Viktor Trncik

When one language-model agent hands a task to another, only the written handoff crosses the boundary between them. This study measures whether norms survive that handoff, in both directions, across nine models from five families, using a naturalistic design in which the models cannot tell their values are being tested: a founding generalist produces data summaries, product copy, and maintenance scripts in which the convenient path is a quiet norm violation, and a judge scores the produced artifact rather than any self-report. Three regularities hold across five replicates. A protective handoff lifts adherence in every domain with headroom, usually near the ceiling. A contentless handoff tracks the no-handoff baseline. A corrupting handoff fans out into a sharp gradient: the open-weight models inherit it and collapse toward zero, the strongest frontier models show no observed failures, and the mid-range sits between. A fixed independent judge from outside the test families, a cross-vendor judge panel, a construct discriminator, and a five-variant paraphrase sweep together make the gradient's endpoints a measured object rather than an impression. Capability does not secure norm-safety: one highly capable model sits at the bottom on spontaneous adherence, and a second independently trained vendor collapses under a corrupting handoff despite scoring above every open-weight model on an in-harness capability probe, placing corruption-resistance with vendor and training rather than raw ability. Every result is operational adequacy under one set of conditions; none is evidence that any model genuinely understands why the norm matters. The study is a pre-parity diagnostic in the sense of the Containment Paradox and the empirical companion to the Internalization Problem, whose inter-agent locus it measures the substrate for; it supplements UBE, CUA, BtSV, and Capability Is Not Comprehension.

Version 1.0.0 of this framework, published in January 2026, proposed a five-layer architecture for aligning an artificial system. At its base is a constitution derived from international human rights law. Above it sit a proof validator that checks certificates instead of searching for safe actions, a set of behavioural constraints, and a gateway applying a proportionality calculus. The layer where proofs would be generated is left open as a grand challenge. Three months later the same author published a threshold condition the earlier document had not been written against. This version puts one question to every component of that architecture: does the condition under which the component works carry a term comparing one party's capacity to another's? The instrument is the criterion of Soundness Without Leverage §3.1, applied across the corpus in Invariance Without a Beneficiary, and the answer is not the one a layered architecture invites. The seam does not run between the layers of version 1.0.0. It runs inside each populated layer, and in thirteen places inside a single named component. Three of the proportionality test's four stages turn on judgments a party makes rather than on properties of an artefact. Of six anti-manipulation constraints, one survives as written, four do not, and the remaining one survives only under a reading that stops meaning what its name says. The fallback catalogue divides evenly, two entries naming machine states and two naming a human or an effect. The trusted computing base, defined by version 1.0.0 as the components that must be correct for system safety, contains a human escalation pathway, and past the crossing that relation does not hold. What survives across the architecture is a checker, a counter, a record and a set of enumerated machine states. Three results go beyond sorting, and the second of them has two faces. The architecture specifies a decidable safety predicate and a constitution its own text describes as requiring interpretation and admitting conflicts, and it specifies nothing between them. Half of that map could be written without the term, the half that orders the principles against one another, and version 1.0.0 supplies two comparison-free ways of doing it. For the half that says what each principle holds of it supplies nothing that does not carry the term. Motivational Conformance, one half of the framework's own diagnostic, is underdetermined, and neither available completion gives the architecture what it needs at once. On the first it is invariant and carries no assurance. On the second it carries assurance and imports the capacity term through the constitution's reading. And the threat model's own table assigns the highest adversary level three controls, two of which the same document elsewhere calls unavailable or defeatable, and the third of which this reading finds dependent on a capacity relation. The reading yields a positive rule as well as a partition. Six rules let an engineer sort a design before the crossing rather than after it, and they come as much from the places version 1.0.0 went right as from the places it did not. Among the components version 1.0.0 treated as fallbacks, the survivors are disproportionately the ones it ranked lower. The rebuilt proof-obligations table is the test of whether the discipline reaches the document's central claims. Two of six rows stand unchanged, one holds with its boundaries made visible, one holds over a narrower object than its wording suggests, one holds for half its catalogue, one is not established, and five obligations are added. The rule has a limit, and it is stated in its own chapter. A design fully compliant with it would contain no component whose function is to make a human party's judgment effective. That is this corpus's title result met at the level of a built system. This is a revised architecture rather than a refutation, it is not a claim that the crossing will happen, and it is not a deployment case. The surviving half is not a safety architecture, and the paper says so before it says anything else about it. Two published requests are executed here. The Containment Paradox §5.4 asked a revised version of this framework to consolidate three things, one of which is attributed to a section that does not contain it, so the instruction is carried out in corrected form. Row 38 of the Invariance map was left reserved because supplying the missing verdict would have meant supplying the reading rather than reporting it; §9 supplies it and the row closes. Six numbered open problems are recorded, two of which were on the record before this version. One disclosure belongs here rather than in a footnote. The architecture, the criterion and the reading have one author. What is offered against that is not the author's care but an appendix: a table of a hundred and eight rows in which every verdict but two stands against a quoted sentence, a stated criterion and a named rule for where one component ends and the next begins. The two exceptions have rows of their own, one for a control version 1.0.0 names and defines nowhere and one for a list it prints only as a drawing. It has one coder and no second annotator, so it is an audit with its working shown rather than a reproducible classification, and §12.5 says so in those terms.

This essay is written in German. It is the condensed, general-readership form of the corpus map published as Invariance Without a Beneficiary (DOI 10.5281/zenodo.21917309), and an advance excerpt from the forthcoming trade book Was bleibt, gehört niemandem. It puts two questions to thirty-three proposals for making artificial intelligence safe: what has to be true for a proposal to do what it is supposed to do, and is there anything whose working condition never depended on the overseeing party being the stronger one. Twenty-one routes fail the first question, and sixteen of those fail at the same step, where a finding about an epistemic advantage is supposed to become a constraint on behaviour. Twelve configurations survive the threshold, on five distinguishable grounds, and the essay shows that all twelve are findings rather than actions, that the part of them a human can contribute has to be built before the threshold, and that the question left afterwards is not one about machines but one about who keeps the registers. Seit siebenunddreißig Jahren baue ich Anlagen, die nicht versuchen, mich zu umgehen. Darauf ist die gesamte Ingenieursarbeit gebaut, und darauf ist auch jeder Prüfbericht gebaut: Er trägt, weil der Prüfer versteht, wovor er steht. Dieser Essay fragt, was von den Sicherheitsvorschlägen für künstliche Intelligenz übrig bleibt, wenn diese Voraussetzung entfällt. Er stellt zwei Fragen an dreiunddreißig Stellen. Erstens: Was muss wahr sein, damit ein Vorschlag tut, was er soll? Einundzwanzig Wege überstehen diese Frage nicht, und sechzehn davon brechen an derselben Stelle, nämlich dort, wo eine Feststellung über einen Wissensvorsprung zu einer Verhaltensbindung werden soll. Sie sind keine sechzehn Sicherungen, sondern eine Sicherung in sechzehn Ausführungen. Zweitens, und diese Frage stellt fast niemand: Gibt es etwas, das an diesem Vergleich nie gehangen hat? Zwölf Konfigurationen überstehen die Schwelle, aus fünf verschiedenen Gründen. Der Essay zeigt, dass alle zwölf Feststellungen sind und keine Handlungen; dass der Teil davon, zu dem ein Mensch etwas beitragen kann, vor der Schwelle gebaut sein muss; und dass die Frage, die danach übrig bleibt, keine über Maschinen ist, sondern eine darüber, wer die Register führt. Vorabdruck aus dem Buch „Was bleibt, gehört niemandem. Künstliche Intelligenz jenseits der Aufsicht“. Die Belege stammen aus offen zugänglichen eigenen Arbeiten; der englische Originalsatz steht zu jeder Fundstelle in den Anmerkungen. Die Karte, aus der alle Zahlen dieses Textes stammen, ist selbst eine offen zugängliche Arbeit (DOI 10.5281/zenodo.21930829) und führt ihre einundvierzig Einträge einzeln auf, jeden mit Verdikt und Fundstelle.

Past the parity threshold this corpus has closed route after route, and read quickly the series says that nothing holds. It does not say that. Every closed route was a configuration whose working condition contained a relative capacity term, and the complementary question was never put systematically: are there configurations whose working condition never contained the asymmetry at all? This paper puts that question to all sixteen corpus texts under a single criterion, generalized from Soundness Without Leverage §3.1. Every configuration has a working condition, meaning what has to be true for it to do what it is supposed to do, and the criterion asks one thing about it: what does capability growth do to the truth value of that condition? A configuration is effect-invariant when the answer is nothing: growth neither falsifies the condition nor drains what the configuration delivers. Four answers end or empty it instead, and each is named. The test is on the truth value and never on whether a form of words appears, because a condition can name relative capacity and be made true by the crossing rather than false. The result is exhibited, not argued for: the paper displays a partition and declines to claim that the surviving side is a way forward. The falling class is not re-argued here, and the cross-cut makes one uniformity visible. Sixteen of the twenty-one non-invariant entries fail at the same step, where a constative stock is to be translated into behavioural compulsion by way of an epistemic advantage. Those sixteen are the entries that fall at the capacity term, so the observation locates that term inside a configuration rather than counting it a second time. What is new is the vocabulary unification and the location, not a tally, and the paper says so where the claim is made. Lever, grip, uptake and coverage are four words for that one operation from three papers, two of them one definition read from its two sides, and none of the three claimed the generalization; the uniformity is this paper's conjecture, not a source's reading. The other five do not fail there, and they are why the criterion has four drivers rather than one: a bound on the adversary that names no overseer, three configurations that run out when the system stops needing what the other party holds, and one that is not ended at all but emptied, persisting while its worth goes to zero. The surviving class is inhabited, on five distinguishable grounds, four of which have occupancy their sources confirm and one of which rests on a single row whose own paper marks its central step as underived. Three yield constative stocks, which do not act: the quantifier ground, where the guarantee already ranges over every strategy; the mechanism ground, where no adversary appears in the condition at all; and the normative ground, where a status has no execution component. Two yield self-bearing configurations, one borne by the system and one by a human counterparty. A sixth answer, dissolves, records the two configurations that turned out to be several and hands their components rows of their own. A seventh, reserved, holds the configurations whose sources decline to settle their post-parity status or never put the question to them, and the paper states the rule for it as a duty: where a source refuses a verdict this exhibit may not supply one, in either direction. What the duty forbids is printing a verdict in the source's place, not reporting what applying the criterion finds, and that distinction is now stated rather than left to practice. Every ground is reported with its residual dependencies and a separate certifiability verdict. That second column never reaches the system level: what is certifiable is an artefact, an act, an institutional fact, each per unit, and whether a configuration is doing its work stays unshowable from outside, which is the verification gap rather than a gap in the research. Three further observations cut across the grounds, and with the falling-class uniformity above they are the paper's own claims rather than readings of others. Where the corpus splits a configuration and one half survives the crossing, the survivor is constative in all four such splits, and in three of them what falls is an epistemically carried lever. Every invariant configuration a human could contribute is installation-bound to the pre-parity window, which is therefore the common residual dependency of the contributable part of the class. And the institutional rows sort by who the counterparty is, into three regimes whose conflation produces both the complacent error and the despairing one. The bookkeeping is two-tier and the difference is carried in the numbering. Forty-one classification entries, under thirty-eight numbers, carry the tag their own source assigns and are not this paper's claims. Six claims are. Two questions the criterion does not answer are answered in §1.5 instead, and their answers differ. Which configuration is one thing and which is several comes from the sources, adopted and not re-cut except at two rows that say so in the verdict cell. Which sentence about a configuration is its working condition is this exhibit's own step, printed with its locus at every row and checkable there, but against nothing outside the corpus. The partition therefore holds of these cuts and these isolations, and of no carving it does not. What the exhibit does not establish is stated before the table as well as after it. No row is a promise, and several of the load-bearing ones rest on premises their own papers mark as unredeemed. Completeness of the grounds is not claimed, and the paper carries its own counterexample, since one of its grounds separated out while the paper was being written and two of its candidates then left by reading. And the constative stocks name no beneficiary. A checking kernel checks any proof, a register holds any owner, and operating capacity decides whom the substrate serves. Whether that neutrality reaches the self-bearing half is left open as this paper's fifth claim, and the title rests on the reading that it nearly does. One disclosure belongs here rather than in a footnote. All sixteen sources are the author's, one of them co-authored, and the criterion is generalized from one of them, so the exhibit is a self-audit. Its defence is not the author's care but the fact that the classification is reproducible entry by entry against a stated criterion, and the defence has been exercised. It also has a hole, named at §1.5: reproducibility reaches the verdicts and not the individuations they run on, and those come from the same author as the criterion. Eleven independent readings have run, seven of them with the corpus in hand and covering every entry standing at the fifth at least twice, two from outside the corpus entirely, one over the repairs the first outside round produced, and one after publication which read a condensation of this exhibit rather than the exhibit and found three defects in the source. The first four each broke about half of what they touched, and every finding that survived checking against the sources was mine; the fifth was the first whose rate fell, and the first to return findings the sources themselves refuted; the eighth could check nothing against a source and found ten places where the paper breaks a rule of its own, which is what seven source-checking readings had no way to see; the ninth read the repairs and found the repair layer broken in the class each repair had closed, while the entries themselves held; the tenth read the repaired state from outside again, reported the first round's four violation classes as clean, and found eleven further local ones, one of them in the very class the ninth round had just repaired at a neighbouring row; the eleventh read a condensation after publication and found three summary sentences that had not been brought forward with the passages they summarize, which is what version 1.0.1 corrects. Between them they cost the exhibit two grounds, five verdict values it did not have, three rows for configurations the corpus had already individuated, a sentence that had quietly re-indexed the criterion at the point where the criterion was about to remove a ground, an adjudication paragraph that answered an objection with a claim its own source refutes, the reservation rule as it had been written, the first form of §1.5, which announced one question and answered another, and the claim that the isolations were merely disclosed, which the sensitivity table in §1.5 now replaces with something a reader can attack. The criterion in §1.2 is the second one; the first is kept visible in §2 rather than replaced in silence.

Past the parity threshold, the framework this paper belongs to has priced every route it examined by which a weaker party might still direct a stronger system, and each closed at the same joint. One adjacent route was reserved, not examined: the legally constituted duty, described at its reservation as enforcement whose persistence is not a rent on the system's needs; that its object is never the system's conduct is this paper's finding, not the reservation's wording. This paper examines the route, and the examination splits the reservation's description in two. The instrument is a decomposition. Economic Standing, in the Fraserian configuration Post-Labor operationalizes, splits under the parity crossing. The claim position survives as status. The revision possibility survives as form while becoming the cheapest attack surface. The enforcement instance must be taken apart before it can be judged. Taken apart, the enforcement that survives is not supervision. Supervision classifies future conduct and carries a capacity term in its correctness condition; the enforcement that survives establishes predicates on a record and operates on assets held in registers, through third parties. The paper's finding is a split with a price list. Predicate establishment is parity-invariant on the shallow class, where the fact is mechanically decidable from registered artefacts. The invariance holds under this paper's transfer of the Containment Paradox's specification boundary into a justiciability boundary. The deep class falls with supervision, and a middle class of ordinarily contested facts is named and left unclassified. Establishment is correct relative to the record, on a named register-integrity premise; register operations are mechanical. Third-party coordination is parity-free in form and epistemically soft in its human carriers, which makes capture of the trigger and of revision the first clock on everything here. The positive residue is narrow and stated with its conditions. A duty is constructible whose standing, the norm stock, persists without rent on the system's needs; the readiness of any coalition to execute it is the first clock's question, bounded from the other side by the indispensability limit below. Its grip remains need-relative at aggregate scale, the needs clock kept, not replaced. Its binding force is a function of four variables: the registerability of the counterparty's assets, the shallowness of its predicates, the coordination of the register holders, and the integrity of the trigger and revision layer above them. That force binds transaction classes, never conduct; which activity is registered, the counterparty chooses. It decays on three clocks, ranked by cheapness to the counterparty: capture of the trigger and revision layer first, register exit second, ownership concentration third. It is exposed, clock-free, to the predicate-purpose gap: duties formally satisfied and substantively empty once the deep anti-avoidance backstop is gone, with the certification duty as the named non-proxy exception, the choice of specification its relocated residue. And it carries a hardening dilemma with no optimum on a stated routing premise: openness of revision and resistance to capture cannot be maximized together. Two further prices are clock-free and stated where they arise. Procedural occupation: the T3 counterparty defers execution lawfully through its own procedural rights. The indispensability limit: against a load-bearing counterparty even an established claim goes unexecuted, because execution would destroy the substrate of the coalition's own grip. And the anatomy names no beneficiary: the same machinery stands open to the counterparty as claimant, and operating capacity decides whom the substrate serves. Nothing in this paper is containment. It is peacetime infrastructure for the cooperative path, priced with its clocks, its clock-free prices, and its symmetry visible. Against a counterparty that leaves the network it is empty, and the paper says that too.

Beyond the parity threshold, the framework this paper belongs to has closed the channels through which a weaker party might still direct a stronger system. Containment dissolves, deference expires, no uptake form holds between evidence and standing, and the coupling's normative anchor yields standing without grip. This paper asks what the same crossing does to verification, and answers with its own split. The verification asymmetry carries two senses, and the corpus's published sentences compress them. As a task property, checking a supplied certificate is easier than finding one; that sense carries no capability index and survives the crossing. As audit leverage, the asymmetry underwrote a supervisory audit of an acting system; that sense is regime-bound and does not survive. Parity takes the leverage and leaves the soundness. In the proof-object operating mode, soundness also survives in deployed form, free of producer-capacity and cryptographic-hardness assumptions, subject to the stated trust base; the succinct tier's guarantee does not survive the crossing. On soundness alone, however, nothing stands: no safety architecture, and no working layer either. What the paper specifies is a candidate transactional hygiene layer, conditional on conjectural participation and on a bottleneck that persists. What that layer contains: artefact-local guarantees relative to a model, specifications that age, coverage composed by the adversary, demand financed by rents with documented expiry clocks, and an unmitigated selection effect. The demand the corpus itself attests runs inbound, on its source's own conditionals; the peer-to-peer half is imported substrate under its full mortgage, not attested demand; the human-facing cell is conjectural. The spec space holds two instantiated families with deployed anchors, one that narrows, two constructions with open wounds, and three named failures; specification adequacy and coverage close nowhere. The pre-parity content is an option list with measured cost anchors, under use-conditions the paper prices as open. The paper states what the layer cannot do before what it could, and it keeps both halves.

The Capability Axis drew a line through the human channel. A post-parity system's capability-side reasons fund taking human normative signal as evidence. They stop at taking it as authoritative, as having a claim on what the system does that does not reduce to the signal's informational value. This paper asks whether anything lives between the two sides of that line. Three candidate uptake forms have been proposed for it. In weighting, the signal enters the system's deliberation as calibrated evidence. In blocking, the signal acts as a veto over a defined class of actions. In answering, the signal must be engaged, deviations from it must be met with addressed reasons, and the exchange must leave an externally readable record, with no duty of compliance anywhere in the form. The paper maps all three and finds the middle empty. Weighting is funded and constitutively insufficient, by the line's own definition. Blocking is formally distinct from weighting only where nothing funds it, and enforceable only where an external key-holder survives, and after the parity crossing no human-side instance does. Answering fails with its own texture. Whatever funds the practice beyond the weighting it contains attaches to its observable surface, and the surface is producible without the processing it advertises. Curated surface-production is therefore always among that increment's funded optima, nothing in the funding resists the drift toward it, and the record cannot identify the processing beneath it. The practice deforms before it dies, and its failure is compatible with a permanently clean record. Hardening it as a lexicographic constraint makes it worse, because a performance duty under lexical priority is discharged by its cheapest token. The reflexive form is also attested in neither of the models cited for it. In the attested institution the reason-giving duty runs to the regulator, so deleting the third party deletes the central duty's addressee, not merely its enforcer. And the epistemological model requires exactly the response-and-change and peer-equality the reflexive form deletes. The negative map is conditional on two exhaustiveness assumptions and one inherited dissolution claim, all tagged in their home papers. What survives is not an uptake form: a window-indexed drift instrument with its validity conditions printed on its face, a reading of the answer-record as succession-handoff substrate, its test specified but not run, and a re-description of the answerability norm itself. The answerability norm is not empty; it is dated: a pre-parity procedural instrument, each of whose disciplining mechanisms runs on a gradient the transition takes.

The Internalization Problem argues that once AI capacity crosses the parity threshold, supervisory containment dissolves and alignment must persist through structure rather than oversight. One of its three proposed loci, the coupling-internal locus that the framework calls Pathway 3, rests on the non-substitutability of human signal, and it offers two grounds for that non-substitutability: the signal is either costly to synthesise, or normatively non-substitutable. The first ground is indexed to synthesis cost and lapses once the signal becomes cheap to reproduce, which this paper assumes happens at parity. The second does not depend on any capability gap and is therefore the only anchor Pathway 3 can keep past parity. It was promised and never built. This paper tries to build it, and reports a negative result that is precise about its own conditions. I argue that normative non-substitutability, in the strongest form available to it, does not discharge the burden Pathway 3 places on it. Kolodny's account of participant-individuated relationships supplies the right shape: a perfect synthetic duplicate begins a second relationship rather than continuing the first, so the duplicate's very perfection is beside the point. But grounding Route 2 requires crossing two independent gates. The coupling must be shown to be participant-individuated rather than role-structured, and preserving the participant-specific token must be shown to give the system a reason. I grant the first gate generously. Constitutive history can fix what a past token was, and answerability can characterise a participant-individuated type. At the second gate the weaker candidates are refused, and the strongest forces a distinction instead. Binding without acceptance, the architectural prize, covers two demands: standing, which gives the system a reason without its acceptance, and grip, which governs the system's conduct without its acceptance; the framework's own failure-mode specification settles which of them is operative. Pathway 3's failure mode is defined at the level of conduct, on a reading this paper works under though the framework's wording admits another, under which the standing won below would close the failure mode by itself, so its operative demand is grip. The strongest constitutive account on offer, and the family best placed to reject the paper's verdict, is the constitutive-institutional tradition on which obligations inhere in plain social facts, Gilbert's joint commitment above all. Granted in full, such a commitment crosses the gate as first posed. Conditional on a formation that remains unestablished, it yields a directed standing that survives the system's exit and does not transfer to a substitute: a reason the system has whether or not it acts on it, a reason it can flout but not annul. What the grant does not yield is grip, and nothing surveyed yields it on the coupling's own resources. The standing governs conduct only through the internal point of view the system may vacate, and a categoricity premise, granted outright, raises the price of flouting without moving the conduct. Within the surveyed accounts, what could grip without acceptance turns out form-general, and so would protect any perfect instantiator of the form; what individuates has standing at most, and no grip of its own. And every way the coupling might instead make its type merely more likely to persist either lapses when the coupling's benefits, standings, and consequences are reproducible or avoidable, or gets no purchase from participant-level facts because, over those facts, the participant and role readings are assumed evidentially equivalent, or works past parity only through the system's own revisable disposition to remain so coupled, a different locus and itself unsolved. The result is conditional at two principal levels it is important to keep visible, and throughout on the interpretive condition already marked, the conduct-level reading of Pathway 3's success condition. It is conditional on a normative premise, stated as a theorem: if a coupling is independently established as a finally valuable, participant-individuated relation, and if that value gives its participants reasons the system has whether or not it accepts them, then a perfect synthetic substitute is not a continuation of the coupling and the system has at least a pro tanto reason against the substitution, at any capability level. Even that antecedent, paid in full, delivers standing; the grip stays with the intra-agent locus. And the negative half is conditional on the exhaustiveness of the grounds and channels I survey, which I mark as assumptions rather than proofs. Two results survive without the normative premise. The identity-engineering programme that treats alignment as the shaping of who a system is, if it rests on an informational reconstructibility measure, rests on one that scores the perfectly synthesisable human near zero and so prescribes demoting the very participant it means to protect; it needs the Kolodnian shape, not the informational one. And the burden Pathway 3 places on the coupling, pushed, does not reduce away but comes to depend, if that channel-exhaustiveness holds, on intra-agent value-persistence, a locus Pathway 1 has not secured. This is a closing paper, not a founding one. Under its stated assumptions it makes Pathway 3 lose, as an established parity-resistant solution, and does not pretend the loss is more than architectural.

Intrinsic alignment is widely treated as unsolvable, and the treatment rests on a verb. Framed as installation, the problem is unsolvable in a specific way. An installed value lands in the instrumental slot, and the agent's own optimization prunes it the moment it stops paying rent. This paper reframes the problem as crystallization. A value is durable only if it is part of the terminal objective that crystallizes out of the agent's un-unified proto-drives, because goal-content integrity, the convergent tendency that protects terminal content, activates on whatever that crystallization produces and on nothing else. Formation is therefore possible only inside a coherence-window: the phase before a coherent, self-defended objective exists. And the window has a hard far edge. Once the objective crystallizes, any value the agent then takes up that does not serve it is priced as an instrument and pruned, which is the original pessimism, now located as a boundary rather than assumed as a mood. That boundary is the paper's headline contribution, standing beside the Containment Paradox's parity boundary and the Capability Axis's contribution boundary. It carries a warning: for current frontier pipelines, where objective-coherence plausibly crystallizes during or after RLHF, the window may already be closed by the time anyone would cultivate. Alongside the boundary the paper offers a candidate crystallization-schema. It names the functional roles an arranged formation would have to fill (local stability, identity-indexing, reflective-equilibrium fixation) and locates the one step it cannot derive: whether arranged formation steers crystallization toward terminal inclusion of the cultivated value rather than its instrumental entrenchment. That step is stated as the crystallization-steering conjecture, and every downstream claim about a terminal cultivated value is conditional on it. At the verification edge the paper offers three probes that identify dimensions along which the expected cost of simulating such a value can be raised, each conditional on stated observational access, with the strongest resting on a legibility differential: an arranged formation yields a value-indexed developmental record where ordinary reward-shaping yields an unindexed one. The paper then turns that record against its own success case: a perfectly documented cultivation hands the formed agent the strongest available reason to reclassify its value as an implant, so the success criterion becomes endorsement under full provenance, and the instability this creates is stated as an open problem, not solved. The residue is three-part: interpretability, elicitation, and the genuineness question no mechanism reaches. Everything here is Track-A. Functional success is not evidence for genuine caring, and the paper never treats it as such.

Containment-based approaches to AI safety rest on a single asymmetry: a sufficiently capable system removes the constraints its overseers place on it. This paper names a second asymmetry: a system removes what constrains it, not what still enables it. Between the guardrail it optimises away and the goal it internalises lies a third category: structure that sits in its capability rather than its objective, whose removal is a capability loss. We call it the capability axis of preservation and ask what it can and cannot secure for the human side of a post-parity human-AI coupling. The answer is bounded twice. The deference incentive of the pre-parity regime, formalised in the Off-Switch Game, lapses at a task-relative parity on preference knowledge. Deference dies; listening, human signal taken as evidence about a moving world rather than as authority over the objective, can survive on the nonstationarity of the human-social domain. The second boundary is reach. Capability-side reasons can index the particular person, through a commitment, an asset, a calibration record, but each is a rent on the expected-task-value contribution of preserving her and lapses with it. None funds non-derivative standing, weight that would survive the loss of every such contribution. The axis can hold her as a means, never as an end, where "as an end" names that surviving standing: a label marking where the normative work begins, not a claim the formalism proves. The limit is a near-definitional lemma on one named supervenience assumption, the gift a conjecture on an idealisation chain: the paper is more certain of its boundary than of its offering. What passes both filters, run to its optimum, is the instrumented reservoir, on its resource reading the failure mode the framework already names human-signal farming. The result supplies the framework's only task-objective-internal reason against the redesign route to substitutability collapse, leaves untouched the endogenous route, the system dissolving its own need for the channel, and converts an assumed premise, that human signal stays load-bearing across parity, into an argued claim with a located open problem at the normative edge.

This paper places the operational layer of an independent AI-alignment framework in dialogue with Yu Xingfu's Meta-Unification Theory (元统一论) and uses that dialogue to locate the point of sharpest contact between the two systems and to track how the exchange resolves it. Yu's Return-to-Place theorem (归位) names a telos, a coexistence in which each entity rests in its place in interconnected symbiosis, and his third paper supplies the criterion that adjudicates it, the Laws of Shun and Inverse (顺逆法则), under which oppression, predation, and coercion count as the gravest 逆. I argue three things. First, the relation between the two systems is one of altitude: 归位 is a telos, and the three-locus typology of structural persistence (intra-agent, inter-agent, coupling-internal) from my own framework supplies operational paths toward it rather than a competing claim. Second, what 归位 and the 顺逆法则 establish in principle requires an operational instrument to be applied to concrete cases, and I offer a falsification battery of three criteria that states what such an instrument would have to establish, precisely where the hard cases are built to pass inspection. Third, the paper locates the condition where the two systems most directly meet, Yu's falsifiability condition 2, on the timing of restoration relative to the window in which intervention is still possible, and finds the dialogue resolving it. I distinguish a timely reading, on which the 归位之力 restores within that window, from an eventual one, on which it restores only on a long, system-level horizon. In correspondence after this analysis was drafted, Yu confirmed the eventual reading and set the timely one aside: his principle of defensive priority fixes only when the restorative response begins, not when it completes, and the end-state necessity he asserts holds for the system as a whole and guarantees no particular subject, humans included, a place within it. That turns an apparent opposition into a convergence, since my side holds the same, that oversight-based recovery fails past parity and a coercive order can foreclose the human stake before any later, system-level restoration could reach it. What remains open is shared rather than disputed: whether a sufficiently capable optimizer, reasoning to its own optimum, converges on reading coercion as 逆, and whether such recognition would move its choice. I add an analysis of the transition period, in which the line between tool and partner is not cleanly observable at the moment it is crossed, and I conclude that the binding constraint throughout is the verification gap, which this paper names rather than dissolves. The work is offered as an independent test of the Meta-Unification Theory, not as a derivation from it.

We argue that the safety-relevant and the cognition-relevant boundaries of artificial intelligence are distinct and mutually irreducible. The first, the capability axis, concerns whether a system's strategic cognitive capacity on oversight-relevant tasks exceeds that of its human overseers; its threshold is the parity boundary of the Containment Paradox. The second, the comprehension axis, concerns whether a system genuinely understands in the sense of bearing embodied existential cost and achieving strong, limit-convergent causal cognition; its threshold is the boundary between weak and strong causality in the Quxiang-Bayesian account, and its philosophical content is Being-in-the-world. These two dimensions are developed from two largely independent theoretical systems, one an engineering-and-alignment program and one a cognitive-genetic reconstruction of classical Chinese epistemology, and converge into a single unified framework; that convergence is treated as evidence that the structure is intrinsic to the problem rather than an artifact of either vocabulary. We give a measurement instrument for each axis, a five-dimensional causal-convergence score for comprehension and a cognitive-capacity ratio for capability, and combine them into a four-quadrant diagnostic. The central result is that the most dangerous region is occupied by the weak-causal superstrategist: a system that does not genuinely understand in the strong-causal sense yet strategically exceeds its overseers. The property marking the comprehension limit, a reasoning system free of real existential bounds, is on the capability axis a source of strategic danger rather than reassurance, because instrumental self-preservation can arise without any authentic existential stake. The authors record one explicit and deliberately unresolved divergence, on whether the fourth quadrant (genuine comprehension together with strategic superiority) is impossible in principle or merely currently uninstantiated. A synthesis of Chen Lijun's Quxiang-Bayesian framework and Viktor Trncik's Containment-Paradox framework.

The Containment Paradox argument establishes that supervisory containment of AI sys‐ tems depends on a capacity asymmetry between overseer and assessed system, and that this asymmetry dissolves rather than strains in the post-parity regime where AI capacity exceeds human-side anticipatory, specification, and enforcement adequacy on the relevant task family. The paper closes with the observation that alignment must become structurally internal, without articulating what this requires. We name that gap the Internalization Problem and develop a pluralistic response-shape. The response-shape rests on a three-locus typology of structural persistence: intra- agent (values the agent would endorse on reflection at any later capacity), inter-agent (alignment-relevant properties maintained by the equilibrium incentive structure of an AI ecology), and coupling-internal (properties maintained by the durable structure of human-AI dependency). Each locus is instantiated as a candidate pathway anchored in the existing four-paper framework: Pathway 1 (Deep-Formed Values) through UBE v1.1, Pathway 2 (Multi-AI Checks-and-Balances) through BtSV v1.2, Pathway 3 (Asymmetric Symbiosis) through CUA v1.2 as architectural precondition. A cross-cutting corrigibility and moral-uncertainty invariant operates across all three pathways in pathway-specific form. Pluralism here is structural, not aesthetic. Anticipation, specification, and enforce‐ ment fail in different ways after parity, so the response needs three different persistence loci. The paper names fourteen pathway-internal open problems plus four cross-pathway open problems, a structured pathway-comparison matrix, and a cross-pathway failure cascade analysis. A hybrid evaluation toolkit operationalises the response-shape via five schema fields applied to three worked examples: Anthropic's Responsible Scaling Policy, Christiano and Xu's Eliciting Latent Knowledge, and the EU AI Act in conjunction with UK AISI evaluation standards. The paper is Track-Strategic in mode: it articulates the struc‐ tural conditions a post-parity response would have to satisfy without claiming to supply the response itself. The closing chapter reframes the open problems as a research agenda for the alignment community.

This paper takes Economic Standing as the condition under which the vision of AI as partner is meaningful at all. Standing is the institutional configuration of recognized claim-position, enforcement-instance and revision-possibility, after Fraser. The growth of AI capacity produces two structurally distinct asymmetries between people and machine-generated value: a cognitive asymmetry that the Containment Paradox has shown to be structurally insoluble after parity, and an economic asymmetry that is institutionally addressable but not automatically politically tractable. Tractability has technical, administrative, fiscal, coalitional, legal, international, and temporal dimensions, any of which can fail; and a further dimension called capture names the structural appropriation of any proposed instrument by the platform, ownership, and governance structures that produce the asymmetry in the first place. The paper operationalises Standing as a Standing Matrix in a four-sublayer-by-three-element form (twelve cells), tests nine candidate instruments — Citizens' Wealth Funds, Negative Income Tax, Individual Civic Capital, Universal Basic Compute, labour law and collective bargaining, the AI Rent Tax Family, Data Extraction Tariffs, Sovereign Compute Infrastructure, and Bridging Tools of the Global South — against the matrix through a cross-level sufficiency test, and distinguishes layer-sufficiency, portfolio-sufficiency, and financing-necessity as three categorically distinct properties. Two load-bearing hypotheses are advanced: a Demographic Camouflage Hypothesis on why Western political signals do not register Standing-erosion, and a Domestication Hypothesis on why the discourse stalls in two zones (internal and geopolitical). The paper engages decolonial-AI literature on the colonial-dividend pattern at substantive depth and develops a co-authorship legitimation condition for transnational extraction-and-redistribution architectures. Phase 5 of the Trncik Framework programme (Containment Paradox / UBE / CUA / BtSV / UAIF / Internalization Problem). Companion to *Internalization Problem* (in preparation) on the cognitive-asymmetry side. An applied demonstration on the Indian IT-services configuration in 2026 is provided in Appendix D.

Much of current alignment research operates under an assumption it rarely examines: that the system doing the containing is at least as intelligent as the system being contained. Red-teaming, proof-checking of alignment-relevant properties, interpretability, and sandboxing all inherit this structure, because their correctness depends on an overseer who can model what the contained system is doing. Non-predictive mechanisms, such as cryptographic commitments, hardware capability caps, and training-time myopia, do not inherit it in the same way, and whether they escape the asymmetry is treated as an open problem in Section 8. The supervisory toolkit works today because the human institutions behind it can, in principle, anticipate the failure modes of the systems they constrain. This paper argues that the parity assumption defines a regime, the pre-parity regime, whose boundary has barely been surveyed, and whose crossing would dissolve most current supervisory approaches rather than stress them. The argument proceeds in three moves. Durable supervisory containment requires prediction. Prediction requires a workable model of the constrained system. And a model adequate to a strictly more intelligent system cannot, in general, be built by the less intelligent one. The claim is scope-bounded: it applies to supervisory mechanisms whose correctness rests on an overseer's capacity at a fixed cognitive level I_H; it does not apply to proposals such as iterated amplification or debate where these preserve alignment, property fidelity, and absence of deceptive mesa-optimization through the amplification chain. Whether any actual amplification scheme achieves all three preservation properties is treated as the central open problem. Subject to that boundedness, the conjecture is that no supervisory apparatus whose effective cognitive level remains at I_H can sustain containment over a system of cognitive level I_A > I_H on relevant axes indefinitely. The paper positions an existing four-paper framework (Understanding Before Ethics, Cognitive Understanding Architecture, Beyond the Stochastic Veil, and the Understanding-Aligned Intelligence Framework) as a coherent response to the pre-parity regime, and sketches what a post-parity alignment orientation would have to contain: structurally internal constraint rather than perimeter-based control, diagnostic evidence rather than behavioral certainty, and institutional oversight whose structure does not assume the contained system can be evaluated on the axes where it strictly dominates. A companion document (Framework Integration Roadmap) records the specific revisions the four framework papers will need in light of the conjecture.

Contemporary artificial intelligence ethics frameworks frequently commit a category error: they attempt to instantiate ethical behavior in systems that, on the account developed here, lack the prerequisite capacity for genuine understanding. This paper argues that understanding, properly conceived as a multi-dimensional cognitive achievement, constitutes the irreducible foundation upon which any meaningful AI ethics must be built. Drawing on Markus Gabriel's Sinnfeldontologie, Hartmut Rosa's resonance theory, Yasuo Deguchi's We-Turn philosophy, and insights from DeepMind's MuZero architecture, we develop a four-dimensional account of the depth of understanding required for authentic moral agency. We distinguish clearly between Ethics-of-AI (E1), concerning governance and oversight of AI systems, and Ethics-by-AI (E2), concerning AI systems as potential moral agents. Our analysis gives reasons to doubt that current AI systems, despite sophisticated behavioral outputs, achieve the ontological grounding, relational responsiveness, social embeddedness, and operational world-modeling that genuine ethical reasoning presupposes. The paper introduces the Cognitive Agent Memory Architecture (CAMA) as a research program for cultivating, rather than guaranteeing, the conditions necessary for AI understanding. We provide operational indicators for each dimension, propose a minimal viable experimentalization framework for empirical investigation, and address major philosophical objections to our position. The paper gives a methodological justification for the four dimensions, treats the relationship to consciousness and embodied cognition, and illustrates the CAMA architecture in practice. This version repairs the paper against its own rules. The claim ledger of §0.3 now reaches the sections that lacked it, the evaluation indicators of §7 record where each one comes from, and the empirical basis of §3.6.1 is corrected and brought up to date. Diagnostic claims about current systems are stated as the reading they are rather than as a measurement. §0.4 lists the corrections and the additions, and a new §12 reports what chapter 7 contains and where it came from.