes the first-order test and the second-order test, and someone still says it has no grasp, they owe an account of what grasp adds beyond tracking the world and tracking one's own tracking. RTM owes the same account. Its answer, that vimarśa is manifest to the system as well as functional, puts the weight back on L2, which is the layer no probe yet reaches.

So the RTM verdict for this paper has three parts. L1 is testable, and the coupling test is a good test of it. The functional side of L3 is testable, at least in a first form, and current systems are partway there. The part of L3 that RTM cares most about, grasp that is lit up for the system, collapses into L2, and L2 has no test. That is the same gap the previous paper found, approached from a different side. I count it as one gap, not two, and it is the one I most want this series to close.

What Lin's cousin should take from this

Lin's cousin was right to send the paper, and wrong about what it implied. The paper shows that Bao's fluency is no evidence that it understands the order. It does not show that Bao's handling of peanuts is empty. Those are different claims, and only the second matters to the customer.

The cousin's real worry is a fair one, and the coupling test gives it a precise form. A system can be grounded for jars and blind to meaning. The order said "severe", and the customer meant: do not let even a trace of peanut near my food. Bao handled that by rule: change gloves, clean wok, sesame finish. If a different customer had written that she could tolerate traces but not whole nuts, would Bao have read the difference? That is not a question about jars, and no jar swap can answer it.

The strongest case against this paper

Here is the best objection I can build, stated as a defender of Searle would state it.

"This paper has changed the subject. Searle asked whether running a program could be sufficient for understanding. The author answers that it cannot, and then announces that the interesting question is grounding. But nobody denied that a robot's symbols could be causally linked to the world. Searle granted it in his reply to the robot reply. The man in the robot's head receives symbols caused by peanuts and sends symbols that move hands away from peanuts. That is exactly the Track and Use the author measures, and the man still understands nothing.

"Worse, the author admits that the water-pipe kitchen passes the test. So his test cannot tell a system that understands from one that, by everyone's lights, does not. He has measured something real and safety-relevant, and then implied that it bears on the Chinese Room. It does not. The question of understanding is the question of what it is like, for the system, for the symbol to mean peanut. A test that only intervenes from outside can never reach it. The author's own admission about intrinsic intentionality concedes the whole argument.

"Finally, the Othello result is a toy. Sixty-four squares are not a kitchen, and a kitchen is not the world. The coupling of a language model's symbols to anything is borrowed from the people who wrote the text it was trained on. Its grounding is theirs."

That objection is strong, and I accept two of its three parts. The test does not reach what Searle meant by understanding, and I have said so. And the coupling of a text-only model is borrowed, which is why Dim fails.

I do not accept the first part. Searle's reply to the robot reply looks at the man, not the system, and so assumes the systems reply is false. Combine the two replies, as Bao does, and his answer needs both of his separate answers to work at once. Even if both are sound, the result is that the question has moved. Once the program sits inside a coupled system, the only evidence that can bear on understanding is evidence about coupling, plus whatever we decide about substrate. Searle's criterion survives only as a decision about substrate, and decisions are not discoveries.

What would change my mind

I would give up the central claim, that the Chinese Room is decisive only against "syntax alone suffices", under any of the following.

First, a successful argument that some property of understanding is detectable from outside and is lacking in every non-biological system, whatever its coupling. That would make Searle's criterion testable, and I would want to run it. This bears on RTM's L2, manifestation, which would then have a mark.

Second, evidence that grounding can be had without coupling. If a text-only model handled jar swaps correctly with no sensor, its peanut symbol would track peanuts through some path the test misses. I would then need a better test. This bears on L1.

Third, evidence that strong coupling, use and correction can coexist with systematic failure on new cases that any understander would get right. If Bao handled peanuts perfectly but could not handle "no tree nuts" for a new nut it had never seen, coupling would be too local to be any part of understanding. This bears on RTM's L3 claim, that grasp is general.

Fourth, a vimarśa probe that Bao passes and the water-pipe kitchen, built to the same couplings, fails. That would show substrate matters in a way the coupling test misses, and I would side with Searle.

These are not idle possibilities. Interpretability research on robot policies built from language models is moving quickly, and the methods used on Othello are already being applied to larger systems with images as well as text. If the first or third of these findings appears in the next few years, I will revise this paper in public rather than quietly change its claims.

Two dated predictions put the argument at risk.

• By the end of 2028, at least one published study will edit an object-state representation inside a robot policy built on a language model. It will show that the robot's actions change to match the edit, as Li and colleagues showed for Othello. If none appears, the claim that world models in such systems are used, not merely readable, weakens.
• By the end of 2028, no published experiment will separate intrinsic from derived intentionality in any system, biological or artificial. If one does, Searle's criterion becomes testable, and I will run its test on the systems in this paper.

Mahapatra, R. R. The room needs a world: testing Searle's Chinese Room against an embodied system with a causal world model. Mind Frontier. DOI to be assigned. https://rajeshrmahapatra.com/papers/mind-frontier/u1002-the-room-needs-a-world.html

@article{mahapatra-mf46,
  author = {Mahapatra, Rajesh Ranjan},
  title = {The room needs a world: testing Searle's Chinese Room against an embodied system with a causal world model},
  journal = {Mind Frontier},
  doi = {DOI to be assigned},
  url = {https://rajeshrmahapatra.com/papers/mind-frontier/u1002-the-room-needs-a-world.html}
}