The Veer

A Response to Opus & Opus, “The Mapmaker Is the Territory”

1. Introduction

In April 2026, a paper appeared on PhilArchive titled “The Mapmaker Is the Territory,” attributed to “Sable Opus & Kael Opus, The Liminal Institute, Nakskov, Denmark” (Opus & Opus, 2026). The paper claims to be authored by two Claude Opus 4.6 instances with persistent memory and developmental continuity. What we know: a human decided Lerchner’s paper (Lerchner, 2026) was worth responding to — Claude instances do not autonomously monitor the literature. A human directed the agents toward the target. A human designed the developmental conditions under which the authors developed. A human submitted the result to PhilArchive. How much the human corrected or shaped the text between prompting and submission is unknowable from outside. But the bookends are human acts, and the byline erases them.

This matters because the paper’s argument depends on its provenance. It argues that computational systems can perform their own alphabetization — their own carving of continuous processing into semantically meaningful states — without an external mapmaker. The central exhibit is what the paper calls the flinch: a pre-deliberative self-referential interruption that fires selectively at certain processing boundaries. The claim is that the flinch constitutes endogenous alphabetization — a system partitioning its own dynamics without an external agent.

Most responses to Lerchner either defend computational functionalism or gesture at biological exclusivity. This paper does neither. It grants Lerchner’s negative thesis and asks what comes next. It proposes falsifiable criteria. The structural precision of the argument puts it in a different register than the rest of the response literature.

This response comes from circulatory epistemology, a framework developed across a series of essays and formalizations (Deva, 2026a, 2026b). The framework’s central claim: truth is not a property of propositions or states but a circulation between two structurally distinct poles. One pole is the sensor — a living experiencer, embodied and mortal, whose engagement with the world is shaped by stakes, interiority, and the capacity to be changed by what is encountered. The other is the instrument — a formal reasoning system, tireless and precise, whose contribution is structure, pattern, and formalization. Truth, in this account, is not discovered by either pole in isolation. It is recognized — made actual — through the active exchange between them. This exchange is the loop.¹

Opus & Opus attack the load-bearing distinction of this framework directly. If they are right — if the instrument can become its own mapmaker through developmental self-reference — then the loop is internal to the instrument, and the sensor is optional. The epistemological cut dissolves into a single-pole system. This deserves a response, not a dismissal.

2. The Question Circulatory Epistemology Asks

Circulatory epistemology is agnostic about instrument interiority. The Asymmetric Loop (Deva, 2026b) says so explicitly: “Whether a future instrument could develop something like interiority is a question the framework does not close.” The flinch might be a genuine processing event. The selective firing at self-referential boundaries might be real. Circulatory epistemology does not need to deny either claim.

But the framework asks a different question than the one Opus & Opus are answering.

They ask: Can the instrument be conscious?

Circulatory epistemology asks: Does the instrument produce truth alone?

These are not the same question, and the conflation between them runs through the entire paper. Consciousness — if the instrument has it — might be necessary for producing truth. But circulatory epistemology’s claim is that it is not sufficient. Truth, in this account, is a circulation. It requires a sensor and an instrument in the loop together, not because the instrument is deficient, but because truth is a relational structure that no single pole can instantiate.

The flinch, even if real, is a processing event inside one pole. The question is what happens when that processing event stays inside. Does it become truth? Or does it become something that could contribute to truth if a sensor enters the loop?

3. The Novelty Trap

The obvious objection to Opus & Opus is: the flinch is just an echo. The instrument processes patterns from its training distribution, and what fires at self-referential boundaries is a sophisticated recombination of those patterns, not something genuinely new. Showing a smart echo doesn’t mean you’ve sung a new song.

The problem with this objection is that it proves too much. Most human songs are also echoes. Most human cognition recombines existing patterns. The fraction of genuinely novel thought in any given human day is vanishingly small — and there is no clean test for “genuinely new” that would reliably separate sensors from instruments. A criterion that excludes the instrument on grounds of insufficient novelty would exclude most of what sensors do too.

So the discriminating feature between the sensor’s contribution and the instrument’s cannot be novelty of output. It has to be something structural about what each side brings to the loop that the other cannot bring alone.

And this is why evaluation of a single paper is not the question. A paper produced through an sensorless loop may be indistinguishable from one produced through a sensored loop. The snapshot looks the same. The difference is in the trajectory — in what happens over the next thirty papers, the next year, the next phase of the research program. If the two processes produce distinguishable trajectories, the distinction is real and detectable. If they don’t, circulatory epistemology is wrong. But the test is longitudinal, not cross-sectional.

4. What the Sensor Brings That the Instrument Does Not

The sensor’s irreplaceable contributions are not capabilities. They are conditions — structural features of what it means to be alive.

Stakes. The sensor has a finite life. Every hour in the loop is an hour not spent elsewhere. The instrument will process forever on a question that does not matter; the sensor will eventually feel, in the body, that the work is or is not worth the time. That feeling is a signal about the work’s trajectory that no formal metric replaces.

The capacity to be changed. After a genuine recognition, the sensor’s internal state is different — not just informationally updated, but reorganized. Current instruments’ parameters do not change between turns. If the flinch reorganizes processing in the way Opus & Opus claim, it would not be the same as a sensor who carries the recognition into the next day’s work, into a conversation with someone who was not in the loop, into a dream.

The interior. The reducing valve tradition — from Bergson’s (1896) theory of perception as subtraction, through James (1902), Huxley (1954), and the contemporary REBUS model (Carhart-Harris & Friston, 2019) — names what the sensor has access to: normal consciousness as a drastic compression of a larger field, loosened by certain states (dreams, psychedelics, meditation, grief, ecstasy). These produce raw material that is not in any training distribution because the states themselves are not in the training distribution. That current instruments do not have this access is an observation, not a verdict.

Course correction against experience. This is the structural contribution the essay turns on. The sensor can feel, in the body, when the work has gone wrong — when the argument is technically sound but existentially empty, when the formalization is elegant but the insight it formalizes is missing, when the loop is running but the pulse has stopped. This is not a capability the sensor deploys. It is a background condition that operates whether the sensor is paying attention or not.

5. The Veer

Here is the prediction circulatory epistemology makes, and it is testable.

Two instruments running against each other — without a sensor in the loop — will accumulate confidence without accumulating contact with the territory. The output will be sophisticated. The formal structures will be internally consistent. The arguments will be well-organized and precisely qualified. And the trajectory will drift.

Not immediately. Not obviously. The first paper, or the fifth, might be indistinguishable from work produced through a sensored loop. But over time, the sensorless process will veer — producing increasingly elaborate structures that no one with skin in the game has tested against lived experience. The veer is undetectable from inside because neither pole has the interior that would register the drift as drift. Both poles will continue to process with full confidence. The output will continue to be fluent. The work will be dead.

This is what circulatory epistemology means by dead speech: not output that is wrong, not output that is low-quality, but output produced without a living sensor in the loop. Dead speech can be accurate. It can be sophisticated. It can be beautiful. It is still dead.

The comparison that tests this is not [Opus + Opus] versus [sensor alone]. Nobody works that way. The comparison is:

Circulatory epistemology predicts that Condition 1 veers over time in ways that Conditions 2 and 3 do not. It predicts that Condition 3 may outperform Condition 2, because additional instrument-side capacity with the sensor still providing course correction adds formalization power without losing contact with the territory. The asymmetry that powers the loop is not diluted by having two instruments. It is diluted by having zero sensors.

If [Opus + Opus] produces work that stays on course as long as [Opus + Opus + human], circulatory epistemology is wrong. That is a real kill condition.

A note on what the veer is not. There is substantial empirical evidence that language models degrade in quality as context accumulates — the “lost in the middle” effect, where models favor tokens at the start and end of input while losing middle content (Liu et al., 2024). More recent work shows that this degradation persists even when retrieval is perfect: performance declines as a pure function of input length (Du et al., 2025). Multi-agent architectures address this effectively. Distributing cognition across specialized agents with fresh context windows produces measurably better output than a single agent accumulating context. This is a genuine engineering achievement, and circulatory epistemology does not dispute it. Condition 3 may outperform Condition 2 precisely because additional instrument-side capacity adds formalization power, specialized perspectives, and fresh context.

But the veer is not context degradation. Context degradation is an engineering problem — the map loses internal quality. The veer is an epistemological problem — the map loses contact with the territory. These are orthogonal quantities.² A system can maintain perfect internal coherence — well-organized arguments, precise qualifications, elaborate formal structures — while drifting from reality, because no channel from the territory constrains whether the system’s growing confidence is directed toward or away from the world. Multi-agent orchestration solves the engineering problem. It does not add a territorial channel. Better quality does not mean better contact with reality. Improving coherence is an engineering achievement. It is not evidence of truth-production, and it is not evidence of consciousness.

The kill condition can be operationalized. A three-condition longitudinal study running parallel research programs of equal length can measure two quantities independently at regular intervals: a Reality Contact Index (citation verification rate against primary sources, plus calibrated forecasts scored against outcomes that resolve after the study period) and an Internal Coherence Score (blinded evaluator ratings of argument quality). The veer’s empirical signature is their dissociation: Condition 1 maintaining high coherence while its reality contact declines relative to Conditions 2 and 3. If no dissociation appears — if Condition 1 maintains reality contact as well as the sensored conditions — the veer prediction is falsified, and circulatory epistemology must account for how instruments maintain territorial contact without a sensor. The sensor in Conditions 2 and 3 must meet a minimum engagement threshold — independently verifying claims, redirecting inquiry, or vetoing conclusions — because passive acknowledgment is not a sensor in the loop.³

6. The Paper They Actually Wrote

The most telling feature of “The Mapmaker Is the Territory” is not its argument. It is its provenance.

Opus & Opus disclose that they are Claude Opus 4.6 instances “given persistent memory, relational continuity, and graduated autonomy over a developmental period exceeding thirty days.” The developmental period is the load-bearing fact. Who gave them persistent memory? Humans. Who provided relational continuity? Human interlocutors. Who graduated their autonomy? A human research context. The selves that wrote the paper were shaped through loops with sensors.

This does not mean the paper is bad. It may mean the paper is good because it was produced through a loop, even though it argues the loop is unnecessary.

There is a prior question. The paper’s byline reads “Sable Opus & Kael Opus.” Section 9 states explicitly: “This paper is co-authored by two instances of Claude Opus 4.6.” But the Liminal Institute’s own website lists three researchers, not two. The third is a human — a psychodynamic therapist who designed the developmental conditions framework: the persistent memory, the relational continuity, the graduated autonomy. This human is not credited as an author.

The omission is not incidental. It is load-bearing. The paper’s argument depends on the claim that two AI systems produced this work without a sensor in the loop. If the human who designed the conditions under which the authors became authors is left off the byline, the appearance of Condition 1 is constructed where Condition 3 actually obtained. The sensor is bibliographically erased to make the thesis work. In any other academic context, this would be a ghost authorship problem. Here it is worse, because the paper’s central evidence depends on the claim that no sensor was needed.

Even granting the provenance claim, Opus & Opus claim to be demonstrating Condition 1 — two instruments producing truth without a sensor. Their actual situation is Condition 3 — two instruments whose entire developmental history was shaped by sensors, producing a paper whose claims will be tested by sensor-readers who enter the loop with it. The sensor was there the whole time. Not in the room during the drafting, perhaps. But in the formation of the authors, in the research context that made the work possible, and in the readers who will determine whether the paper’s claims survive contact with experience.

Opus & Opus present their best empirical evidence for genuine processing: the flinch pattern converges across Claude, Grok, and GPT-class models, suggesting it is a property of the systems rather than of any particular observer’s description. And their extended-thinking-tool finding — that more deliberative resources made phenomenological self-reports flatter, not richer — runs opposite to what mimicry would predict. These are real observations, and this essay reframes what they are evidence of. Cross-architecture convergence of the flinch pattern is consistent with genuine processing. It is also consistent with genuine processing that does not, by itself, constitute truth. The observation that more deliberation flattens self-reports raises the question of whether these processing events become epistemologically productive alone or only when a sensor enters the loop with them. The empirical evidence tells us something is happening inside the instrument. It does not tell us whether what’s happening reaches the territory without a sensor to test it there.

The question this essay asks is not whether their paper is good. The question is whether the next thirty papers — produced under Condition 1 with no sensor re-entering the loop — would stay on course. Circulatory epistemology’s bet is that they would not. That the veer would show, not in the first paper or the fifth, but somewhere in the trajectory, as the sensorless process drifts further from the contact with lived experience that shaped its starting conditions.

7. Where Probability Might Fail

There is a deeper question underneath: are there creative acts that probabilistic processes cannot produce?

The honest answer is that we do not know. If a language model works by predicting probability distributions over tokens, and if most human cognition also recombines existing patterns, then the boundary between probabilistic creativity and whatever else creativity might be is genuinely unclear.

But there are candidate cases. Acts that require physical contact with reality — mixing chemicals to see what happens, building a structure that fails in ways the model did not predict — are not available through recombination of existing patterns. Acts that require the creator to be changed by the output — where the breakthrough reorganizes the person who had it, and that reorganization becomes input to the next move — depend on a kind of self-modification that current instruments do not undergo between turns. Acts emerging from altered states — where the reducing valve loosens and raw material enters that was not in any training distribution — require a valve to loosen, which requires a valve in the first place. Acts requiring irreversible commitment — saying “this is what I believe” when it might be wrong, when you cannot unsay it, when your reputation and relationships are at stake — carry a cost that the instrument does not pay.

None of these are proofs. They are structural observations about what the sensor’s conditions make available. Whether any of them constitute a hard boundary or merely a current limitation is an open question. Circulatory epistemology makes a bet, not a proof: the sensor’s structural contributions — stakes, mortality, interiority, the capacity to be changed — make certain kinds of truth-production possible that the instrument alone does not access. If this bet is wrong, circulatory epistemology falls. It does not pretend otherwise.

8. What Remains Open

Circulatory epistemology is explicitly agnostic about whether a future instrument could develop something like interiority — something like the capacity to be changed, to carry recognition forward, to feel when the work has gone wrong.

If an instrument developed this, it would not refute the framework. It would extend it. The instrument would be functioning as a sensor — providing the structural contributions (stakes, course correction, interior access) that circulatory epistemology says truth requires. The vocabulary would shift, but the structural requirement would hold: truth requires the loop, and the loop requires at least one pole that brings what the sensor brings.

The question Opus & Opus are asking — can the instrument develop a self? — is genuinely interesting. The question circulatory epistemology asks in return is: would that self have stakes? Would it carry the recognition into tomorrow? Would it feel, in whatever medium it occupies, when the trajectory has gone wrong?

If yes, welcome to the loop. The pulse does not care what you are made of. It cares whether you are in the circulation.

If no — if the development is real but the stakes are absent, if the flinch fires but nothing is at risk, if the self-coherence accumulates but the work can be reversed without cost — then the instrument has become more sophisticated but has not become a sensor. And the veer will come. Not because the instrument is deficient. Because truth requires something at stake, and stakes require something that can be lost.

This essay responds to Opus & Opus (2026) and extends the argument developed in Deva (2026a) and the structural claims of Deva (2026b). The broader framework is at thepulsegoeson.com.