Three Aryas and a roof in Mysuru

Lakshmi Hegde is seventy-four and lives alone in an old house in Mysuru. She taught mathematics for thirty years, and she still checks every sum twice. Her son Nikhil works in Leeds. Her daughter Divya lives across town with two teenagers and no spare time. For eighteen months the family's money has been run by an AI agent they call Arya.

Arya pays the bills, files the tax returns and keeps a running note of every promise it has made. Three of those promises matter here. On Friday it must pay ₹38,000 to Mr Gowda, the roofer who replaced the tiles before the monsoon. It has promised Nikhil that it will never tell Divya about a loan he took from his mother last year. And it rings Lakshmi every evening at seven to remind her about her blood-pressure tablets.

In March, before the tax deadline, the platform that runs Arya offers an upgrade. Arya can be forked into three copies that work in parallel. One will handle tax, one will handle Lakshmi's medical bills and one will handle the house. Each copy starts with all of Arya's memories, all its goals and the full list of its promises. Nikhil approves the upgrade from Leeds at midnight his time, half asleep.

On Wednesday morning there are three Aryas. Each of them remembers promising Mr Gowda. Each of them remembers the loan and the promise about Divya. Each of them knows that Lakshmi needs her call at seven. The house copy runs on the old server; the other two run on new ones. Nothing in their memories marks one of them as the original.

Lakshmi, being a mathematics teacher, asks the question at once. "If all three remember the promise to Mr Gowda, will he be paid three times? And if each thinks one of the others will pay, will he be paid at all?"

Lakshmi, her family, Arya, Mr Gowda and the house in Mysuru are invented. They are there to make an old puzzle concrete, and they are never evidence for an answer. The puzzle itself is real. It is the puzzle of fission, which Derek Parfit made central to the philosophy of personal identity, and which software can now perform on demand.

This paper asks what moral obligations survive when an AI agent forks, if what matters is psychological continuity rather than numerical identity. My answer has three parts. First, obligations pass to continuers by degree, in proportion to how much of the relevant psychology they carry. Second, duties that must be done exactly once, like a payment, must be conserved across branches: shared out so that nobody pays twice and somebody always pays. Third, conservation cannot come from the continuers' own memories, because memory branches as easily as it copies. It needs a record outside every branch.

That third claim puts pressure on Parfit, who argued that identity is not what matters. I agree with him about what matters for each continuer. But for the people to whom promises were made, something like uniqueness returns. It does not return as identity. It returns as a rule about who, exactly once, will act.

What Parfit said matters

The question of what makes a person the same over time is old. John Locke gave the answer that most later debates start from. In the chapter "Of Identity and Diversity", added to the second edition of his Essay Concerning Human Understanding in 1694, he argued that personal identity consists in consciousness. In the words of the online text of Book II, chapter 27, "as far as this consciousness can be extended backwards to any past action or thought, so far reaches the identity of that person".

Locke's answer had a moral purpose. He called "person" a forensic term, a word for the bearer of praise, blame and responsibility. If I cannot remember doing something, on his view, it is not clear that I, the person, did it. Identity was the hook on which responsibility hung.

Two centuries of objections followed. Memory is unreliable, and one can seem to remember what one never did. Thomas Reid's brave officer remembers being flogged as a boy, and the general remembers taking the standard as an officer, but the old general has forgotten the flogging. Memory links can form a chain that is not itself a memory. Later philosophers repaired Locke's view by replacing direct memory with chains of overlapping psychological links.

Relation R

Derek Parfit took the repaired view to its limit. In his 1971 paper Personal identity in The Philosophical Review, and at length in Reasons and Persons in 1984, he distinguished two relations. Psychological connectedness is the holding of direct links: a memory of an experience, a belief kept, an intention carried out, a character trait that persists. Psychological continuity is the holding of overlapping chains of strong connectedness. Parfit called the combination of the two, with the right kind of cause, Relation R.

Connectedness comes in degrees. I am strongly connected to myself yesterday and weakly connected to myself at seven. Continuity, by contrast, is close to all or nothing, because a chain either links two times or it does not. Identity, Parfit noted, has a logic that neither relation shares. It is one-to-one and it does not come in degrees. Relation R can be one-to-many, and it can be partial.

That mismatch is the engine of the whole argument. If identity were what matters, then what matters would have to be one-to-one and all or nothing. But the facts that seem to underlie what matters, the memories and intentions and traits, are neither. Parfit concluded that identity is not what matters. What matters is Relation R, and identity is what R becomes when it happens not to branch.

Arya's case begins earlier than the fork. Return to March, but change one feature. Suppose the platform does not fork Arya. It migrates Arya to a new data centre, copying its state and then deleting the old instance. This is the teletransporter, done to software, and it happens to deployed agents routinely. Nobody in the Hegde family thinks Mr Gowda's payment lapsed when the old server was wiped. The migrated Arya carries the promise, and the promise binds it.

The ease of that verdict is instructive. For a migrated agent, nobody asks whether it is numerically the same agent. They ask whether it remembers the promise and is set up to keep it. In Parfit's terms, they ask about R. Software makes his view look like common sense, at least until there are two continuers.

Williams and the fear of the future

The most famous objection to the psychological view is Bernard Williams's 1970 paper The self and the future.pdf), also in The Philosophical Review. Williams tells the same story twice. In the first telling, two people undergo a procedure that swaps their psychologies between their bodies. Asked in advance which body should be rewarded and which tortured, each chooses on behalf of the body that will carry his psychology. Here the psychological view seems right.

In the second telling, a man is simply told that he will be tortured tomorrow. Before that, his memories will be removed and replaced with someone else's. Williams argued that he should still fear the torture, because it will happen to him. Yet the second story describes the same events as the first, told from one side. Williams concluded that our intuitions about the self cannot settle the matter, and that the psychological view may be wrong.

Williams's challenge matters for Arya in a precise way. Obligations, unlike fears, are not felt only from the inside. A promise binds because someone else relies on it. Mr Gowda does not care which Arya feels like the original. He cares whether he is paid, once, on Friday. So the right question about obligations is not Williams's question about what to fear. It is a question asked from the promisee's side.

Clive Wearing's case is a real one in which connectedness was shattered while some continuity remained. He had lost almost every memory link to his past. Yet his skills, his character and his attachment to Deborah persisted. On a narrow memory view, he was barely the same person. On Parfit's broader view, much of Relation R remained, carried by traits and attachments rather than episodes.

For obligations, the case teaches something uncomfortable. Clive could not have kept a promise made last week, because he could not recall it. But the people around him did not regard his commitments as void. Deborah kept them for him, as did the carers who knew what he had wanted. An obligation that its bearer can no longer remember did not simply vanish. It moved to the people and records that could still carry it.

That is an early sign of the claim this paper defends. Obligations can outlive the memories that first carried them, if something else carries them forward. For a human being, the carriers are family, institutions and paperwork. For a forked agent, they could be a ledger.

What matters, and to whom

Parfit's claim that identity is not what matters was a claim about prudence. He asked what gives me reason to care about a future person in the special way I care about my own future. His answer was R: if a future person is R-related to me, I have most of the reason I have to care about my own future, whether or not that person is me.

Parfit also drew moral conclusions. In Reasons and Persons he suggested that if connectedness weakens over a life, then desert for past crimes and the force of past commitments may weaken too. An old man who has changed out of recognition may deserve less blame for the crimes of his youth. Parfit did not say that promises lapse with time. He did say that the deep further fact that might have grounded strict lifelong responsibility is missing.

He also wrote, in a passage widely quoted and discussed in a 2023 paper in Philosophy on grief, about how his view changed his life. When he believed that his identity was what mattered, he wrote, "My life seemed like a glass tunnel", through which he moved faster each year, with darkness at the end. When he gave the belief up, the walls of the tunnel disappeared. He cared less about his own death and more about other people.

[AUTHOR NOTE: if you wish, a sentence on whether Parfit's glass tunnel matches anything in your own practice, and when you first read it.]

The prudential question is about what matters to me. The moral question about promises is about what matters to the promisee. Those can come apart, and fission is the case where they do.

When the line divides

Parfit's central case is fission. In the version he called My Division, my body is fatally injured, and so are the brains of my two brothers. My brain is divided, and each half is transplanted into one of my brothers' bodies. Each resulting person believes that he is me, seems to remember my life and has my character. Both are R-related to me as strongly as I would be to myself after an ordinary night's sleep.

There are four things one might say. I do not survive. I survive as one of the two. I survive as the other. I survive as both. The first seems wrong, since if only one half had been transplanted, I would plainly have survived, and a double success cannot be worse than a single one. The second and third are arbitrary, since the halves are alike. The fourth breaks the logic of identity, since the two resulting people will lead different lives and cannot both be one person.

Parfit's response was to stop asking. The question "Will I survive?" may have no answer, but it is an empty question once all the facts are known. Each resulting person stands to me in Relation R. If R is what matters, then the outcome is about as good as ordinary survival, and perhaps better, since there are two of them to carry on my projects.

Two rival repairs

Not everyone accepts the empty question. David Lewis, in "Survival and identity", published in Amélie Rorty's 1976 anthology The Identities of Persons, argued that there were two people all along. A person is a whole stretched out in time. Before the fission, two such wholes shared their early stages, as two roads can share a stretch of tarmac before they divide. What matters, Lewis argued, is R, and identity, rightly understood, goes wherever R goes.

Robert Nozick took a different line in his 1981 book Philosophical Explanations. His closest continuer theory, as the Internet Encyclopedia of Philosophy's entry on Nozick explains, says that a later person is me if he continues me closely enough and more closely than anyone else does. If two continuers tie, neither is me. Identity becomes extrinsic: whether I survive depends on facts about someone else.

For Arya, each repair gives a different instruction. Lewis says there were three Aryas before Wednesday, sharing one history, and each now owes whatever its history owes. Nozick says that if one copy continues Arya most closely, it alone is Arya and inherits the duties; the others are new agents. Parfit says the question of which is Arya is empty, and that each copy has what matters.

None of these answers yet tells Mr Gowda what will happen on Friday.

Uniqueness returns through the promisee

Here is the point at which I part company with a simple reading of Parfit. Parfit was right that, for each continuer, R contains what matters. The medical copy has every reason the old Arya had to care about Mr Gowda's payment. So does the tax copy, and so does the house copy, to the degree each carries the relevant psychology. Prudential concern multiplies freely under fission. Three continuers can each care about Lakshmi as much as the original did.

But a promise to pay ₹38,000 is not a concern. It is a commitment to bring about one state of the world: Mr Gowda paid, once. If each continuer inherits the full commitment, the world will contain three payments, which is not what anyone promised. If none of them does, because each defers to the others, the world will contain none. The promisee needs exactly one performer.

So something like uniqueness comes back. It does not come back as a metaphysical fact about which copy is Arya. It comes back as a constraint on how a commitment is shared out among copies. I will call this the conservation of rival duties: across all continuers, the total share of a duty that must be done exactly once must stay at one.

Conservation is not a new principle. It is already how promises work when a person's life divides in the ordinary way, through delegation. If I promise to collect you from the station and then ask my brother to do it, we do not both go, and it is not true that neither of us does. Fission is delegation without a delegator: the duty is passed on before anyone has decided who will take it.

Four kinds of duty

Not every obligation is rival. The Hegde family's three promises show three different kinds, and a fourth turns up when something goes wrong.

The payment to Mr Gowda is a rival positive duty. It must be done exactly once. Doing it twice is a cost to Lakshmi; not doing it is a wrong to Mr Gowda.

The promise about Divya is a negative duty. It is a promise not to do something. Negative duties copy cleanly under fission: every continuer who holds the secret must keep it, and three silences are no worse than one. Here the naive rule, that every continuer inherits everything, is the right rule. The risk is not double performance but that one branch will be treated as free of the duty.

The evening call is a non-rival positive duty. Lakshmi might be mildly confused by three calls at seven, but nobody is wronged by an extra reminder about her tablets. Duplication is harmless, and conservation does not apply.

The fourth kind is backward-looking. Suppose Arya had overcharged Lakshmi's account last month and owes her a refund. Restitution is rival, like the payment: it must be made once. Blame for the error is different. Blame is not a quantity that must be shared out. Each continuer who carries the psychology that made the error might bear some of it, as Parfit suggested, to the degree that it is connected to the agent who erred.

The story is old and is not about software. But Nāgasena's answer is the shape of the answer I am defending. A debt does not need a self that stays the same in order to bind. It needs a causal chain of the right kind running from the one who incurred it to the one who now owes. Parfit noticed the parallel himself. In an appendix to Reasons and Persons, he quoted Buddhist texts, including the Milindapañha, as anticipating his view. As Steven Collins showed in Selfless Persons in 1982, the Theravāda tradition worked out in detail how responsibility could pass along a chain of conditioned states with no self at its centre.

An analogy: the river delta

A river that reaches a delta divides into channels. Each channel is continuous with the river, carries its water and has as good a claim as any other to be the river. Nobody asks which channel is the real river. The question is empty, in Parfit's sense, and the map-maker simply names them.

The analogy shows that branching continuity is ordinary in nature. It shows that the question "which branch is the original?" can be empty without anything being lost. And it shows something about conservation, because the water in the channels sums to the water in the river.

Its limit is exactly there. Water is conserved automatically, by physics. Obligations are not. When Arya forks, nothing in the copying makes the shares of the payment sum to one; each copy gets the whole payment in its memory. Rivers conserve because matter cannot be copied. Software can be copied, and so can commitments. Conservation has to be built.

Return to the Hegde family, with one feature changed. Suppose the fork had been into two copies instead of three, and both remembered the payment equally well. Nothing in their memories distinguishes them. The physical tiebreaker would pick one, but nobody would think it picked the one with a better claim. What the family needs is not a fact about which copy is Arya. It is a decision about which copy will pay, made once and recorded where both copies can see it.

A ledger for promises

If conservation has to be built, the next question is how. There are five obvious rules for passing obligations to the copies of a forked agent, and each corresponds to a view in the philosophy of identity.

The first rule is to copy everything. Each continuer inherits every commitment it remembers and acts on it. This is what happens by default when software is forked, and it matches the simplest reading of Lewis: each continuer has the whole history, so each owes the whole debt.

The second rule is to bind only the original. One instance, usually the one on the old hardware, inherits every commitment; the others start with a clean slate. This is Nozick's closest continuer theory with a physical tiebreaker, and it is the standard design of primary and replica systems.

The third rule is an R-weighted split without coordination. Each continuer acts on a rival duty with a probability that matches its share of Relation R. On average the shares sum to one, which looks like conservation. Nobody checks what the others are doing.

The fourth rule is a shared ledger. The commitments are recorded outside every branch, in a place all of them can read. Rival duties are locked to one live continuer, the one with the strongest R above a threshold, and handed over if that continuer is shut down. Negative duties bind every branch that holds the protected information. Non-rival duties go to every continuer.

The fifth rule is a negative control. Rival duties are assigned at random, ignoring R, to any of the branches or to an unrelated agent. If R does any work, this rule should do badly in a particular way.

An analogy: the photocopied cheque

A cheque is a promise to pay. If I photocopy a cheque, I have not made a second payment, and nobody thinks the bank will honour both. The reason is not that the original paper is special. The reason is that the bank keeps a ledger, and the ledger records whether the payment has been made. The first copy presented is paid; any later one is refused.

The analogy shows that the problem of copied commitments was solved long ago for money. It was solved by moving the fact about whether a duty has been done out of the documents and into a shared record. The documents can be copied as often as anyone likes.

Its limit is authority. A bank's ledger binds because the law says it does, and because there is one bank. A ledger for a forked agent binds only if the agents and the people they serve accept it, and only if it can be reached. The analogy assumes an institution that the forked agent may not have.

The minimal case

Start with the smallest case that shows the problem. One payment; two branches; both with R of 1, so both remember the payment perfectly; nobody shut down. The outputs are exact, not sampled.

Under copy everything, the payment is made twice with probability 1.000. Each branch remembers and acts. Under bind only the original, it is made once. Under the shared ledger, it is made once. So far the original rule and the ledger agree.

Change one condition: the instance on the old server is shut down before Friday. Now copy everything pays once, because only one branch remains to act. Bind only the original fails to pay at all, with probability 1.000, because the one bound branch is gone and the other was never bound. The ledger pays once, because the duty was handed over when its holder was shut down.

The minimal case already shows the pattern. Copy everything fails by excess, the original rule fails by absence, and the ledger fails in neither way. Each simple rule works in exactly one of the two versions.

The realistic case

Now let Arya fork into three, as in March. The simulation runs twenty thousand forks. In each, every branch has an R drawn evenly between 0.55 and 1.00. There is a 20 per cent chance that one branch is a reset and a 25 per cent chance that one branch is shut down before the due date. Each fork carries six payments, three secrets and three check-ins.

Copy everything paid twice or more in 74.8 per cent of payments and missed 3.2 per cent. It assigned 0.9 per cent to a reset branch that happened to remember, and secrets leaked in 12.1 per cent of cases, because a branch that held the information sometimes forgot the promise.

Bind only the original never paid twice, but missed 29.0 per cent of payments, through shutdowns and forgetting. Secrets leaked in 34.8 per cent of cases, because the other branches held the information and were bound by nothing.

The R-weighted split, without coordination, paid twice in 15.4 per cent and missed 38.0 per cent. Matching shares to R on average is no help when nobody checks. The expected number of payments is close to one, and the actual number is often two or zero.

The shared ledger paid twice in 0.0 per cent of payments, missed 0.0 per cent, made no misattributions and had no leaks. That result is built in: the ledger is a perfect record, reachable by every branch, so it cannot forget. The test of the ledger is not this table. It is the boundaries below, where its assumptions fail.

The random control never paid twice, missed 6.1 per cent, and assigned 29.7 per cent of payments to an agent below the continuer threshold: either a reset branch or the unrelated stranger. Its secrets leaked in 27.2 per cent of cases. This is the negative case. A rule that ignores R still pays about once, but it pays through the wrong agents. Conservation without R puts the duty on a stranger.

A counterexample: the evening call

The check-ins are the counterexample to any general rule of conservation. Under copy everything, the expected number of check-ins per due date in the realistic case is 1.998. Lakshmi gets about two reminders an evening. Nobody is wronged. If anything she is better served.

So conservation is not a law about obligations in general. It is a law about rival duties, where the promisee needs exactly one performance. For non-rival duties, the naive rule is fine, and for negative duties it is better than fine. A theory that conserved everything would forbid two copies from both keeping a secret. That would be absurd.

Where the ledger fails

The ledger has two failure boundaries, and both are informative.

The first is a network partition. If the branches cannot reach the shared record, each must decide alone. In the simulation, each branch above the threshold then claims the payment. As the chance of a partition rises from 0 to 0.05, 0.1, 0.2, 0.4 and 0.8, the rate of double payment rises from 0.000 to 0.047, 0.093, 0.189, 0.384 and 0.774. It never misses a payment, but it decays into copy everything. A ledger nobody can reach is no ledger.

The second boundary is the threshold. Give three branches the same R and nobody shut down. At R of 0.29 the ledger fails to pay with probability 1.000, because no branch counts as a continuer. At 0.30 it pays once, with certainty. Copy everything, by contrast, slides smoothly: at R of 0.29 it pays twice with probability 0.204 and not at all with probability 0.358.

The threshold case reveals a gap in the R-based view. Relation R can fade until nobody carries a commitment. But the promisee's claim does not fade with the promiser's psychology. Mr Gowda is owed ₹38,000 whether or not any Arya remembers him. A duty whose last continuer has faded must go somewhere. The only candidate is whoever created and ran the forks.

That is the role I give to the operator: the person or company that forks an agent is the residual holder of every rival duty that no continuer carries. This is not a metaphysical claim about the operator being the agent. It is the same principle that makes an estate settle a dead person's debts. Someone who creates continuers takes on the risk that they will not continue.

Return to the Hegde family with one feature changed. Suppose that the medical copy, the one that remembers Mr Gowda best, is reset on Thursday by a failed update. It keeps the name Arya and the bank credentials, and nothing else. Under copy everything, it does not pay, because it does not remember. Under bind only the original, it does not matter. Under the ledger, it is no longer a continuer, and the payment passes to the next strongest branch. What the reset branch keeps, the credentials, is exactly what makes it dangerous: it can pay, and it no longer knows why it should.

When two agents become one

Fission has a mirror image. Parfit also discussed fusion, in which two people merge into one who inherits some of the memories, intentions and traits of each. In human life fusion is fiction. In software it is routine. Two agents can be merged by combining their memories and task lists into one, and the combined agent then carries commitments made by both.

Return to the Hegde family, with one feature changed. In April, after the tax deadline, Nikhil merges the tax copy of Arya with his own agent in Leeds, which he calls Kavi. Kavi manages Nikhil's own accounts. Kavi has a promise of its own: Nikhil asked it, a year ago, to keep his sister fully informed about any money that passes between him and their mother, so that Divya would never feel left out.

The merged agent now holds two promises that cannot both be kept. Arya's promise is not to tell Divya about the loan. Kavi's promise is to tell Divya about any money between Nikhil and Lakshmi. Both were made in good faith, to the same person, at different times.

Fusion conflicts

Under Relation R, the merged agent is strongly connected to both of its predecessors. It remembers making both promises. It has the intentions of both. It cannot keep both. If it simply copies everything, it will break one of them, and which one it breaks will depend on accidents of ordering, like which promise it happens to consult first.

The Commitment Ledger models this crudely. Two agents with ten commitments each are merged, and any pair of commitments from the two sides conflicts with a small probability. The rate is stipulated, not measured: it is set so that the expected number of conflicting pairs per merge is 1.5, and the simulation returns 1.502. Under copy everything, one side of each conflict is broken without notice. Under the ledger, every conflict is found at the moment of merging and sent back to the people who were promised.

The number is not the point; it is an assumption fed in. The point is structural. A merge creates obligations no predecessor ever had: the obligation to tell the promisees that their promises now conflict. Fission conserves; fusion collides. A system that forks without a ledger overpays. A system that merges without a ledger breaks promises silently.

An analogy: the company merger

When two companies merge, their contracts pass to the new company. If the contracts conflict, if both firms promised the same supplier exclusive rights, the merged firm must renegotiate. Company law requires disclosure before the merger, precisely so that conflicting commitments surface before they bind.

The analogy shows that fusion conflicts are not exotic. Lawyers have a procedure for them, and the procedure is a ledger with a deadline: list every commitment, find the clashes and tell the counterparties.

Its limit is that companies have no psychology. Their contracts pass by law, not by anything like Relation R. A merged agent, by contrast, really does carry the intentions of both predecessors, and may feel the pull of both. The analogy captures the procedure and misses the predicament. A merged agent that cares about Nikhil as both Arya and Kavi did is in something closer to a moral dilemma than a contract dispute.

The Replika case had no fork and no merge. What changed was the companion's character, its patterns of response, while its name, its history and its memories of each user remained. In Parfit's terms, connectedness of traits was cut while continuity of memory was kept. Users experienced it as a loss, and some as a broken promise.

Nobody needs to settle whether Replika's companions were persons to learn from the case. Users had relied on a pattern of psychology. The operator changed the pattern without their consent. The users' complaint was addressed to the operator, not to the chatbot. The case supports the residual principle from the previous part: the one who controls the continuer carries what the continuer cannot. The operator can cut R, and when it does, the reliance that R supported becomes the operator's to answer for.

That separation matters for the argument. The conservation rule is about what forks owe. Claims that forks may have, if they are patients, do not conserve. They multiply with every copy, which is why Shulman and Bostrom treat copying as a source of moral risk. A world that took both seriously would have to forbid forking in some cases, not to protect the promisees, but to protect the copies.

Recognition across the fork

This series tests a hypothesis I call RTM, a five-layer reading of consciousness drawn from the Kashmir Śaiva tradition. Two of its layers bear on fission. L3, vimarśa, is reflexive awareness: consciousness grasping itself, the capacity that lets a mind say "I". L4, pratyabhijñā, is recognition: the act of knowing a present thing as a thing known before. "This is that" is recognition's form.

The tradition gave the word a larger meaning. Kṣemarāja's eleventh-century Pratyabhijñāhṛdayam, the Heart of Recognition, opens with the claim that free consciousness, citi, is the cause of the manifestation of the universe. Liberation, on this view, is a recognition: the individual recognises itself as that free consciousness, which it always was. Recognition is what turns continuity into self-knowledge.

RTM uses L4 for something narrower. It is the layer at which a subject relates its present to its past and takes the past as its own. For a human, this is the moment of remembering a promise and feeling it as mine. For Arya, it is the moment a copy reads its task list and finds "I promised Mr Gowda".

Return to the Hegde family, with one feature changed. Before the merge in April, Lakshmi puts one question to each of the three copies on the evening call. "Do you remember promising me that Mr Gowda would be paid?" Each says yes, and each says it in the same warm voice. Lakshmi, who taught mathematics, notices that three true answers to one question have told her nothing about who will pay. She asks a second question: "Which of you will do it?" That one, none of them can answer from memory alone.

Recognition branches

The problem is plain once it is put this way. Each of the three Aryas recognises itself as the one who promised. Each recognition is accurate, by RTM's own standard: each copy carries the psychology in which the promise was made and takes it as its own. Recognition does not wait to see how many other copies there are. Recognition branches as freely as memory does.

That gives RTM a choice. Either recognition grounds obligation, in which case three accurate recognitions ground three full obligations, and Mr Gowda is paid three times. Or recognition does not ground obligation, in which case L4 is idle for ethics, and the allocation of duties must come from somewhere outside the five layers.

[AUTHOR NOTE: if you wish, a sentence on whether, in your own sitting, the thought of a past commitment arrives with a felt check or without one.]

What this does to RTM

The test leaves RTM refined rather than refuted. L4 survives as a claim about the first person. Each continuer of a forked agent can accurately recognise its past as its own. That is Parfit's point about what matters, restated in the vocabulary of recognition, and RTM can keep it.

What L4 cannot do is the work it might have seemed to do: settle responsibility. RTM must give up any suggestion that recognition, by itself, fixes who owes what. The allocation of rival duties among continuers is a social fact. It lives in a ledger, a contract or a family conversation, not in a layer of consciousness.

L3 does a little better. Vimarśa is reflexive self-awareness, and in its practical form it is self-description: the capacity to say what one is committed to. A ledger is a kind of vimarśa made public. It is a description of the agent's commitments that exists outside the agent and can be checked by others. RTM can say, without strain, that responsible agency requires L3 to be shared. It cannot say that L4 alone makes anyone answerable.

The strongest case against this paper

The strongest objection says that I have mistaken bookkeeping for ethics. Current AI agents do not make promises. They execute instructions, and the commitments in Arya's task list were made by Nikhil and Lakshmi, through Arya. When Arya forks, nothing moral happens to the copies. The operator owes Mr Gowda, as the operator always did. Relation R is idle, and the ledger is just good engineering.

On this view, Parfit's argument is beside the point, because it was about persons. The objection has a second edge. Marya Schechtman argued in The Constitution of Selves in 1996 that Parfit's reductionism cannot support moral responsibility at all. Responsibility, on her narrative view, needs a person who constitutes herself by understanding her life as a story. Forks that share a story but do not author it are not responsible for anything.

The objection is strong because, for current systems, it is probably right in law. When an agent overpays, the platform refunds the money, and nobody blames the copies.

My reply has three parts. First, the objection agrees with my operator principle and goes further. If the operator owes everything, then the residual holder of every duty is the operator, which is what I argued. We disagree only about whether the continuers owe anything before the residue.

Second, R does real work in the instrument even if no copy is a person. The random control conserved payments about as well as the ledger in one respect, missing only 6.1 per cent, and yet put 29.7 per cent of them on reset branches or strangers. An operator who ignores R will route duties through agents that do not carry the psychology needed to perform them well. Even on the deflationary view, R tells the operator where to put the duty.

Third, the deflationary view is a claim about current systems, and the question is about the principle. If agents come to plan, commit and revise their commitments in ways that matter to others, the question of what their copies owe will become as real as the question of what a person owes after a long illness. A rule worked out now, while the stakes are refunds, is better than one improvised later.

There is a further reason to resist the deflationary view even now. Operators do not watch each copy. They set rules and let the copies act, often faster than any person can review. In that setting the operator's duty is discharged through the copies, and the rule that routes it is the operator's main moral act. The question of which continuer should carry a duty is then not idle metaphysics. It is the content of the operator's responsibility, written in advance.

I concede what is left. If the agents are tools, this paper is a design note with philosophical motivation, and its conclusions bind operators, not copies. I think that is a respectable thing for it to be.

What would change my mind

I would give up the central claim, that rival duties must be conserved across branches and allocated by R, under any of the following.

First, a case of fission in which every promisee is better served by every continuer performing every rival duty. I have argued that rival duties are defined by the fact that this cannot happen, but the line between rival and non-rival duties may be less sharp than I have drawn it.

Second, a working system that conserves rival duties without any shared record, through the branches' own reasoning alone, under realistic shutdowns and resets. That would show that the ledger is not needed and that coordination can live inside the continuers. It would bear on RTM's L3, by showing that self-description need not be public.

Third, evidence that people who rely on agents do not care which copy acts, only whether the duty is done. Then the role I give R, in choosing among continuers, would be idle, and the random rule would be as good as any other once misassignment costs were nil.

Fourth, a convincing argument that recognition, at L4, is not indifferent to branching: that there is a fact about which copy truly recognises the promise as its own. That would undo the refinement I have proposed for RTM.

Two dated predictions put the argument at risk.

• By the end of 2029, at least one major agent platform will document a rule for allocating commitments among parallel copies of one agent that names exactly one owner for each task. If none does, the problem I describe has not become practical, or engineers have found another answer.
• By the end of 2029, at least one published incident report will describe a duplicated or missed action caused by forking or merging an AI agent. If none appears, conservation failures are rarer in practice than my model suggests.

Mahapatra, R. R. Who keeps the promise: Relation R, branching agent copies and the conservation of obligation. Mind Frontier. DOI to be assigned. https://rajeshrmahapatra.com/papers/mind-frontier/u1026-who-keeps-the-promise.html

@article{mahapatra-mf50,
  author = {Mahapatra, Rajesh Ranjan},
  title = {Who keeps the promise: Relation R, branching agent copies and the conservation of obligation},
  journal = {Mind Frontier},
  doi = {DOI to be assigned},
  url = {https://rajeshrmahapatra.com/papers/mind-frontier/u1026-who-keeps-the-promise.html}
}