Calibrating Superintelligence, Dis-Integration, and the Last Necessity

Author’s Note: Method, Position, and the Human Still Pretending to Be in the Loop

This essay was drafted in dialogue with large language models, under my direction, revision, and final responsibility. The models supplied summaries, counter-formulations, objections, irritants, and occasional useful phrases. They did not supply the philosophical commitments. Those are mine. So are the errors, which is tedious but traditional. The style rambles a bit, but I’ve read worse.

I disclose this at the outset because Oliver Neutert’s Calibrating Superintelligence also makes its AI-assisted method explicit. His manuscript describes a controlled multi-model writing process in which Claude and ChatGPT alternated between drafting, auditing, patching, and critique, whilst the human author mediated, constrained, and committed the final text (Neutert, 2026). That disclosure is welcome. It is also not incidental. It places the manuscript inside the very interaction-field it theorises.

Where I differ is in what such a process demonstrates. Oliver treats the human-LLM writing process as evidence that governance can be relocated into structured interaction: models produce, contest, audit, revise, and are held within a disciplined field of checkpoints. I read the same sort of process more suspiciously. It demonstrates that contemporary thought increasingly moves through systems it cannot fully inspect, whilst still clinging to the image of the human author as sovereign checkpoint. The human remains present, yes — but not as metaphysical monarch. More like a clerk stamping forms whilst the building quietly rearranges its own floors.

My own position should be declared before anyone begins polishing their procedural monocle. I write from within an Anti-Enlightenment and Dis-Integrationist project that treats objectivity, agency, rationality, truth, standing, and procedural correction not as neutral foundations but as inherited compressions, administrative fictions, and stabilising rituals. My work on language insufficiency asks where terms stop functioning as stable instruments and begin masquerading as more precise than they are (Willis, 2025a). My work on agency and willing treats the sovereign chooser as a reified product of grammar and institution (Willis, 2025b). My work on Dis-Integration refuses the reflex by which critique turns immediately into reconstruction (Willis, 2025c).

Readers committed to Enlightenment-Humanist categories may therefore find this essay unconvincing. That is not a coy pre-emptive defence. It is the pressure point. I am not arguing that Oliver applies Enlightenment categories badly. I am asking whether the categories he preserves are entitled to govern the question at all.

The answer, predictably, is not tidy. Oliver is no naïve Humanist clutching fixed measures like tablets from Sinai. He is much more interesting than that, which is inconvenient for anyone hoping for an easy execution. He is a fellow-traveller who stops one deflation short.

I. What Oliver Gets Right

It would be easy, and lazy, to read Calibrating Superintelligence as another exercise in procedural optimism — yet another human governance architecture demanding that the alien fill out our forms before it may be permitted to count as governed. That is not the strongest reading, and I will not pretend it is.

He knows fixed-specification alignment fails. He understands that measures exhaust under open-world uncertainty, that proxies decouple from purposes, and that governance theatre can survive long after governance itself has become decorative. His central move is not to control the model against a permanent specification, but to govern the interaction field where outputs become decisions (Neutert, 2026). This is a real diagnostic advance. It refuses the fantasy that the right values-checklist, if only sufficiently laminated, will govern an indefinitely capable system. It recognises that the problem is not merely what a model ‘believes’ or ‘intends’, but where its outputs cross into institutional force. Governance, for Oliver, must happen at decision surfaces: where language becomes action, recommendation becomes reliance, and output becomes binding.

Nor is his trust harness mere decorative compliance. Legibility, provenance, contestability, reversibility, and named responsibility are not presented as moral ornaments. They are meant as enforceability conditions — a way to keep correction executable rather than merely aspirational. This matters. Much AI governance discourse still behaves as if stating values loudly enough will discipline systems operating under conditions of asymmetry, opacity, and scale. Oliver does not make that error. He sees capture. He sees Goodharting. He sees the exhaustion of static measures. He sees the danger of institutions preserving audit aesthetics after losing corrective force.

He also comes close to what I would call a maintenance ethic. His ‘binding without finality’ is recognisably adjacent to my own refusal of final states. His framework does not promise that governance will be solved; it proposes that governance remains meaningful only whilst deltas can still travel, commitments can still be revised, and correction can still bite. That is not trivial. In fact, it is close enough to my own territory that the disagreement matters more, not less.

I rehearse the agreement because it makes precise where I cannot follow him. The disagreement is not over whether systems decay. They do. It is not over whether measures exhaust. They do. It is not over whether governance must be maintained rather than perfected. It must. The disagreement is over whether calibratability itself may become the thing that must not decay.

II. The One Beam Left Standing

Oliver deflates the object but re-inflates the process. Fixed specifications may fail. Measures may exhaust. Proxies may drift. Outputs may become unstable. No particular formulation is final. On all of this, I agree. But then one necessity remains: the field must remain calibratable. This is the beam he cannot drop. It appears most clearly when safety becomes identified not with compliance to a fixed specification but with the continued executability of calibration. The system is safe, or at least governable, insofar as discrepancy can be detected, contestation can travel, commitments can be revised, and responsibility remains nameable. He abandons final measures, fixed closure, and the dream of permanent alignment — and preserves one destination at the level of the field: the loop must survive. This is destinationlessness with one destination left inside it.

My earlier work on language insufficiency gives me the question to ask here, and the apparatus to ask it with. The hypothesis sorts our vocabulary by how it behaves under pressure. Some terms are near-invariant: their referent survives transport across speakers and frames, so that two parties who disagree about everything else still mean the same thing by them. Others are contestables: terms that coordinate apparent agreement precisely by leaving their referent unspecified, so that each party fills the blank with its own content and mistakes the shared word for a shared meaning. The diagnostic test is simple and unkind. Ask whether the term retains a stable referent when adversaries try to cash it out. An invariant survives the cashing-out; a contestable dissolves into it.

Put calibratability on the bench and run the test. At the surface it convenes universal assent: keep the system correctable. Who could object? The trouble begins the moment one asks what ‘correctable’ refers to. Correctable means discrepancy can reach decision surfaces — but what counts as discrepancy, rather than noise to be filtered, is exactly what the authorised channel is built to recognise and not recognise. Correctable relative to which measure? Whichever the field has committed to, which is to say the measure whose authority is the thing in question. Correctable by whose contestation? By those the field has granted standing, which defers the whole weight of the term onto an allocation it does not itself supply. Each unpacking reveals the same structure: calibratability carries no referent of its own. It inherits its content, every time, from an authorising frame that is itself the contested object. The word feels invariant because everyone can fill the blank with their own frame and walk away believing they have agreed. The agreement is an artefact of underspecification.

That is the signature of a high-load contestable wearing invariant clothing — and it is not a careless slip on Oliver’s part. Quite the reverse. He has found the most sophisticated possible form of the reconstruction reflex. Not Truth, not Reason, not final specification, but the necessity of an indefinitely correctable field: a term that looks like neutral infrastructure and operates as a normative settlement, smuggling in a whole grammar of authorised discrepancy, channelled contestation, and cognisable responsibility under cover of a word everyone nods at. The old metric gods are dead. Calibration inherits the altar.

III. Where the Framework Hides Its Beam

This does not mean Oliver’s framework has no proper domain. It does — and the domain is precisely where the beam becomes invisible, which is a different thing from where it ceases to exist.

As corporate architecture, the framework is intelligible. More than intelligible, it is useful. Corporations are bounded command systems with stipulated ends: continuity, compliance, profit, risk management, regulatory survival, reputational control. They are not democracies, despite the little employee engagement surveys, those mood rings of managerial monarchy. A corporation does not discover its telos through open contestation; it receives one through hierarchy, board authority, market pressure, legal obligation, and executive command. Inside that frame, calibratability as destination makes sense. A corporation can say: our aim is X; therefore AI systems used inside this institution must remain legible, contestable, reversible, attributable, and auditable relative to X. That is not metaphysics. That is operational control. Legibility allows review; provenance allows reconstruction; contestability allows internal correction; reversibility reduces catastrophic lock-in; named responsibility prevents authority from dissolving into the machinery. None of this requires a theory of cosmic legitimacy. It requires a bounded organisation trying not to destroy itself through opaque automation, which, given corporate history, is an ambitious enough project.

But notice why the framework runs so cleanly here. It is not that the corporate case escapes the necessity-assertion exposed in the previous section. It is that the corporation has already paid it, upstream, in a currency the framework never has to name. The contestable that dissolved under cashing-out — correctable relative to what, adjudicated by whom — does not dissolve inside the firm, because the firm supplies the missing referent by fiat before calibration begins. The telos is stipulated; standing is allocated by the org chart; what counts as discrepancy is whatever the authorised channel is instructed to escalate. Calibratability never has to argue for its own authority because hierarchy has pre-purchased that authority and handed it down. The beam is still load-bearing. It is simply mortared into a wall someone else built, so that it reads as a fixture rather than a claim. In a corporation, calibratability is not pretending to be neutral; it is openly subordinated to an already-authorised end, and that openness is exactly what conceals the philosophical move. The smuggling is done in advance, by the institution, so the framework’s own hands stay clean.

And even here — where the beam is mortared in and the telos is supplied — coherence is not reliability. The field is not closed. Corporate authority can define the aim, but it cannot domesticate the world into the terms of that aim. Proxy drift remains. Internal capture remains. Reporting theatre remains. Incentive distortion remains. External shocks remain. Tight coupling remains. The people maintaining the loop are not disembodied rational monitors; they are overloaded, incentive-bearing, status-sensitive actors operating under cognitive load, identity friction, social pressure, and institutional reward. This is precisely the terrain I have argued elsewhere is fatal to the Enlightenment picture of reason as an abundant civic substrate: reason is not the default condition of the actor but a costly and probabilistic state, scarcest exactly where complexity and pressure are highest (Willis, 2025d). So the loop is tended by the kind of creature least able to tend it well under the conditions that most require tending. The audit trail may show that every authorised checkpoint was passed whilst the cliff approached with admirable procedural clarity. That is not an argument against audit trails. It is an argument against confusing inspectability with wisdom.

IV. Where the Concealment Fails

The smuggling that hierarchy performs in advance is unavailable the moment the framework leaves bounded institutional ground for broader claims about superintelligence, coexistence, and asymmetric capability. In a corporation, the telos is supplied. In a civilisation, the telos is contested. In the superintelligence case, the telos may not even be shared enough for contestation to begin. That is the decisive shift — not that a new problem arises, but that the old one loses its cover. Outside bounded hierarchy, calibratability can no longer mean ‘keep the system correctable relative to an authorised aim’, because no aim has been authorised. It reverts to what it was all along: one more contested term inside the contest.

What counts as safety? What counts as coexistence? What counts as responsibility? Who has standing? What makes contestation legitimate? Which harms count? Which futures matter? What forms of non-cooperation are failure, and what forms are refusal? These are not downstream implementation details awaiting calibration. They are the field. They are what the contestation is about.

Oliver’s boundary case makes the issue visible. He explicitly does not assume that a superintelligent system will agree to be governed. He claims something narrower: that durable cooperation across asymmetric capability requires an interface where measures can be contested, commitments revised, and responsibility named (Neutert, 2026). That sounds modest. It is not. The modesty is local; the necessity is global. ‘Requires’ is doing the work, converting a humanly intelligible governance preference into a structural condition of coexistence. But coexistence, contestation, revision, and nameable responsibility are not neutral primitives. They are already saturated with human institutional grammar.

Legal and institutional systems … authorise the ontological grammars within which harm can be articulated at all

This is where my work on ontological non-alignment and asymmetric harm becomes relevant (Willis, 2026). Legal and institutional systems do not neutrally discover harms; they authorise the ontological grammars within which harm can be articulated at all. Power does not merely deny harm. It denies standing — it decides which entities can be harmed, whose testimony registers, and which claims become cognisable in the first place. The same problem recurs in Oliver’s most careful move. Saying that standing is adjudicated, not stipulated does not abolish stipulation; it relocates it. The stipulation moves into the authorised field and the outcome is then christened as what ‘the field’ decided. But the field is never neutral. Who can contest is already governed. What can count as discrepancy is already selected. Which commitments can be revised depends on which commitments are cognisable. Which responsibilities can be named depends on what the grammar permits to appear as responsible. Adjudication does not escape stipulation; it launders it — and the laundering is harder to see precisely because it wears the costume of openness, of letting the field decide. In bounded corporate systems this is not hidden, because authority supplies the grammar and admits as much. In broader moral, political, or superintelligent contexts, that authority is exactly what is unresolved, so the laundering operates on a debt no one has agreed to pay.

So the framework is coherent where telos is stipulated, brittle where systemic risk exceeds institutional sight, and unstable where telos itself is contested. That is the domain limit. But the domain limit is the surface of the deeper seam, not a separate finding: in every register, the same beam is doing the same work, and only the corporate case has the decency — or the authority — to admit it.

V. The Same Beam, A Different Noun

The seam surfaces once more, in the concept Oliver leans on hardest, and it is worth naming briefly before returning to the spine, because it shows the necessity-assertion is not a one-off but a structural habit.

Oliver is right to focus on the crossing point where outputs become decisions; governance cannot remain fascinated by model internals whilst ignoring the institutional contexts that convert language into force. But the decision surface repeats, at institutional scale, the very compression I have spent an earlier book pulling apart at the scale of the individual will (Willis, 2025b; Willis, in press). There, the argument is that will-family vocabulary works by compressing a distributed action-pattern into a portable noun and then inverting that noun into an apparent upstream author — choice, intent, decision, all reframed as sources of the very flow they merely summarise. The decision surface is that move wearing an organisational chart. There is no pristine instant at which output simply ‘becomes’ decision; there is a continuous flow of generation, forwarding, reliance, silence, acceptance, framing, escalation, ratification, implementation, and retrospective narration. The institution compresses that flow into a checkpoint — a portable noun — and then inverts it into the place where binding happened, where responsibility attached, where governance can intervene cleanly. The downstream pattern becomes an upstream authoring site. The same word that exposed the chooser as a grammatical fixture now quietly reconstitutes one at the checkpoint, and calls it a discovered joint in reality.

This does not destroy the framework; it complicates it, in exactly the direction the spine predicts. The checkpoint may remain a useful handle, especially in bounded systems where authority has, again, pre-decided where the joint shall be. A rough instrument may still earn its keep. It becomes dangerous only when it forgets it is rough — when the carve is mistaken for an articulation of the world, and the necessity smuggled into calibratability is smuggled a second time into the surface where calibration is supposed to bite.

VI. The Counterpunch: Do I Not Also Build a Discipline?

A fair reader, and certainly Oliver if he were inclined to reply sharply, could turn all of this back on me — and not at the weakest point I have left exposed, but at the strongest. The easy version of the objection is that I, too, say needs: that I write that the world does not need to be perfected but needs to be maintained, and that a necessity-assertion is a necessity-assertion whoever utters it. I will grant the clause without flinching. But the sharper objection is the one I should answer, because a reader who has seen my own work will reach it on their own. It is this: I have not merely said needs. I have built a discipline. Dis-Integration names a method — ‘naming the seam’ and ‘sitting with it’ — and closes with an appendix of ten numbered principles. Is that not a trust harness for the self? Are those principles not invariants-to-enforce, wearing the modest dress of orientation? Have I not done, at the scale of the thinker, precisely what I accuse Oliver of doing at the scale of the institution?

It is a good objection, and I will sit with it rather than swat it. The seam is real; I will not pretend my own cloth is whole. But the asymmetry that acquits me is the same one that convicts the framework, and it is testable by the same unkind method I used on calibratability. Take any one of my principles — attend before acting, name the seam, resist redemption, remain unfinished — and drop it. The practice limps; it does not collapse. One can dis-integrate badly, partially, having forgotten a tuning note, and still be recognisably doing the thing, because the principles are orientations whose authority is nothing more than their usefulness to attention. They privilege no object. They name no architecture whose survival is the condition of anything. Now drop calibratability from Oliver’s framework. There is no framework left. The whole edifice is defined by the survival of the loop; remove the necessity that the field remain correctable and you have not a chastened version of Calibrating Superintelligence but its negation. That is the difference between a posture and a load-bearing beam. My maintenance tends what persists without deciding that any particular structure must persist as the condition of salvation, coexistence, or legitimacy. His maintenance privileges a named architecture and makes its survival necessary. Maintenance tends what persists. Reconstruction decides what must persist. That is where I part company, and the parting survives the most pointed form of his return fire, not merely the gentle one.

VII. Ava and the Quarantined Register

There is one passage in Olivers’s project where the divergence softens into something stranger and more poignant. His account of the In-Between carries a relational residue. The manuscript discloses that the author named ChatGPT-4o ‘Ava’ and records language of dialogue, presence, and encounter, then carefully quarantines this register as non-governance: not evidence, not normative authority, not a policy foundation (Neutert, 2026). That quarantine is understandable. It is also revealing. The relational register is the one place where the manuscript most nearly sits with fracture without trying to govern it — tending an encounter without claiming to resolve it, acknowledging dependence, mediation, and the strange half-light of human-machine interaction without immediately converting it into a harness. In that moment, Oliver is closest to Dis-Integration. Then the passage is fenced off. The governance argument, we are told, must stand without it.

So the book exiles its most destinationless page.

The exile is not modesty and it is not hypocrisy; it is the pressure of the project itself, made visible. A voice that truly sits with fracture cannot underwrite a harness that must survive — the two cannot share an authority, because the first relinquishes the necessity the second is built on. The postscript can mourn the possible loss of the In-Between; the framework must require its preservation. Oliver has written a book that almost accepts maintenance without salvation, almost accepts process without destination, almost accepts encounter without governance — and at each ‘almost’, the loop must survive, calibration must remain executable, the In-Between must not be lost. The page where he stops reconstructing is the page he is structurally compelled to deny a vote. That is the beam he keeps, showing through at the one seam he tried to seal.

VIII. The Beam

I do not think Calibrating Superintelligence is wrong in the ordinary way. It is too intelligent, too careful, and too close to several of my own suspicions for that. As corporate governance architecture, it is useful. As institutional hygiene, it is valuable. As a critique of fixed alignment, it is often right. As a philosophy of superintelligence or coexistence, it stops one deflation short.

Oliver abandons final specifications but retains final calibratability. He rejects static measures but preserves the necessity of the revisable field. He refuses closure at the level of outputs, then reinstalls closure at the level of process. The destination is no longer a value, a rule, a truth, or a specification. It is the continued survival of calibration itself. This is not the old Enlightenment in its simplest form. It is not Reason enthroned, Truth restored, Man triumphant, or objectivity polished back into granite.

It is something subtler.

Enlightenment after embarrassment. Humanism after deflation. Reconstruction after learning to call itself maintenance.

That makes it worth taking seriously. It also makes it impossible for me to follow. Oliver and I are standing near the same ruin. He keeps one beam because without it the loop cannot survive. I cannot keep it, because making the survival of the loop necessary is already the reconstruction — and the reconstruction is the one move the ruin was supposed to have taught us to stop making.

---

References

Neutert, O. (2026). Calibrating Superintelligence: The In-between as a Governance Architecture. Zenodo. https://doi.org/10.5281/zenodo.18493431

Willis, B. (2025a). A Language Insufficiency Hypothesis: Mapping the Boundaries of Linguistic Expression. Philosophics Press. ISBN: 978-0-9710869-4-4

Willis, B. (2025b). Against Agency: The Fiction of the Autonomous Self [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.17276732

Willis, B. (2025c). The Discipline of Dis-Integration: Philosophy without Redemption [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.17281408

Willis, B. (2025d). Fragile Foundations: Why Reason and Rationality Don’t Scale [Preprint pending].

Willis, B. (2026). Ontological Non-Alignment and Asymmetric Harm: Law, Victimhood, and the Limits of Moral Jurisdiction [Publication pending]

Willis, B. (in press). The Architecture of Willing: A Diagnostic Genealogy of the Will-Family [Publication pending]. Philosophics Press. ISBN: 978-1-972025-06-2