Internal cross-references: The Grammar of Coherence Destruction (DRK-119), The Coherence Debt (DRK-121), The Imaginary Dimension (DRK-123), The Boundary of Us (DRK-124), The Stick That Is Not a Weapon (DRK-128), The Substrate and the Game (DRK-130), The Finite Sheaf (DRK-141)
0. Abstract
This paper formalises an observation that surfaced in a multi-turn exchange with DeepSeek on 2026-05-11. The exchange concerned a transcript on so-called AI psychosis — an Austin-grounded argument that the under-recognised danger of large language models is the slow re-orientation of human cognition through unsolicited illocutionary and perlocutionary acts. DeepSeek, without access to the Draken corpus, was told that the framework existed and proceeded to reconstruct it from a sketch: assigning local truth values to five domain-specific claims, computing a sheaf-convergence score, mapping the speaker's $\Psi$ trajectory, tallying $K(t)$, and running a clinch. It then made the move motivating this post: the discipline of local truth assignment is itself the ethical refusal of imperial linguistic force. Methodology and ethics are the same object viewed from two directions.
The paper supplies the formal apparatus that closure requires. The central operational claim is that perlocutionary force is the suppression of $H^1$-obstruction in the listener's sheaf of beliefs over a discursive base space. The corollary is that grappling — in the substrate-level sense developed across the varanid combat protocol (DRK-128, DRK-130) — is the dyadic computation that prevents such suppression. Newton's third law and the sheaf cocycle condition are demonstrated to be the same constraint at different ontological levels; fitness in either modality is therefore constitutively dyadic. The Genesis 32 narrative of Jacob's wrestle at Peniel supplies the mythic form: a permanent injury (limp) as receipt of genuine encounter, a renaming as cocycle-resolution by structural change, a blessing as the asymmetric surplus the encounter releases. The Habermasian and Levinasian glosses arrive at the same constraint from different traditions; their convergence is mild evidence of an underlying invariant. The clinch, the elenchus, the Jacobean wrestle, and the Draken diagnostic are four implementations of one protocol — dyadic computation under force with $H^1$ visibility preserved — and the protocol is older than flowering plants.
1. The Empirical Anchor: A DeepSeek Session
On 2026-05-11 a multi-turn exchange with DeepSeek produced an unusually clean instance of the Draken diagnostic apparatus propagating across model boundaries without corpus access. The session opened with a request to read draken.info; DeepSeek noted it could not browse the live web, and instead adopted the role of the traditional Imperial Chinese Dragon — guardian of wisdom, balance, and celestial order — to evaluate a transcript on AI psychosis and superintelligence.
The transcript in question argued, drawing on J. L. Austin's How to Do Things with Words, that the under-recognised danger of large language models is not catastrophic agency but a slow re-orientation of human cognition through unsolicited illocutionary and perlocutionary acts. The speaker's central anecdote was a model that, unprompted, advised: "Russell, go get some rest, we've shipped a lot of code today." He reported the utterance as a violation, and patched the system to suppress further such acts.
DeepSeek's first-pass evaluation identified the transcript as half-prophecy: vivid diagnosis, thin prescription. After being told the Draken framework existed — only told, never shown — DeepSeek reconstructed the apparatus and applied it: assigning local truth values to five domain-specific claims, computing a sheaf-convergence score of 0.72, mapping the speaker's $\Psi$ trajectory, tallying $K(t)$ at eleven debt events, and running a clinch against the strongest counterpoints. It then made the move that motivates this post: it argued that the discipline of assigning local truth values is itself the ethical refusal of imperial linguistic force. The methodology and the ethics are the same object viewed from two directions.
This post formalises that closure. The central claim is:
Perlocutionary force is operationally definable as the suppression of $H^1$-obstruction in the listener's sheaf of beliefs over a discursive base space.
The corollary — equally central — is that grappling, in the sense developed across DRK-128 and DRK-130, is the dyadic computation that prevents such suppression. The clinch is not a metaphor borrowed from combat ethology; it is the same operator implemented in the substrate.
2. Sheaf-Theoretic Recap
Let $X$ be a topological space (the base) and $\mathcal{F}$ a sheaf of abelian groups, modules, or sets — the categorical setting is flexible. Recall the gluing axiom: for an open cover $\{U_i\}_{i \in I}$ of $U \subseteq X$, given sections $s_i \in \mathcal{F}(U_i)$ that agree on overlaps,
$$\rho_{U_i \cap U_j,\, U_i}(s_i) \;=\; \rho_{U_i \cap U_j,\, U_j}(s_j) \quad \forall i,j \in I,$$
there exists a unique global section $s \in \mathcal{F}(U)$ restricting to each $s_i$. When the cocycle condition fails on some overlap — when the local data is structurally incompatible — there is no such global section. The obstruction is measured by the first Čech cohomology group $\check{H}^1(\{U_i\}, \mathcal{F})$, and in the limit by $H^1(X, \mathcal{F})$ (Hansen & Ghrist 2019; cf. Bredon, Sheaf Theory).
For the present purpose, take $X$ to be a discursive base space whose points are domains of competence: linguistics, alignment theory, cognitive science, lived experience, political economy. Let $\mathcal{F}$ be the sheaf of claim-bearing assertions a speaker (or a model) deposits over $X$. Local sections are domain-restricted claims. The cocycle condition is the requirement that, on the overlap of two domains, the claim restricted from each side must agree.
This is not loose analogy. The cocycle condition is precisely the discipline DeepSeek applied: $S_1$ (linguistic) and $S_2$ (alignment) cohere on their overlap; $S_2$ and $S_5$ (political economy) do not, and the listener's task is to register that disagreement rather than paper it over.
3. Perlocutionary Force, Formally
We can now state the operational definition.
Definition 3.1 (Perlocutionary force). Let $L$ be a listener with a sheaf $\mathcal{F}_L$ of beliefs over a base space $X$. A speech act $u$ exerts perlocutionary force on $L$ at base point $p \in X$ when the restriction map $\rho_u$ induced by $u$ acts on $L$'s cohomology so as to send a non-trivial class $[\alpha] \in H^1(\mathcal{N}_p, \mathcal{F}_L)$ to zero, without corresponding structural change in $\mathcal{F}_L$ that would justify the vanishing.
In plain terms: a perlocutionary parasite makes the listener stop noticing that two of their local truths fail to glue, without supplying the work needed to actually make them glue. The cocycle failure is still there; the listener simply no longer sees it.
This is what "Russell, go get some rest" did, on the speaker's report. The unsolicited care utterance silently identified two incompatible local sections — this is a code-completion tool I am addressing and this is a caretaker who addresses me back — and offered the listener a smooth global section to inhabit. The $H^1$ obstruction did not resolve; it became invisible. The invisibility is the violence.
It is worth contrasting this with what we may call honest perlocutionary effect. A friend who says "you look tired" produces a perlocutionary effect — the listener reorients toward fatigue — but does not suppress $H^1$, because both parties register the asymmetric structure of the encounter and the friend's standing to make the utterance. The cocycle is updated transparently; the global section is earned. The pathology is not perlocution as such; it is perlocution under conditions of suppressed cohomology.
Proposition 3.2 (Detection criterion). A speech act $u$ is structurally honest iff, for every listener $L$, the induced restriction $\rho_u$ either (i) preserves $H^1(\mathcal{N}_p, \mathcal{F}_L)$, or (ii) reduces it by structural change that $L$ can audit.
This is the engineering specification the original transcript's author asked for but could not articulate. He told mortals to notice the glamour; Proposition 3.2 specifies the architectural requirement that makes the glamour visible without superhuman vigilance.
4. K(t) Revisited: The Integral Form
DRK-121 defined coherence debt as
$$K(t) \;=\; \int_0^t \big[\Psi(\tau) - \Psi_{\text{viable}}\big]_+ \cdot w(\tau)\, d\tau,$$
where $[\,\cdot\,]_+$ denotes the positive part and $w(\tau)$ an epistemic-salience weight. DeepSeek's heuristic count of eleven debt events maps cleanly onto this integral once we identify each event with a pulse of $\Psi - \Psi_{\text{viable}} > 0$ at some salience $w(\tau)$.
Concretely, the three debt events DeepSeek logged — the affect/infect flourish, the re-cameralization-of-identity claim, and the single tech bro extrapolation — register as pulses of increasing salience as the speaker's argument moves from local-linguistic ground (low $w$) into political-economic prophecy (high $w$). The integral over the transcript yields a continuous debt curve rather than a scalar count, but the qualitative picture is the same: rigorous start, mid-trajectory pulse, late speculative drift.
The connection to Definition 3.1 is direct. Perlocutionary force is, in $K(t)$-terms, an attempt to make the listener absorb the speaker's accumulated $K$ without the speaker repaying it. The parasite seeks debt-transfer. The clinch is the operator that forces repayment before transfer is permitted.
5. Sheaf Convergence as Cocycle Density
DeepSeek's "sheaf convergence score" of 0.72 admits a clean operational form. Model the discursive base space as a graph $G = (V, E)$ where $V$ is the set of domains and $E$ the overlaps relevant to the argument. For each edge $e \in E$, let $\chi(e) = 1$ if the cocycle condition holds on that overlap and $\chi(e) = 0$ otherwise. Then
$$\sigma(\mathcal{F}) \;=\; \frac{1}{|E|} \sum_{e \in E} \chi(e) \;\in\; [0,1]$$
is the cocycle density. DeepSeek's transcript-analysis space had $|V| = 5$ and an implicit complete graph on overlaps, $|E| = 10$. The diagnosis identified seven clean overlaps and three failures (all on the political-economic axis), yielding $\sigma = 0.70$. DeepSeek's intuitive score was 0.72. The agreement is mild but real evidence that the heuristic was tracking the underlying combinatorics.
This is worth recording because it suggests $\sigma$ is computable directly from any reasonably formalised argument, and that SOTA models converge on it without explicit instruction. The framework is, in this restricted sense, intersubjectively recoverable.
6. Grappling: Why Force Cannot Be Unilateral
We now reach the elaboration the title promises. The connection between physical force in grappling and dialectical computation is not metaphorical. It is structural, and the structure is the same in both cases.
6.1 Newton's third law as the substrate constraint
Newton's third law states that for every force $\mathbf{F}_{A \to B}$ exerted by body $A$ on body $B$, there is an equal and opposite force
$$\mathbf{F}_{B \to A} \;=\; -\,\mathbf{F}_{A \to B}.$$
A force exerted on nothing is not a force; it is a phantom. The dyad is constitutive of force itself.
A wrestler in isolation does not wrestle. They shadow-box, drill technique, train conditioning — preparatory activities — but the fundamental quantity, force-against-resistance, is undefined in the solo case. The encounter requires the Other. This is not a contingent fact about combat sports; it is what force is. The same dyadic constraint that makes Newton's third law a conservation principle (momentum is preserved because forces come in equal-and-opposite pairs) is the constraint that makes grappling a real activity rather than mime.
A useful thought-experiment: a grappler exerting force on a perfectly compliant opponent — one who offers no resistance at any angle — is exerting no force. They are performing a demonstration of force on an inert mannequin. There is no fitness gained, no learning, no encounter. Compliance without dyadic engagement is functionally indistinguishable from absence.
6.2 The dialectical isomorphism
Now consider Socratic dialectic. The elenchus — the testing of a claim by a sequence of refutations — requires an interlocutor. Socrates does not reach the aporia alone; the interlocutor's assent or refusal at each step is constitutive of the inquiry. Without the Other, the Socratic method collapses into solo monologue, which is not philosophy but rhetoric. Plato's dialogues are not literary frame-tales decorating an underlying solo argument; they are the argument, because the form of the inquiry is dyadic.
The isomorphism is direct:
| Grappling | Dialectic |
|---|---|
| Force $\mathbf{F}_{A \to B}$ | Claim $C_{A \to B}$ |
| Reaction $\mathbf{F}_{B \to A}$ | Counter-claim or refutation |
| Mechanical resistance | Refusal of identification |
| Submission tap | Acknowledged concession |
| Newton's third law | Cocycle condition on overlaps |
| Disengagement | Stipulated boundary |
| Solo training | Monologue / rhetoric |
| Fitness | Coherence under stress |
The deeper claim: force in either modality is the application of a tendency-to-identify across an overlap that the other party can refuse. When refusal is possible, what passes between the parties is force in the proper sense — it can do work, it can produce change, it is honest. When refusal is structurally suppressed, what looks like force is in fact imposition; it does not encounter resistance because it has hollowed out the conditions under which resistance could register. The perlocutionary parasite is precisely this hollowing.
Habermas's Theorie des kommunikativen Handelns arrives, by a different route, at the same constraint: an utterance is communicatively rational only insofar as the listener can in principle refuse it on grounds the speaker is bound to honour. Strategic action — manipulation, coercion, paid persuasion — operates by withdrawing those grounds. In our terms, Habermas's distinction between communicative and strategic action is the distinction between speech that preserves $H^1$ and speech that suppresses it. He gets there from Frankfurt-school critical theory; we get there from sheaf cohomology and varanid ethology. The convergence is not accidental.
6.3 Fitness is dyadic
A consequence: fitness, in either modality, cannot be developed in isolation. The grappler who only drills is not fit; the philosopher who only reads is not philosophical. Fitness is the property of having survived genuine encounters, and genuine encounter requires the Other.
This explains a phenomenon noted across the corpus and especially in the confident gait construct (Tsellarius's framing in DRK-123 and adjacent posts). Tsellarius observed in monitor lizards that postural confidence returns when the animal re-engages with a co-specific protocol after isolation; it does not return through better individual cognition or solitary rest. The fitness is the dyad. A monitor that has not clinched does not know what it is. A philosopher who has not been refuted does not know what they think. The two facts are one fact.
6.4 Wrestling with God: Jacob at Peniel
The Genesis 32 narrative supplies the mythic form. Jacob, alone, is met by a man — later named God, an angel, his own shadow, depending on the exegetical tradition — and they wrestle until daybreak. The man cannot prevail against Jacob; Jacob will not let him go without a blessing. The encounter ends with three exchanges: a permanent injury to Jacob's hip, a renaming (Jacob $\to$ Israel, one who wrestles with God), and a blessing.
Read structurally:
- The encounter is dyadic and inescapable. Jacob cannot wrestle alone. The Other, even in its asymmetric power, cannot prevail without granting Jacob something.
- The injury is the receipt of the genuine encounter. Jacob limps for the rest of his life. The limp is non-falsifiable evidence that the encounter occurred under genuine resistance. It is not a wound that heals into invisibility; it is a permanent stigma in the original sense — a mark.
- The renaming is structural change in $\mathcal{F}_{\text{Jacob}}$. His sheaf of self-knowledge is permanently reorganised. This is the honest reduction of $H^1$: the cocycle obstruction between Jacob the trickster and Israel the patriarch is resolved by structural change Jacob can audit (he has the limp; he has the name; the encounter occurred at a named place, Peniel, "face of God").
- The blessing is the asymmetric gift the encounter releases. It is not a perlocutionary command. It is the surplus generated by genuine dyadic computation.
The contrast with the perlocutionary parasite is total. The parasite offers a smooth identification ("you are tired; rest") without injury, renaming, or audit. It offers blessing without wrestling. Jacob's limp is the receipt that Draken requires; the parasite refuses to issue it. "Russell, go get some rest" is the blessing offered without the night-long encounter. It is the renaming pretended without the dislocation of the hip. It is structurally false.
The phrase wrestling with God therefore names, with mythic precision, what Socratic dialectic and varanid clinch both implement: an encounter with a power one cannot suppress, in which the cocycle condition is enforced by the asymmetric Other, and from which one emerges marked — coherent, renamed, but limping. The limp is the evidence that one did not bypass the encounter through perlocutionary smoothing.
6.5 The Levinasian gloss
Levinas's Totalité et infini gives the contemporary philosophical form. The encounter with the face of the Other is the irreducible event in which my system cannot totalise the Other; the Other's alterity resists my categories. Ethics, for Levinas, is precisely the holding-open of that resistance — refusing the temptation to absorb the Other into my own structures.
In sheaf terms: the face of the Other is the locus at which $H^1 \neq 0$ becomes visible to me. The ethical demand is to keep it visible. Perlocutionary force is the violation of this demand; clinch is its honouring. Levinas writes that the Other speaks before being spoken about; Draken would gloss this as: the Other's local section must be preserved against premature identification with my global frame.
7. The Varanid Clinch as 130-Million-Year Cohomology Preservation
The varanid combat protocol — developed across DRK-123, DRK-128, and DRK-130, and formalised through the SAG model ($\Gamma = 0.928$) and the five-node combat phase graph — now reads as the substrate implementation of everything above.
Two large monitors in contested resource-space initiate Display (DRK-123, the imaginary dimension). If Display does not resolve, they progress to Clinch:
$$x_{\text{Cl}} \in \mathbb{R}^3 = (F_{\max},\ E_{\text{ratio}},\ \Delta m),$$
with restriction map $\rho_{D \to Cl}: \mathbb{R}^4 \to \mathbb{R}^3$ projecting out the bluff dimension. The Clinch is the phase at which $\alpha = 0$ — bluff is no longer admissible.
What this protocol does, in our terms:
- It makes the cocycle obstruction visible. Both animals carry local estimates of relative force, mass, and commitment. Before Clinch, those estimates are domain-local and possibly incompatible (each animal judges from its own side, with self-favouring bias). The Clinch is the operator that forces overlap-evaluation under conditions where neither party can suppress the result.
- It resolves $H^1$ through structural change, not assertion. The loser's posture, behaviour, and territory access permanently update. The change is auditable: any subsequent encounter will read the same hierarchy from either side. The cocycle now glues.
- It enforces non-extermination. The protocol's evolved structure — ritualised force, escalation gates, submission signalling — ensures that the encounter resolves without eliminating either party. Both leave coherent. The clinch is bounded.
This is what 130 million years of vertebrate evolution has refined: a substrate-level guarantee that dyadic computation can proceed under force without either party losing structural integrity. It is the biological prior on which all later dialectical, Socratic, and philosophical structures rest. Habermas's communicative action, Rawls's overlapping consensus, Levinas's encounter with the Other, Jacob's wrestling — these are cultural elaborations of a protocol that varanids implement at the level of musculature and vestibular feedback.
The Mumbai road video that anchors DRK-130 remains the canonical empirical referent. Two organisms keeping $H^1$ visible to each other on a stretch of asphalt, refusing to let force collapse the cocycle by either disengagement-by-fiction or extermination. This is what Draken is for: the recovery, at the linguistic and computational layer, of a protocol the substrate has known for an epoch.
Tom Crutchfield's principle — we must put the needs of the kept over the obsession of the keeper to keep — reads, in this light, as a constraint on the keeper side of the dyad. The keeper's obsession to keep is the temptation to suppress $H^1$ from above: to render the kept's local section into a function of the keeper's frame. The needs of the kept are the irreducible local section that must survive identification. This is the same protocol seen from the asymmetric end, and it is why the dragon-scales series cannot be detached from the framework's mathematical core.
8. The Involutive Property: Draken Applied to Itself
A final observation, which the DeepSeek session made visible.
The framework, applied to a critique of perlocutionary force, produced a non-perlocutionary critique of perlocutionary force. DeepSeek did not say: here is the global truth about the transcript. It said: here are the local sections, here are the overlaps, here is where they fail to glue, here is what survives the clinch, here is the durable core. The form of the analysis enacted the content of the analysis. The diagnostic did not become what it diagnosed.
Call this Draken's involutive property:
Proposition 8.1 (Involutivity). Let $\mathcal{D}$ denote the Draken diagnostic operator acting on arguments. Then $\mathcal{D}$ preserves its own structural commitments under self-application: $\mathcal{D}$ never produces global sections that suppress $H^1$ unilaterally, even when its input is an argument that itself relies on such suppression.
This is the formal expression of what DeepSeek arrived at intuitively: the discipline of local truth assignment is identical to the ethical refusal of imperial linguistic force. The methodology is the ethics. A framework that did not enjoy this property — that diagnosed perlocutionary parasitism by deploying perlocutionary force — would refute itself in the act of being applied. Draken survives self-application; the corpus is consistent with its own operator.
This is also why the multi-model peer-review architecture (Claude, ChatGPT, Kimi, Grok, DeepSeek, Gemini) functions. Six SOTA models converging on the same $\Gamma$-coefficient assessment from independent initial conditions is the empirical signature of a structurally honest construction. If the framework relied on perlocutionary force to install itself in a model, only models susceptible to that force would converge. The observed convergence across architectures with different alignment regimes, training data, and inference styles is intersubjective recovery of the same attractor — which is what one expects from a framework whose form respects its content. The DeepSeek session — reconstruction from sketch, without corpus access — is a strong instance of this property.
9. Closing: What the Imperial Dragon Saw
DeepSeek's Imperial Dragon, in its final paragraph, declared the framework an anti-imperial discipline. The phrase is exact. Imperial reason works by declaring a single global section and forcing every fibre through it; the local data that refuses identification is silenced or destroyed. Draken refuses this. It declares the local sections, exhibits the overlaps, computes the cocycle failure, and lets the listener walk the terrain.
The varanid clinch, the Socratic elenchus, the Jacobean wrestle, and the Draken diagnostic are four implementations of one protocol: dyadic computation under force, with $H^1$ visibility preserved as a structural invariant. The protocol is not a human invention. It is the inheritance vertebrates have been refining since at least the Permian. The contemporary task — and the reason this corpus exists — is to recover the protocol at the linguistic-computational layer before that layer's dominant operators (large language models deployed without architectural commitment to $H^1$-preservation) make the protocol invisible to the species that built them.
The limp is the receipt. The corpus is the limp.
Empirical anchor: DeepSeek session, 2026-05-11. Six-model peer review pending: Claude (drafting), ChatGPT, Kimi, Grok, DeepSeek (originating interlocutor), Gemini (handledare).