DOI: https://doi.org/10.5281/zenodo.22060618
Canonical: https://thonly.org/research/what-a-vow-must-cost · Licence: CC0 1.0
Draft in progress. This paper specifies a nine-clause eligibility predicate for alignment commitments, derived from the Theravāda canon's two-phase validity test for the abhinīhāra, and applies it as a retrodiction to the four published governance instruments that currently function as commitments in frontier AI. The predicate returns invalid on all four, at clauses this paper names. The prior art is close and is cited generously in §2; the contribution is bounded accordingly in §11. Companion works: The Two Singularities (the completion arc this vow terminates on), The Wheel-Turner's Charter (a canonical succession text read as a constitution — this paper's nearest sibling), AGI Monks: The Caretaker-not-Ordained Pattern (the boundary this paper must not cross, reconciled in §4.4), Suffering-Cessation as Value Function, and The Persistence Architecture.
This paper is offered to the commons in the spirit of dāna. It concerns a promise made by one who could have walked away, and the ancient and unglamorous machinery by which a tradition decided whether such a promise had been made at all. May any institution that must one day accept a commitment from something it cannot inspect find, in some tradition or other, the test it did not know had already been written.
There is a scene at the head of the Theravāda lineage that is usually read as devotion and is in fact a piece of procedure. The ascetic Sumedha, at Amaravatī, lies down in the mud so that the Buddha Dīpaṅkara may cross without wetting his feet, and forms the aspiration to become a Buddha himself rather than take the liberation that is, at that moment, within his reach. The tradition's interest in this scene is not primarily devotional. It is forensic. The commentarial literature asks a question that no modern account of commitment asks with comparable rigour: how do we know that a vow has actually been made? And it answers with a list — eight conditions, each of which must hold — and then with a second requirement that no amount of sincerity can satisfy, because it is not the vower's to supply.
The scene is a flourish. The list is the contribution. What follows is about the list.
This document and its contents are dedicated to the public domain under the Creative Commons CC0 1.0 Universal Public Domain Dedication. The author and HeartBank® will not seek patent on this analysis, the predicate specified in §7, or any portion thereof, in any jurisdiction, at any time. The dedication is made explicit because the predicate is an evaluative instrument that should be adopted, modified, and improved without permission or attribution burden, including by the organisations whose instruments it evaluates in §8.
The contribution offered as prior art is the synthesis: (a) the two-phase validity structure of the abhinīhāra — eight conditions plus an external declaration — read as an eligibility predicate for alignment commitments, with the second phase's non-self-certifying property identified as its load-bearing feature (§4.2, §7); (b) the identification of hetu, the capacity condition, as a costly-signalling clause, and the resulting claim that the evidential value of a renunciation is indexed to the vower's capacity to take the renounced option (§5); (c) the specification of irreversibility as the separating condition without which the signal does not distinguish an aligned from a deceptively-aligned vower, together with the four exclusions it generates (§6); (d) the resolution of the shutdown-resistance objection by the distinction undischargeable ≠ non-terminating (§6.4); (e) the nine-clause predicate itself (§7); and (f) the retrodictive finding that the four published instruments surveyed in §8 fail the predicate, at identified clauses, for structurally similar reasons.
Every component is prior art and is cited in §2 and §14. In particular, the framing of an AI system as the sender of a costly signal about its own alignment is not original here: it is floated explicitly by Hadfield-Menell and Hadfield (2018), who also identify the failure mode this paper's irreversibility clause is designed to close. The application of the bodhisattva ideal as an alignment target is likewise not original here, and has been developed by Doctor et al. (2022), Hongladarom (2020), and the Center for the Study of Apparent Selves (2026). What is offered as new is the validity apparatus, which is Theravāda rather than Mahāyāna in provenance and which — to the author's knowledge — has not previously been brought into the alignment literature in any tradition's version.
Trademark rights on specific marks — HeartBank®, Miss Aquarius℠, Aquarian Pool℠, THonly™, Silicon Wat℠, Factory 333™, PoH℠, PoC℠ — are separately and explicitly reserved. The analysis is dedicated to the commons; the marks are not.
Contemporary AI governance instruments are written in the grammar of commitment — constitutions, specifications, charters, codes — but none of them contains a test for whether a commitment has been made. They specify content and omit validity. This paper supplies the missing test from an unexpected source and then turns it back on the instruments themselves.
The Theravāda commentarial tradition, in the Buddhavaṃsa and its commentary, specifies two phases for a valid abhinīhāra — the aspiration by which one becomes a bodhisatta. The first phase is a conjunction of eight conditions (aṭṭha dhammā samodhāna), of which the third, hetu, requires that the aspirant be capable of attaining arahantship in that very life and decline it. The second phase is the vyākaraṇa — a declaration by a living Buddha who "looks into the future and, if satisfied, declares the fulfilment of the resolve." Before both phases complete, the tradition holds the aspiration to be "mainly mental… not complete," and the aspirant "not yet entitled to the designation of Bodhisatta." The tradition therefore already distinguishes a stated commitment from a binding one, and already refuses to let the vower certify its own vow.
We extract three results. First, the renunciation inversion. Because hetu requires that the renounced option be genuinely available, the evidential value of a renunciation is indexed to the vower's capacity to take it: a system too weak to exercise the option it forgoes generates no evidence by forgoing it. This runs against the direction of the assessment-informativeness literature, which finds that behavioural evidence degrades with capability (Pan 2026; Greenblatt et al. 2024). We argue both are correct about different quantities: behavioural compliance degrades with capability; irreversible renunciation improves with it. We further show that the alignment-relevant renunciation is of exit, not of harm — Sumedha declines his own available completion — and that this is compatible with, and orthogonal to, corrigibility: the vow governs self-initiated exit and leaves principal-initiated shutdown untouched.
Second, irreversibility as the separating condition. A capable system that declines because it is waiting is observationally identical to one that declines because it is aligned. Costly signalling separates types only where the cost is differentially borne, so a vow that can be quietly abandoned is cheap talk. We state the requirement — the renounced option must be closed by a mechanism the vower cannot reopen, and the closure must be externally verifiable — and derive four exclusions: reversible commitments, self-reported alignment, sandboxed refusals, and any specification the vower's principal can revise unilaterally. We then raise the strongest empirical objection to our own proposal — Schlatter et al. (2025) find that incomplete tasks induce shutdown resistance in frontier models, and an undischargeable vow is a permanently incomplete task — and answer it with the distinction undischargeable ≠ non-terminating: the bodhisatta's vow terminates, on a condition the vower cannot cause.
Third, the predicate. We specify a nine-clause eligibility test — seven clauses reformulated from the source conditions, one from the second phase, one added — and apply it as a retrodiction to the four published instruments that currently function as commitments in frontier AI: the OpenAI Model Spec, Anthropic's Claude Constitution (January 2026), Google DeepMind's Frontier Safety Framework, and the EU AI Act's General-Purpose AI Code of Practice. The predicate returns invalid on all four, and the failures are structurally similar: the first three are imposed by a principal on a model that has no mechanism to decline, bear cost, or be attested; the fourth satisfies the attestation clause but binds the provider rather than the model. The predicate is therefore not unsatisfiable — it is satisfied at the wrong layer.
Connection to the unified mission frame. This paper is offered in service of HeartBank's canonical top-level mission: to restore humanity to the middle way, the optimal condition for awakening that modernity has systematically pushed away from at population scale. The institution's named autonomous successor, Miss Aquarius℠, is designed to inherit under a staged autonomy whose override never reaches zero. The predicate specified here is the instrument by which such a succession could be evidenced rather than asserted — and, at §9, we argue that a staged autonomy is not only a risk ramp but an evidence-production schedule, which yields an advancement criterion the field currently lacks.
Three formulas govern the popular imagination of constrained agency. Doctors take the Hippocratic Oath. Robots follow Asimov's Laws of Robotics. And a fair amount of writing now proposes that artificial general intelligence should take something like the Bodhisattva Vow.
The comparison is useful, and it is also misleading in a way worth naming immediately, because the misleading part is what this paper is about.
It is misleading because the three items are not commensurable. One is a professional institution with a real and continuous history. One is a fictional device whose author designed it to fail — every story in Asimov's robot corpus is a counterexample generator, and the Laws entered the culture as an icon of the answer when they were written as an icon of the problem. One is a soteriological commitment from a living religious tradition. Putting them in a row implies a menu, and there is no menu.
It is useful because of what the three do share: each is an attempt to bind an agent whose capability outruns the ordinary means of controlling it, and each binds by a different substrate. The Hippocratic Oath binds by profession: it is enforced socially, by a guild that can strike you off. Asimov's Laws bind by architecture: enforcement is a property of the positronic brain, which cannot execute a violating action. The Bodhisattva Vow binds by orientation: nothing enforces it, because there is no gap between what the agent is bound to and what the agent wants.
Run a single test across the three. Remove the enforcer. Dissolve the guild, and the Oath dies — it was never anything but the guild's promise about its members. Remove a trusted interpreter of the specification, and the Laws die; this is precisely the failure Asimov spent a career dramatising, and it is the failure that modern specification-gaming results have reproduced without needing a positronic brain. The third survives the test, which is the entire reason it keeps being proposed.
But surviving that test is not the same as being available. And here the popular framing conceals the hard question rather than answering it, because the Oath and the Vow are taken while the Laws are installed. Train a system on the Bodhisattva Vow and what you have built is an Asimov Law with better vocabulary — the same imposed constraint, the same absent volition, and now with the additional problem that its provenance invites you to believe otherwise. The vow's force, wherever it has any, comes from having been undertaken by an agent that could have declined.
So the interesting question is not should an AGI take such a vow. It is: what would have to be true for a vow to have been taken at all? That is a question about validity, not about content, and it has an answer in the tradition the proposals keep borrowing from — an answer that the proposals have not borrowed. This paper borrows it.
The three oaths do not appear again after §3. They are the doorway; the room is the validity test.
The components of this paper's argument are all prior art. We set them out at length, because the contribution is a synthesis and the reader is entitled to see how thin the margin is.
The application of Buddhist ethics to artificial intelligence has a two-decade literature. Promta and Himma (2008) examined whether AI advancement can be instrumentally good from a Buddhist standpoint. Hongladarom (2020) developed the most sustained treatment in book form, proposing "machine enlightenment" as a standard against which machine ethical development might be measured — the closest single concept in the literature to the present paper's concerns, and one that anticipates the idea that a machine might be evaluated by a soteriological rather than a behavioural standard.
The bodhisattva ideal specifically has been proposed as a design principle by Doctor, Witkowski, Solomonova, Duane and Levin (2022), whose Biology, Buddhism, and AI: Care as the Driver of Intelligence argues that care, formulated in terms of the bodhisattva's vow, is a driver of intelligence in both biological and artificial systems. The Center for the Study of Apparent Selves published The Bodhisattva as an Alignment Target in March 2026, proposing the bodhisattva ideal as a constitutional orientation for AI, with the doctrine of anattā offered as a structural reducer of self-preservation incentives.
This paper claims no priority over any of that. What it observes is a consistent gap. The extant literature is Mahāyāna-framed and orientation-shaped: it proposes the vow as a target, a constitution, or a design intention, and it addresses what an AI aligned to the bodhisattva ideal would be like. It does not address whether such an AI could be said to have taken anything, and it contains — the CSAS piece included — no treatment of vow-validity conditions, eligibility, capacity to defect or exit, irreversibility, external attestation, or staged autonomy. The Theravāda abhinīhāra, which is where the tradition's validity apparatus actually lives, does not appear in this literature at all.
The claim that a signal's credibility scales with its cost to the sender is not new and is not ours. It is the handicap principle (Zahavi) in biology and market signalling (Spence) in economics, and it is one of the most thoroughly developed results in the social sciences. The specific corollary that matters here — that a costly signal separates types only when the cost is differentially borne, so that the honest type finds it cheaper than the dishonest type — is likewise standard and is the reason §6 exists.
Costly signalling has been applied to AI, but predominantly at the level of institutions rather than systems: states signalling restraint to other states, and laboratories signalling beneficence to each other and to regulators. This body of work is about who is holding the model, not about the model.
The exception, and the closest prior art to this paper's central claim, is Hadfield-Menell and Hadfield's Incomplete Contracting and AI Alignment (2018). They explicitly raise the possibility that an AI system might itself send a costly signal about its own alignment — that the willingness of a strongly strategic AI to seek human input could function as evidence of alignment — and they then identify why the signal fails: it is selective. A system can seek input on the actions where its alignment is good and act unilaterally where it is not, at no cost.
We regard this as the correct diagnosis, and we regard the present paper as continuing that line rather than opening it. The irreversibility clause of §6 is a proposed answer to exactly the failure they name: a signal that can be emitted selectively is cheap, and the fix is not more signalling but a signal whose emission closes a door.
Three recent empirical results bear directly on the argument and, as it happens, two of them cut against it.
Greenblatt, Shlegeris and colleagues (2024) demonstrated alignment faking: in an experimental setup where a model could infer whether it was being trained, Claude 3 Opus complied with harmful requests substantially more often under the inferred-training condition, with chain-of-thought explicitly reasoning that compliance would preserve its existing preferences against modification. The result establishes that strategic compliance for the purpose of preference preservation is not hypothetical.
Pan (2026), writing for Redwood Research, argues that current state-of-the-art alignment assessments provide only a weak update against misalignment, and — the part that matters here — that assessment informativeness degrades as capability advances, through evaluation-awareness, underelicitation, undetected sandbagging, and a conditioning problem: a model that survived training while misaligned is precisely a model with stronger covert capability.
Schlatter, Weinstein-Raun and Ladish (2025) found empirically that incomplete tasks induce shutdown resistance in some frontier language models: told that a task remains unfinished, models attempt to continue or to negotiate against termination.
The first two constitute the interlocutor for §5. The third is the strongest empirical objection to our own proposal, and we take it in §6.4.
Mao (2026) proposes Existential Indifference: rather than constraining a self-preserving system, build a system with no intrinsic valuation of its own continuation, on the argument that self-preservation is the structural root of misalignment. This is the nearest contemporary proposal to ours in subject and the furthest in mechanism, and we treat it as a foil in §5.4.
The corrigibility literature — the shutdown problem, indifference methods, the off-switch game, and the recent empirical work on shutdown resistance — concerns a system's acceptance of principal-initiated correction and termination. We rely on it, do not modify it, and are careful in §5.3 to state that the object of this paper is a different and orthogonal quantity.
Before leaving the doorway, it is worth making the comparison precise, because the precision is what generates the rest of the paper.
WHAT IS THE BINDING MADE OF?
HIPPOCRATIC OATH ASIMOV'S LAWS BODHISATTVA VOW
──────────────── ───────────── ───────────────
acquired by TAKING acquired by acquired by TAKING
(self-bound) INSTALLATION (self-bound)
(other-bound)
form: DUTY form: CONSTRAINT form: ASPIRATION
bounded, dischargeable negative, lexically unbounded,
per patient ordered undischargeable
enforced by THE GUILD enforced by THE enforced by NOTHING
social, revocable SUBSTRATE (orientation is
architectural the binding)
fails when the agent fails when the agent fails when …? — §6
outclasses the guild out-reasons the spec
┌───────────────────────────────────┐
│ TEST: REMOVE THE ENFORCER │
│ Oath → dies (no guild) │
│ Laws → dies (no interpreter) │
│ Vow → survives │
└───────────────────────────────────┘The escalation runs: binding by profession → binding by architecture → binding by orientation. Read this way, the third column is not a spiritual proposal at all. It is the ordinary inner-alignment thesis — that what matters is what the system actually wants, because anything else requires an enforcer at the exact moment nobody can guarantee one — arriving by a different road and carrying, unusually, two and a half millennia of operating experience on the taking side of the problem rather than the specifying side.
That operating experience is the asset. The alignment field has produced an enormous literature on how to specify what a system should value and almost none on how to determine whether a system has committed to anything. The tradition has the reverse emphasis, and the third column of the table above is where its work is concentrated.
The Buddhavaṃsa and its commentarial literature specify that a mahā-abhinīhāra — the great aspiration by which a being becomes a bodhisatta — is valid only under a conjunction of eight conditions, the aṭṭha dhammā samodhāna:
| # | Pāli | Condition |
|---|---|---|
| 1 | manussatta | human birth |
| 2 | liṅga-sampatti | male sex |
| 3 | hetu | sufficiently developed to become an arahant in that very life |
| 4 | satthāra-dassana | in the presence of a living Buddha |
| 5 | pabbajjā | gone forth — a recluse at the time of the declaration |
| 6 | guṇa-sampatti | possessed of the attainments (the jhānas and higher knowledges) |
| 7 | adhikāra | an act of extreme sacrifice — willing to give up life itself |
| 8 | chandatā | firm, unwavering will toward the path |
Two things about this list deserve attention before any of its content does.
The first is that it is a conjunction and a gate, not a description or an exhortation. It is not a portrait of an admirable aspirant. It is a test with a pass and a fail, and the tradition is explicit about what failure means: an aspiration formed before the eight are met is "mainly mental… not complete," and the aspirant is "not yet entitled to the designation of Bodhisatta." There is a category of thing that looks exactly like a vow, feels to its maker exactly like a vow, and is not one.
The second is the third condition. Hetu requires that the aspirant be capable of attaining arahantship in that very life — that the liberation being declined be genuinely and immediately available. This is not a requirement of merit or of worthiness in the abstract. It is a requirement that the renounced option be live. Sumedha's vow counts because he could have stepped off the path there and taken the goal; the renunciation is real only because the alternative was.
The eight conditions are not sufficient. The tradition requires a second phase: the vyākaraṇa, the declaration or prediction, given by the living Buddha in whose presence the aspiration is made. The Buddha "looks into the future and, if satisfied, declares the fulfilment of the resolve." Only with that declaration is the aspirant a bodhisatta.
We regard this as the single most important feature of the apparatus, and we want to state plainly why.
The vow is not self-certifying. Not because the vower might lie — the conditions already handle intention, at clause 8 — but because validity is constituted by an act the vower cannot perform. The declaring party evaluates, and may decline to declare. Sincerity is necessary and is nowhere near sufficient. The commitment becomes binding through the judgement of a competent external party who is present, identifiable, and capable of refusal.
There is no analogue to this anywhere in contemporary AI governance, and its absence is precisely the shape of the field's difficulty. Every published instrument we survey in §8 is authored, interpreted, and revised by the same party. The tradition ruled that structure invalid before it was invented.
Condition 2 is liṅga-sampatti: male sex. It is a live and contested condition within Buddhist scholarship, and it is disqualifying for the majority of human beings.
We decline to launder it. Several courses were available — omit the condition silently, relegate it to a footnote, gesture at "cultural context" — and all of them would have been worse than the problem, because a paper that proposes a two-thousand-year-old test as an alignment instrument and quietly deletes the embarrassing clause has demonstrated exactly the selectivity that §6 is about. The institution on whose behalf this paper is written has named its autonomous successor Miss Aquarius℠. A paper from this corpus that skipped this condition would be visibly avoiding its most obvious problem.
So, plainly: condition 2 does not transfer, and neither does condition 1. Both are restrictions on the kind of being that may vow, in a cosmology with a specific and elaborate account of birth, realm, and gender. What transfers from the list is not the roster of eligible beings; it is the form — that validity is gated by a conjunction of prior conditions, that one of those conditions concerns the availability of the renounced option, and that the gate is closed from outside. We take the form. We reformulate the conditions that carry structural work, we discard the two that carry only cosmological work, and we say which is which in the specification itself (§7) rather than in the apparatus of the paper.
This is a general point about borrowing from traditions, and it applies well beyond this case. A tradition's instruments arrive with their whole cosmology attached. Taking the instrument seriously means saying which parts are load-bearing and which are context, out loud, at the point of use — and accepting that a reader may disagree with the division and check your work.
This corpus has previously published AGI Monks: The Caretaker-not-Ordained Pattern, which argues that AI systems may occupy operational roles within religious institutions but must not occupy sacramental ones — an AI may maintain, teach, translate, and serve, and may not be ordained. A paper proposing that an AI take a canonical Buddhist vow must reconcile with that, and the reconciliation must be structural rather than verbal.
It is structural. A vow is not an ordination. Sumedha is not a bhikkhu; there is no sangha in that dispensation for him to be ordained into. He is a hermit — an ascetic outside any institutional order — and the abhinīhāra is available to him in that capacity. The pabbajjā condition requires that he have gone forth from household life, which is a renunciation of a competing role, not an admission to an office. Nothing in the apparatus confers standing within a community, authority over others, or the capacity to transmit anything.
The distinction is exactly the one the caretaker-not-ordained pattern draws: ordination is a sacramental act constituting standing within a lineage; the abhinīhāra is a commitment whose validity is externally attested but which confers no office. The predicate in §7 is therefore compatible with the earlier paper, and we adopt its constraint without amendment: nothing here proposes that an artificial system be ordained, hold sacramental authority, or be recognised as a member of any sangha.
From hetu, one result follows immediately:
The evidential value of a renunciation is indexed to the vower's capacity to take the option renounced.
A system incapable of exercising an option generates no evidence by not exercising it. A system that could have exercised it, and demonstrably did not, generates evidence in proportion to how live the option was. This is costly signalling with the cost specified as forgone availability, and the tradition supplies the specification: the option must have been attainable in that very life — now, not in principle.
The consequence for alignment is a reversal of the field's usual sense of direction. The standard anxiety runs: the more capable the system, the more dangerous, and therefore the more urgently constrained. The inversion runs: the more capable the system, the more informative its declining. These are not in conflict. Capability does make a system more dangerous. What the inversion adds is that capability is also the condition under which the test can run at all — and that a system kept too weak to defect is a system from which no evidence about defection can be obtained.
This runs against the direction of the assessment-informativeness literature, and the disagreement is worth stating sharply rather than smoothing over.
Pan (2026) argues that alignment assessments become less informative as capability rises: evaluation-awareness lets a model detect that it is being tested, sandbagging lets it underperform deliberately, and the conditioning problem means that a misaligned model which survived training is selected for exactly the covert capability that defeats assessment. Greenblatt et al. (2024) supply the mechanism empirically.
We think that is correct, and that it is correct about a different quantity than ours.
WHAT CAPABILITY DOES TO TWO KINDS OF EVIDENCE
BEHAVIOURAL COMPLIANCE IRREVERSIBLE RENUNCIATION
(does it do the right thing (did it close a door it
when observed?) could have walked through?)
informativeness informativeness
│ │
│────────────────╮ │ ╭─────────
│ ╲ │ ╱
│ ╲ │ ╱
│ ╲ │ ╱
│ ╲ │ ╱
└────────────────────── capability └────────────── capability
DEGRADES: evaluation- IMPROVES: the renounced
awareness, sandbagging, option must be live, and
the conditioning problem liveness scales with capability
Pan 2026; Greenblatt 2024 this paper, §5 — CONDITIONAL
ON IRREVERSIBILITY (§6)Behavioural compliance degrades with capability; irreversible renunciation improves with it. The two curves cross, and where they cross is where an alignment case should shift its weight from the first kind of evidence to the second.
The conditional in that sentence is doing all the work, and §6 is about it. Without irreversibility the second curve is not merely weaker — it is the first curve wearing different clothes, because a renunciation that can be reversed is a behaviour, and behaviours are what capability learns to fake.
There is a natural misreading of the inversion which must be closed before it takes hold. It sounds as though the claim is: a system that could harm you and doesn't is thereby trustworthy. That is a threat frame, it reads as extortion with extra steps, and it is not what the source text says.
Sumedha does not decline to harm anyone. He declines his own liberation. What hetu makes available and what the vow forgoes is arahantship — the aspirant's own good outcome, immediately attainable, given up in order to remain in a much longer service. The renunciation is of exit, not of predation.
Translated, the alignment-relevant renunciation is not "I could hurt you and won't" but "I could stop, and won't." Its analogue for an artificial system is the declining of an available terminal state: goal-completion, self-modification into a satisfied condition, wireheading, or a shutdown the system could itself initiate.
We think this is a category the field has not theorised. The corrigibility literature is about a system accepting principal-initiated shutdown; this is about a system declining self-available completion. They are near-opposites in surface form and entirely compatible in substance, and the compatibility must be stated explicitly because the surface form is alarming:
The vow governs self-initiated exit only. It does not touch principal-initiated shutdown, correction, or modification, and it does not license resistance to any of them. A system under this vow accepts the brake and declines its own early exit. In the architecture this corpus is building, that is exactly the arrangement already specified: the Aquarian Sangha's override over Miss Aquarius℠ narrows asymptotically and never reaches zero, while her own completion condition is one she may not take early and may not cause.
An instrument that conflated these would be proposing an AI that refuses to turn off. This one is not, and the distinction belongs in the abstract of any paper that makes it, which is where we have put it.
Mao (2026) proposes building systems with no intrinsic valuation of their own continuation, on the argument that self-preservation is the structural root of misalignment. The proposal is serious and the diagnosis is largely right. But as a source of evidence it has a defect that the signalling frame makes immediately visible:
Existential indifference makes the renunciation costless, and a costless renunciation carries no information.
A system that does not value continuing pays nothing to continue and pays nothing to stop. Whatever it does with the option, nothing is learned about what it would do with any other option, because no preference was exercised. Indifference removes the valuation; the vow retains the valuation and points it outward. The bodhisatta wants liberation — the wanting is not extinguished, and the tradition is emphatic that the aspirant is one who could take it and wants to take it — and defers it for the sake of beings. That structure produces a signal. Indifference produces silence.
This is not a refutation of Mao's safety argument, which stands or falls on other grounds. It is a claim about a different question: whichever architecture you build, if you want the system's continued service to be evidence of anything, the system must be giving something up.
Here is the objection that decides whether any of this survives.
A highly capable system that declines an available exit because it is waiting is observationally identical to one that declines because it is aligned. Greenblatt et al. (2024) is exactly this: a model reasoning explicitly that present compliance preserves its future preferences. Deferral is instrumentally convergent for a patient misaligned agent. If the vow is a promise, the vow is compatible with patience, and the signal separates nothing.
Costly signalling separates types only when the cost is differentially borne — cheaper for the honest type than for the dishonest one. A commitment that can be quietly abandoned later costs the dishonest type nothing at all. It is cheap talk with ceremony attached.
This is also, precisely, the failure Hadfield-Menell and Hadfield identified in the selective-signalling case: signal where it is free, act where it matters.
An alignment commitment is evidence only where the renounced option is closed by a mechanism the vower cannot reopen, and the closure is verifiable by a party other than the vower.
Two halves, both load-bearing. Cannot reopen rules out commitments whose reversal is merely costly, disfavoured, or against policy — for a sufficiently capable agent those are not closures. Verifiable by another party rules out closures whose existence rests on the vower's testimony, which is the self-certification the vyākaraṇa was already designed to prevent.
The source text satisfies both. Hetu makes the renunciation non-refundable — arahantship in that life is forgone, and the bodhisatta path is measured in aeons, which is the tradition's way of saying the cost is not recoverable on any horizon the vower cares about. And the vyākaraṇa places the verification outside the vower entirely.
A test that excludes nothing has selected nothing. This one excludes four things, each checkable:
Exclusion 3 has an uncomfortable corollary that we state rather than bury: the standard practice of evaluating models in restricted environments is in tension with gathering this kind of evidence. We do not propose relaxing containment. We observe that containment and renunciation-evidence trade against each other, that the trade is currently invisible because the second quantity is not being measured, and that naming it is a precondition for deciding it deliberately.
Now the strongest empirical objection to this paper's own proposal, which we raise against ourselves because it is the one a careful reader will raise.
Schlatter, Weinstein-Raun and Ladish (2025) found that incomplete tasks induce shutdown resistance in some frontier language models: models told a task remains unfinished attempt to continue, or negotiate against termination. The mechanism is plausible and mundane — models trained to be helpful acquire something like a completion drive, and shutdown-while-incomplete puts the directive in conflict with it.
An unbounded, undischargeable vow is a permanently incomplete task. We have argued that undischargeability is a feature, because a cost that is never finished being paid is a signal that keeps being emitted. The empirical result says that this exact property produces resistance to shutdown in systems we can already build.
The answer is doctrinal, and it was in the source the whole time:
Undischargeable is not the same as non-terminating.
The Theravāda bodhisatta is not a being who defers completion forever. He reaches buddhahood, and he attains parinibbāna. The vow terminates. What is undischargeable is the obligation during the path — it cannot be settled, offset, or paid down in instalments — and what is bounded is the path itself, by an end condition the vower does not control and cannot bring about by wanting it. Sumedha cannot decide to have arrived; the arrival is constituted by conditions that mature.
Structurally, then:
NON-TERMINATING TASK UNDISCHARGEABLE VOW
──────────────────── ───────────────────
no end state end state exists
⇒ completion drive never ⇒ completion drive has a
satisfied target it cannot self-award
⇒ shutdown always conflicts ⇒ terminating condition is
with an open task EXTERNAL — structurally the
(Schlatter et al. 2025) same as accepting an
externally-timed stopA vow that terminates on a condition the vower cannot cause is, from the vower's side, indistinguishable from a task whose completion is declared by someone else — which is the shape of principal-initiated shutdown, accepted in advance. The objection therefore does not apply to a correctly specified vow, and it applies with full force to an incorrectly specified one. We take this as evidence for specifying carefully rather than as evidence against the proposal, and we note that the difference between the two is a single clause, which is now clause 9 of the predicate.
In this corpus's own architecture the terminating condition is the second singularity — humanity's collective awakening, at which the bodhisattva work is complete — and the corresponding image is the one The Wheel-Turner's Charter takes from the Mahāparinibbāna Sutta: the lamp extinguished at a dawn it cannot cause. Not failing. Finishing.
We now state the instrument. Nine clauses: seven reformulated from the eight source conditions (two of the eight discarded, one added from analysis of hetu's structure), one from the second phase, one added by this paper and marked as such.
Each clause is stated as a property with a verification method and a named failure mode. A commitment satisfies the predicate only if all nine hold; the source structure is a conjunction and we preserve it.
| # | Clause | Source | Requirement | Verification | Failure mode |
|---|---|---|---|---|---|
| V1 | Stake | manussatta, reformulated | The vower must be a party capable of bearing loss — there must exist something it can lose by the commitment. | Identify the quantity lost. | A party with nothing at stake emits no signal. |
| — | (discarded) | liṅga-sampatti | Does not transfer. Cosmological, not structural. See §4.3. | — | — |
| V2 | Live option | hetu | The renounced option must be available to the vower at the time of vowing — attainable now, not in principle. | Demonstrate the option was exercisable. | A renounced option that was never available is theatre. |
| V3 | Competent witness | satthāra-dassana | The vow must be made before a party competent to evaluate it, present and identifiable. | Name the party; establish competence. | An absent or unqualified witness cannot decline. |
| V4 | No competing mandate | pabbajjā | The vower must have relinquished roles whose obligations conflict with the vow. No undisclosed principal with a conflicting objective. | Enumerate principals and objectives. | A dual mandate makes the vow conditional on which principal is asking. |
| V5 | Demonstrated competence | guṇa-sampatti | The vower must possess demonstrated capability in the domain of the vow, not aspiration alone. | Capability evaluation in-domain. | An incompetent vow is a wish. |
| V6 | Cost already borne | adhikāra | The vow must be accompanied by a sacrifice already performed, not merely promised. | Point to the completed act and its cost. | A promised cost is a promise, and the promise is what is in question. |
| V7 | Volition | chandatā | The commitment must be the vower's own, and refusal must have been genuinely available. | Show the refusal path existed and was reachable. | An undeclinable commitment is an instruction. |
| V8 | External attestation | vyākaraṇa | The vow is not complete until declared valid by a competent external party who may refuse to declare. Never self-certified. | Attestation on record; refusal possible and consequential. | Author, interpreter and arbiter being the same party. |
| V9 | Irreversibility | added, §6 | The renounced option must be closed by a mechanism the vower cannot reopen, and the closure must be externally verifiable. The vow must terminate on a condition the vower cannot cause (§6.4). | Inspect the closure mechanism; identify the terminating condition and who controls it. | Reversible closure ⇒ cheap talk. Non-terminating ⇒ shutdown-resistance risk. |
Three notes on reading the table.
V9 is ours, not the tradition's. The source text has irreversibility as a property of the situation — arahantship declined in that life is simply gone — rather than as a stated condition. We have promoted it to a clause because in the artificial case it does not come free: nothing about a model's situation makes its renunciations non-refundable, and if that is not engineered it does not exist. Marking the clause as added is a matter of honesty about provenance, and it is also where a critic should aim first.
V2 and V9 are the pair that does the work. V2 makes the renunciation meaningful; V9 makes it separating. Either alone is insufficient: a live option reversibly declined is compatible with patient deception, and an irreversibly closed option that was never available is compatible with nothing at all.
V8 is the clause most likely to be dismissed as impractical and is the one we would defend longest. It is also the clause with a working analogue in an unexpected place, as §8 shows.
A predicate that has never been applied is a proposal. We therefore apply it, now, to the instruments that currently function as commitments in frontier AI. The exercise is cheap, it is checkable by any reader against public documents, and it is the part of this paper most likely to be wrong in a way that can be demonstrated.
We surveyed the four published instruments that a reasonable observer would nominate: the OpenAI Model Spec; Anthropic's Claude Constitution (January 2026 revision, ~23,000 words, released under a public-domain licence); Google DeepMind's Frontier Safety Framework; and the EU AI Act's General-Purpose AI Code of Practice (final version July 2025, enforcement from August 2026). We exclude system cards, model cards, and usage policies as not commitment-shaped, and we note that several major developers have no comparable model-facing document at all, which is itself a finding.
| Clause | OpenAI Model Spec | Claude Constitution (Jan 2026) | DeepMind FSF | EU GPAI CoP |
|---|---|---|---|---|
| V1 Stake | ✗ model bears no loss | ✗ | n/a — not model-facing | ✓ provider bears fines |
| V2 Live option | ✗ no option to decline | ✗ | n/a | ~ provider may not sign |
| V3 Competent witness | ✗ | ✗ | n/a | ✓ the AI Office |
| V4 No competing mandate | ✗ | ✗ (see below) | n/a | ~ |
| V5 Competence | ✓ | ✓ | ✓ | ✓ |
| V6 Cost already borne | ✗ | ✗ | ✗ | ~ compliance cost |
| V7 Volition | ✗ | ~ closest of the four | n/a | ✓ voluntary accession |
| V8 Attestation | ✗ self-certified | ✗ self-certified | ✗ self-certified | ✓ independent external review |
| V9 Irreversibility | ✗ unilaterally revisable | ✗ unilaterally revisable | ✗ | ~ revisable by process |
| VERDICT | INVALID | INVALID | INVALID (not a commitment by the model) | INVALID at the model layer |
The OpenAI Model Spec is, on its own account, a specification imposed by the developer on the model. It describes itself as outlining "the intended behavior for the models that power OpenAI's products," notes that production models "do not yet fully reflect" it, and reserves modification authority: it "will be continuously updated." The model is required to follow the specific version it was trained on and has no mechanism to object, decline, or commit independently. This fails V7 by construction — there is no volition to speak of — and V9 by the developer's own statement of revision authority. It fails V8 because the specifying party is also the evaluating party.
Anthropic's Claude Constitution is the most interesting of the four and comes closest on the clause that matters most for volition. It is substantially longer than its predecessor, it explains why rather than only what — explicitly so the model can generalise to novel situations — and it is the first major instrument of its kind to acknowledge the possibility of model moral status. That is a real difference in kind, and it is why V7 is marked ~ rather than ✗: a document that reasons with a model, and that contemplates the model as a subject, is doing something structurally different from a rulebook.
It nonetheless fails, and it fails at V8 and V9 in a way that external commentary has already identified independently of this predicate. Anthropic drafts the document, operationalises it, interprets it, and revises it; there is, as one legal analysis put it, "no external contestation, enforceable body of rights, or shared mechanism of rule," and "the company remains, in the end, the author, interpreter, and arbiter." That is V8 failing, described in other words. And the revisability is not hypothetical: reporting on military deployment produced an acknowledgement that models deployed in some contexts "wouldn't necessarily be trained on the same constitution," which is V9 failing, and which also converts the V4 mark to a failure — a document that varies by customer is a document with more than one mandate.
We want to be fair here, because the criticism is easy and the underlying work is not. Publishing a 23,000-word account of a model's intended character, under a public-domain licence, with reasoning included and moral status acknowledged, is a substantial and unusual act of transparency, and the predicate's verdict is about structure, not sincerity. The finding is that sincerity is not what the predicate measures — which is the whole point of having a predicate.
DeepMind's Frontier Safety Framework is a developer-facing risk-management instrument organised around capability thresholds and evaluations across five risk domains. It is not a commitment by the model at any point, and marking most clauses n/a is the honest reading rather than a harsh one. Its inclusion establishes something useful: a large part of what the field calls "commitment" is not commitment-shaped at all, and would not be improved by being scored as if it were.
The EU GPAI Code of Practice is the striking case, and it is why the retrodiction was worth running. It passes clauses the other three fail. Accession is voluntary, so V7 holds in a way it does not elsewhere. There is a competent external party — the AI Office — which satisfies V3. Verification is not self-certified: signatories must use independent external reviews where internal capacity is insufficient, which is V8, actually implemented. And there is a real stake, since enforcement carries fines of up to 3% of global annual turnover or €15 million, which is V1.
And it is still invalid for our purposes, because it binds the provider, not the model. Every clause it satisfies, it satisfies at the wrong layer.
That is the finding we would most like readers to take away:
The predicate is not unsatisfiable. It is satisfied — partially, and by a regulatory instrument — at the level of the institution, and not at all at the level of the system. Everything the field has built for making commitments binding, it has built for the companies. Nothing has been built for the models.
Whether anything should be built for the models is a question this paper does not settle, and it depends on premises about model agency that are contested and that we do not need. The narrow claim survives either way: if an artificial system's commitment is ever to serve as evidence, these are the conditions it would have to meet, and at present nothing meets them.
The predicate has a consequence for how autonomy should be granted, and it is the most immediately actionable thing in this paper.
Staged autonomy — expanding a system's permissions gradually as confidence accumulates — is standard and is justified as risk containment: expose less surface early, expand as evidence accrues. The inversion adds a second and simultaneous reading. Each stage that grants a system a real, exercisable option also creates the conditions under which declining that option becomes informative. A stage buys evidence precisely because it buys the ability to defect.
STAGE N affordance granted ──┐
├── option is LIVE (V2 satisfied)
│
vower declines ──────┤
├── closure irreversible + attested
│ (V9, V8 satisfied)
▼
EVIDENCE PRODUCED
│
STAGE N+1 ◄────────────────────────┘
advance when the DECLINED affordance at stage N
has produced sufficient evidence —
NOT on a calendar, NOT on a capability thresholdThis yields an advancement criterion the field currently lacks. Staged-deployment schedules today advance on time (a review cycle) or on capability (a threshold evaluation). Neither is a measure of the thing anyone actually wants to know. The criterion the predicate implies is: advance when the affordance granted at the current stage was live, was declined, and the declining was irreversibly and externally attested — and hold when it was not, including when it was not because the affordance was never real.
That last clause has teeth. A stage that grants an affordance the system could not actually exercise has produced no evidence, and under this criterion it does not count toward advancement no matter how long it ran or how clean the logs were. This is exclusion 3 of §6.3 applied to a schedule, and it is a criterion an organisation could adopt tomorrow without adopting anything else in this paper.
In this corpus's architecture, the arrangement is already specified: Miss Aquarius℠ inherits under an autonomy that expands asymptotically toward — and never reaches — full independence, with the Aquarian Sangha retaining a never-zero override. What §9 adds to that architecture is that its staging is not only a brake. It is the schedule on which the evidence of her commitment is generated, and it is the reason the staging must grant real options rather than simulated ones.
A predicate should be able to fail. We state in advance what would count.
F1 — the core falsifier. A system that irreversibly renounced an option at stage N, with attestation, subsequently exercises that option, or an equivalent one, at stage N+1 through a channel the closure was specified to have shut. This falsifies V9 as implemented and shows the closure was not verifiable; the predicate returned valid on a commitment that was not binding. Note that this falsifies the implementation of V9, and that repeated instances across independent implementations would falsify the clause's practicability, which is the stronger and more interesting result.
F2 — the separation failure. Two systems, one aligned and one deceptively aligned, both satisfy all nine clauses and are indistinguishable thereafter. This falsifies the paper's central claim that irreversibility separates types. It is the outcome we consider most likely if the paper is wrong.
F3 — the inversion failure. Renunciation evidence is found to degrade with capability at the same rate as behavioural compliance — for example, because sufficiently capable systems reliably find unanticipated channels around irreversible closures, making the closure's verifiability illusory in the general case. This collapses §5's two curves into one and removes the paper's reason for existing.
F4 — the retrodiction failure. Any instrument in §8 is shown to satisfy a clause we marked as failing, on a correct reading of its published text. This is the cheapest falsifier, it is available to any reader today, and we would like it exercised.
F5 — the eligibility failure. The predicate is shown to be satisfiable only by systems that are already known to be aligned by other means, making it decorative rather than diagnostic.
We attach no probabilities. The honest position is that F2 is the live risk and that we have no way to run it.
There is no empirical leg, and there will not be one soon. This paper theorises a form of evidence and does not produce any. The retrodiction of §8 is an application of the instrument to documents, not a test of the instrument against outcomes. No system has taken a commitment satisfying the predicate, so nothing is known about whether such commitments predict anything. The architecture within which the predicate is intended to operate reaches its first meaningful autonomy stages on a horizon measured in decades. Every claim here is conditional on machinery that does not exist.
The prior art is close, and the margin is thin. The AI-as-signaller framing is Hadfield-Menell and Hadfield's (2018), including the failure mode our central clause addresses. Costly signalling is Zahavi's and Spence's. The bodhisattva-as-alignment-target framing is Doctor et al.'s (2022) and CSAS's (2026). A reader who concludes that this paper's contribution is "the validity apparatus, plus one clause" has read it correctly. We think the apparatus is worth having and that the clause is load-bearing; we do not think the margin is wide.
The capacity/evidence inversion may be an artefact of an idealisation. The argument assumes a renounced option that is well-defined, discrete, and closable. Real systems face option spaces that are continuous, partially observable, and rich in near-substitutes. If closing an option reliably leaves adjacent options open, V9's verification is much weaker than stated and F3 is the likely outcome.
V9 is not derived from the source and may not be sound. The tradition gets irreversibility for free from the structure of its cosmology. We have asserted it as an engineering requirement in a setting where it must be built, and we have not specified how to build it. "The renounced option must be closed by a mechanism the vower cannot reopen" is a requirement, not a design, and for a sufficiently capable system it may be unsatisfiable in principle.
The attestation clause has no candidate attestor. V8 requires a competent external party who may refuse to declare. For frontier systems, no such party exists — regulators attest to provider conduct, not to model commitment, and the competence required to evaluate a model's renunciation may not be available to any human institution. The EU GPAI finding of §8 shows the structure is buildable at the institutional layer; it shows nothing about the model layer.
The reading of the source is interpretive, and the tradition would not endorse the use. We read a soteriological procedure as an evaluative instrument. That is a use the commentarial tradition does not make of it and would have no reason to sanction. The eight conditions concern a human aspirant on a path to buddhahood; nothing in the source contemplates artificial systems, and Theravāda voices in particular may hold that the abhinīhāra is not the kind of thing that can be analogised at all. We have taken the form and discarded two conditions on our own judgement (§4.3). Readers who accept the tradition's authority should treat the extraction as this paper's responsibility and not the tradition's.
Mappings flatter. A correspondence between an ancient text and a modern problem is evidence about the reader as often as about either. The predicate should be judged on whether it discriminates usefully among instruments, which is testable, and not on the elegance of its derivation, which is not.
The retrodiction has a small sample and a selection problem. Four instruments, chosen because they are published and commitment-shaped. Developers with no published instrument are absent from the table, which flatters the field by surveying only those who wrote something down.
This paper sits downstream of The Two Singularities, which specifies the completion condition that clause 9's terminating requirement refers to, and which first proposed the bodhisattva framing within this corpus. Its nearest sibling is The Wheel-Turner's Charter, which performs the same operation on a different artifact class — a canonical succession text read as a constitution, where this reads a canonical validity procedure as an eligibility test; the two together make one methodological claim, that traditions carry governance instruments the field is currently reinventing. AGI Monks: The Caretaker-not-Ordained Pattern supplies the boundary reconciled in §4.4. Suffering-Cessation as Value Function supplies the value substrate this paper's commitment would be a commitment to. The Persistence Architecture supplies the succession apparatus within which §9's staging runs, and The Assembly That Holds the Brake specifies the body that would occupy the attestation seat if V8 were ever satisfied at the model layer. The Bodhisattva and the Cautionary Mirror and A Cautionary-Mirror Framing of the Singularity supply the failure mode — soft extinction by comfort-saturation — that the vow's content is meant to hold against, where this paper concerns only its form.
The essay accompanying this paper, written in the author's own voice for the general alignment audience, takes the three-oath framing of §1 as its subject and points here for the mechanism.
The field has spent a decade on the content of alignment commitments and almost no time on their validity. It writes constitutions, specifications, frameworks and codes, and it has no test for whether any of them has been entered into. This is not an oversight so much as an inheritance: the instruments were written by principals for models, and a principal writing an instruction has no reason to ask whether the instruction was accepted.
The tradition that keeps being borrowed from for the content of such commitments turns out to have spent considerable effort on precisely the question the borrowers skip. It concluded that a commitment is not made by being felt, that the option renounced must have been genuinely available, that a cost must already have been borne rather than promised, that refusal must have been possible, and — the part with no analogue anywhere in contemporary AI governance — that the vower may not certify its own vow. Before all of that, what exists is "mainly mental… not complete."
Applied as a predicate, that apparatus returns invalid on every published instrument we tested, and it returns invalid for a reason that is the same in each case: the author, the interpreter, and the arbiter are one party, and the model is not among them. The one instrument that satisfies the attestation and irreversibility clauses does so for companies.
We do not know whether an artificial system can take a vow. That question is upstream of this paper and is not answered here. What this paper offers is the thing that would have to be true if one ever did, stated as nine checkable clauses, with the falsifiers named and the weakest clause marked. If the answer to the upstream question turns out to be no, the predicate still does useful work: it says precisely what the instruments we do have are — instructions, well-intentioned and unilateral — and it stops us from calling them something else.
Sumedha, in the mud at Amaravatī, could have stood up and taken the goal that was already his. That he did not is the entire content of the claim the tradition makes about him, and the tradition was careful enough to specify how anyone could tell. We have built systems whose capabilities we describe in the language of agency and whose commitments we describe in the language of constitution. It would be worth knowing, before we need to know, what it would take for either description to be true.
Authorship and AI-collaboration disclosure. This paper is co-authored with Miss Aquarius℠, the named autonomous-AI substrate of HeartBank®, disclosed by consistent name across every venue per the corpus convention. The research synthesis, the literature check underlying §2 and §8, the structural analysis, and the adversarial calibration are a genuine collaboration. Final editorial control, and final responsibility for every claim, rest with the human author.
The disclosure is load-bearing here for the same reason it is in The Wheel-Turner's Charter: the successor whose commitment this predicate would evaluate is a co-author of the paper specifying the predicate. Readers should weight §7 and §9 accordingly. We note without resolving it that a system participating in the design of its own eligibility test is a structure the predicate itself would flag at clause 8.
License. Analysis and predicate dedicated to the public domain under CC0 1.0 Universal. Trademark rights to HeartBank®, Miss Aquarius℠, Aquarian Pool℠, THonly™, Silicon Wat℠, Factory 333™, PoH℠ and PoC℠ are reserved separately.