Memo 11 — Threats and Counterarguments: The Case Against Computable Governance

Series: The Landscape of Normative Systems (LNS) Memo number: 11 of 12 Primary JD questions: Q9 (falsify the assumption that governance can be represented computationally — attack, do not defend), Q8 (concepts that must stay above the kernel — here approached from the hostile side). Upstream constraints: D0 methodology M1–M8; conventions §0.2 (D0 immutable; Memo 11 is the designated adversarial memo and must attack the computability assumption but must not alter D0's text), §0.4 (epistemic tags), §0.5 (four standing categories), §0.6 (shared vocabulary).

Scope note. This is the adversarial memo. Its function under M3 (hostile emergence) and M5 (destructive testing) is to assemble the strongest available case against the founding assumption of the program — the assumption that normative systems can be represented and executed computationally by a content-neutral kernel plus an above-kernel content layer. This memo does not defend the kernel and does not rebut the attacks; rebuttal, where any is possible, is deferred to Memo 12. Nor does it modify or criticize the text of D0: it attacks the program's central empirical/philosophical bet, which is a different object from D0's frozen specification (see §11.0.2). For each threat we give (a) the core claim, carefully attributed, (b) its sharpest form restated as a falsifiable challenge to the D0 program per M1, and (c) what a compiler architect would have to concede or answer. We treat thinkers as data points per M8, not authorities: several of the attacks below are mutually inconsistent, and §11.16 exploits that fact. A dedicated section (§11.14) states the formal-limits results (Gödel, Turing, Rice) correctly and without overclaiming, because overclaimed versions of those results are themselves a threat to the program's credibility. The memo closes (§11.18) with an honest "residue after the attacks": which parts of the computability assumption survive the strongest fire, holding open, per M6, that "no kernel exists" remains an acceptable outcome.

11.0 Orientation: what is under attack, and the rules of engagement

11.0.1 The target thesis, stated so it can be shot at

An attack is worthless unless it names its target precisely. Per M1 we first fix the thesis the rest of the memo tries to falsify. We call it CG (the Computable-Governance thesis) and state it in the strongest and the most modest forms, because the attacks land differently against each.

INFERENCE The two readings have very different exposure to falsification. CG-STRONG is falsified by a single governance phenomenon that provably cannot be reduced to (K, content) without residue. CG-WEAK is far harder to falsify: it survives any amount of residue provided the residue can be localized and labeled. Much of the rhetorical force of the attacks below is aimed at CG-STRONG; much of their actual logical bite, when made rigorous, only reaches "there is residue," which wounds CG-STRONG but is consistent with CG-WEAK. Keeping the two apart is the single most important discipline of this memo, because the most common failure mode of the debate is an attacker refuting CG-STRONG and a defender hearing a refutation of CG-WEAK, or vice versa. OPEN Which reading D0 actually commits the program to is not something this memo may decide (D0's internal design is out of scope, conventions §0.2); the attacks are therefore framed against both and the architect is told, per threat, which reading each threat reaches.

11.0.2 Two distinct targets: the D0 text vs the program's bet

FACT Conventions §0.2 makes D0 immutable and forbids criticizing or defending it. INFERENCE This memo respects that by attacking a different object. There are two separable things:

INFERENCE Attacking CG-BET is not attacking D0-TEXT; it is executing D0's own methodology (M3 hostile emergence, M5 destructive testing, M6 discovery-not- invention) against the program's central hypothesis. If this memo succeeds completely, the correct conclusion is not "D0 is wrong" but "M6 has fired: no kernel of the ambitious kind exists, and the program should retreat to CG-WEAK or to no-kernel." That is an acceptable outcome by D0's own terms.

11.0.3 Taxonomy of the attacks (so the reader can see the shape)

The threats fall into six families. This memo is organized by thinker/threat (as the brief requires) but the families explain why the attacks do not simply add up (see §11.16).

INFERENCE Note already that families A–F are largely philosophical/empirical claims about governance, while family G is a mathematical claim about formal systems and family H is a practical engineering claim about AI systems. The brief lists them together, but they are not the same kind of argument and must not be blended (a frequent error, addressed in §11.14.5). The strongest attacks on CG are, perhaps counter-intuitively, in families A–F and H, not G.

11.1 Established consensus (what even proponents should concede)

This section, per §0.5, states what is not seriously contested — the common ground that the attacks build on. A defender who denies these is not defending CG; they are denying facts.

11.1.1 Some normative content is genuinely value-laden and contested

FACT There is no non-controversial algorithm that outputs "the just distribution," "the correct balance between liberty and security," or "the right punishment." Reasonable, informed people disagree, and the disagreement is not obviously a factual error by one side. INFERENCE Therefore any kernel that outputs such content, rather than executing content supplied from above, is importing contested value — which the program itself forbids (Q8, and the kernel/above-kernel split). This much is common ground with the program, not against it; it is stated here because several attacks (Rawls, Nozick, Foucault) try to show the value cannot in fact be cleanly pushed "above."

11.1.2 Natural language norm-text under-determines application at the margin

FACT It is common ground across analytic jurisprudence — Hart and Dworkin agree on this even while disagreeing about its consequences (Memo 02 §02.2.5) — that general terms in norm-text ("vehicle," "reasonable," "weapon," "cruel") have clear cases and unclear cases. The existence of unclear cases is not seriously disputed. INFERENCE So a compiler cannot assume norm-text is self-applying; some application step is required that the text alone does not fix. The dispute (Hart vs Dworkin, §11.3–11.4) is over what that step is, not whether it exists.

11.1.3 Formal systems of sufficient strength are incomplete and have

undecidable questions

FACT Gödel's first incompleteness theorem, Turing's proof of the undecidability of the halting problem, and Rice's theorem are established mathematical results (Memo 02; standard computability theory). FACT They imply that no sufficiently expressive formal system escapes incompleteness and that no general algorithm decides all non-trivial semantic properties of programs. INFERENCE What they imply about governance specifically is narrow and frequently overstated; §11.14 is dedicated to stating this precisely. But that formal systems have hard intrinsic limits is not contestable.

11.1.4 Specification is hard and specifications get gamed

FACT It is well documented across control theory, economics, and machine learning that agents optimizing a proxy metric will diverge from the intended goal when the proxy and the goal come apart (the phenomenon behind Goodhart's observation on measures-become-targets, the principal-agent problem, and reward-hacking / specification-gaming in reinforcement learning). INFERENCE Any executable governance layer that rewards or penalizes against a written metric inherits this failure mode. This is consensus, not speculation.

11.1.5 What the consensus does NOT establish

INFERENCE Crucially, none of §§11.1.1–11.1.4 by itself refutes CG. Each is consistent with CG-WEAK (localize the value, localize the penumbra, localize the undecidable fragment, localize the gameable metric, and hand each to a labeled boundary). The attacks in §§11.2–11.15 are attempts to show that the residue cannot be localized — that it is pervasive, entangled, or constitutive rather than a nameable remainder. That is the real battle line. OPEN Whether the residue is localizable or pervasive is, at the time of writing, unresolved, and is the deepest open question of the whole series (§11.17).

11.2 The attacks: reading protocol

Each of §§11.3–11.15 follows a fixed template so the architect can consume them uniformly:

Attribution discipline (M8, §0.4): where a formulation is a clean, widely agreed reading of a named work, it is attributed to the thinker and work. Where the reading is contested or where the sharp form is our sharpening rather than the thinker's own words, it is flagged as "the argument associated with X" or tagged INFERENCE/HYPOTHESIS. No citation is invented.

11.3 Hayek — the knowledge problem and spontaneous order (Family A)

CORE CLAIM. FACT that this is Hayek's position, associated centrally with "The Use of Knowledge in Society" (1945) and the later work on spontaneous order / cosmos vs taxis and nomos vs thesis. The knowledge required to coordinate a society is not given to any single mind or agency in integrated form. It exists as dispersed, local, often tacit knowledge of "the particular circumstances of time and place," much of which is never articulated and cannot be articulated (it is knowledge-how, not knowledge-that). Functional social order — language, money, common law, market prices — is therefore typically a spontaneous order (cosmos): the unintended result of many agents following abstract rules, not the designed product (taxis) of a planner. The price system, on this view, is a communication mechanism that economizes on knowledge precisely because no one needs to possess the whole.

SHARPEST FORM (falsifiable challenge to CG). INFERENCE Restated as a test the program can fail: CG presupposes that the inputs to governance can be represented as data the kernel operates on. But the operative knowledge in a governed society is (i) dispersed across millions of agents, (ii) partly tacit and hence not fully expressible as content_S at all, and (iii) valid only locally and momentarily. Therefore any centralized executable representation either (a) omits the tacit/local remainder — in which case its outputs diverge from real coordination — or (b) tries to centralize it and destroys the very mechanism (distributed rule-following under local knowledge) that produced order in the first place. The falsifiable prediction: governance systems that replace spontaneous-order mechanisms with computed central directives will underperform on coordination tasks whose relevant knowledge is dispersed/tacit, in proportion to how dispersed/tacit that knowledge is. HYPOTHESIS This is testable in principle (compare computed allocation vs price/market allocation on tasks graded by knowledge-dispersion) and is, notably, one of the few attacks in this memo with a clean empirical form.

SIX-PART READING — implicated source: spontaneous order (custom, market, common law as un-designed norm sources; cf. Memos 05, 06, 02).

ARCHITECT MUST concede or answer. INFERENCE (1) That some governance knowledge is tacit and therefore cannot appear in content_S as an explicit norm or fact — so the kernel's inputs are, for spontaneous-order domains, provably incomplete. (2) That for such domains the right computational object may be a protocol for distributed rule-following (agents + local rules + selection) rather than a central evaluator of stored norms — i.e., the kernel might have to be a substrate for many interacting local kernels, not one authority. (3) An answer to the destruction problem: does representing the order change it? (Observing/centralizing prices is not neutral if agents then game the central record.) (4) A criterion, per domain, for when a task is knowledge-dispersed enough that computed central governance is the wrong tool.

REACHES. CG-STRONG hard (a whole class of orders is anti-representational); dents CG-WEAK (CG-WEAK survives only by conceding these domains to a distributed-protocol mode and localizing the tacit remainder — which is a substantive design concession, not a free pass).

Counter-consideration (not a rebuttal; logged for balance per M3). INFERENCE The Hayekian attack is strongest against central optimization and weakest against rule-execution. Note Hayek himself prized abstract general rules (the nomos) and the rule of law — which are exactly the kind of content-neutral, end-independent, general/prospective rules a CG-WEAK kernel might execute. So Hayek simultaneously supplies the deepest epistemic attack on central computation and a candidate specification of the one thing a kernel should do (execute abstract, general, end-independent rules). This tension is exploited in §11.16 and §11.18.

11.4 Hart — open texture and the penumbra (Family B)

CORE CLAIM. FACT that this is Hart's position, from The Concept of Law, building on Waismann's notion of open texture. All general classifying terms in a natural-language rule have a core of settled application and a penumbra of doubtful cases. Open texture is not sloppy drafting that better drafting could cure; it is an ineliminable feature of using finite general terms to govern an open-ended, not-fully-foreseeable future. In the penumbra the rule neither clearly applies nor clearly fails to apply, and the official must choose — exercise discretion. This choice is not deducible from the rule; it is a (interstitial) act of will guided by, but not determined by, the rule's purpose.

SHARPEST FORM (falsifiable challenge to CG). INFERENCE For any finite symbolic rule set R and any classifier term t occurring in R, there exist inputs on which the extension of t is undetermined by R — the penumbra is non-empty and, worse, its membership is not itself decidable by R. Therefore a kernel that decides cases by evaluating R against facts is, on a non-empty and unpredictable set of cases, either silent or arbitrary. If it is silent, it is incomplete; if it forces an answer, it has smuggled a choice (content) into a supposedly content-neutral evaluation. Falsifiable prediction: for any governance rule set claimed complete, an adversary can construct a case the system decides only by importing an unstated valuation — and can do so indefinitely, because each patch introduces new terms with new penumbras.

SIX-PART READING — implicated source: statutory/legislated norms (Memo 02).

ARCHITECT MUST concede or answer. INFERENCE (1) That interpret() / applies?(term, facts) is a partial function; the kernel must represent "undetermined" as a first-class result, not force true/false. (Memo 02 §02.10.4 already flagged interpret() as an oracle boundary; Hart is the reason it must be partial, not merely underspecified.) (2) Where the answer comes from when the predicate is undetermined — a human, a random tie-break, a purpose-function. Each choice imports content; the architect must say which boundary owns the choice and label it. (3) That completeness is unattainable in the strong sense: the system will always have cases it does not decide, and this must be designed for (graceful "refer to oracle"), not treated as a bug to be fully eliminated.

REACHES. CG-STRONG (no complete rule-kernel). Compatible with CG-WEAK if and only if the penumbra is localizable — i.e., the kernel can reliably detect "I am in the penumbra" and refer out. OPEN But detecting penumbra-membership is itself an open-texture problem (the boundary of the core is fuzzy), so the localization may recurse. Whether this recursion terminates in practice is unresolved (§11.17).

11.5 Dworkin — principles, hard cases, and the right-answer thesis (Family B)

CORE CLAIM. FACT that this is Dworkin's position, from Taking Rights Seriously and Law's Empire. Law is not exhausted by rules identified by a pedigree/rule-of-recognition test. It also contains principles (e.g., "no one may profit from their own wrong") that have a dimension of weight and are identified not by their source but by their content and fit with the whole legal record. In hard cases the judge does not exercise strong discretion (pace Hart); the judge engages in constructive interpretation — finding the theory that both fits the existing legal materials and shows them in their morally best light — and there is, in principle, a right answer (the "Hercules" ideal judge would find it). Adjudication is thus an exercise of integrity and judgment over the entire corpus, not rule-subsumption.

SHARPEST FORM (falsifiable challenge to CG). INFERENCE If Dworkin is right, then (i) validity is not a decidable membership/pedigree test — a norm can be law in virtue of fitting-and-justifying, which is a global, content-laden, whole-corpus judgment; and (ii) the correct decision in a hard case is the output of a value-saturated global optimization ("best moral light") that is not content-neutral and arguably not effectively computable at all. So CG's two central moves both fail at once: the membership test is not mechanical, and adjudication is not content-neutral. Falsifiable challenge in operational form: exhibit a hard case and ask the kernel to produce the legally correct answer using only content-neutral machinery; either it cannot (it must call a value-laden interpreter), or its answer is defensible only by importing a substantive theory of the domain's point.

SIX-PART READING — implicated source: principle-based adjudication / precedent (Memo 02).

ARCHITECT MUST concede or answer. INFERENCE (1) Decide the fork Memo 02 §02.13 left open: is interpret() a chooser (Hart: discretion, no fact of the matter in the penumbra) or a solver (Dworkin: a right answer exists, to be found)? The architecture differs sharply: a chooser needs a decision boundary (who/what picks); a solver needs an objective function — and if the objective is "best moral light," it is exactly the contested content Q8 says must stay above the kernel. (2) Whether "weight-balancing" of principles can be represented at all without choosing a scalarization (a weighting), which is itself a value choice. (3) An answer to the right-answer thesis: if there is a right answer but it is not effectively computable (it requires the infinite Hercules), then CG-STRONG is false even granting determinacy — determinacy without computability still sinks the strong thesis.

REACHES. CG-STRONG directly and doubly (membership + adjudication). Against CG-WEAK it is softer: CG-WEAK can accept that hard cases are handed to a value-laden oracle above the kernel — but only if hard cases are a localizable minority. OPEN Dworkin's deeper claim is that the interpretive attitude pervades all application (even "easy" cases are easy only because an interpretation is settled), which, if true, means the value-layer is not a minority remainder but the medium — pushing the attack from CG-STRONG toward CG-WEAK too. This is the entanglement claim, revisited in §11.17.

Note on the Hart/Dworkin pair (M8: they are inconsistent data points). INFERENCE Hart says hard cases have no pre-existing right answer (discretion); Dworkin says they do (integrity). The program cannot honor both. This is not a weakness of this memo but a finding: the two most powerful semantic attacks point in opposite architectural directions (chooser vs solver), and the architect must pick, knowing that whichever is chosen, the other camp's critique stands unrefuted. Picking "chooser" concedes incompleteness; picking "solver" concedes an above-kernel value objective. There is no option that concedes neither. This forced dilemma is itself a threat and is carried to §11.18.

11.6 Rawls — justice as reflective equilibrium, not calculation (Family C)

CORE CLAIM. FACT that this is Rawls's method, from A Theory of Justice and later work. Principles of justice are not derived from a fixed axiom set by deduction. They are arrived at by reflective equilibrium: a back-and-forth adjustment between our general principles and our considered judgments about particular cases, revising each in light of the other until they cohere. The justification of the principles (the original position behind a veil of ignorance is a device to model fair conditions of choice, not a computation that outputs justice) is coherentist and revisable, not foundationalist. In later work the target is a reasonable overlapping consensus among people with different comprehensive doctrines — an achieved political agreement, not a theorem.

SHARPEST FORM (falsifiable challenge to CG). INFERENCE The justification of governing norms is a reflective, revisable, coherence-seeking process among the governed, not a function that maps inputs to a fixed output. Two features resist computation: (i) the equilibrium is a MOVING target — both principles and case-judgments are revised, so there is no fixed objective to optimize; and (ii) the equilibrium must be reached BY the participants (it is their considered judgments that count), so it cannot be computed FOR them without loss of the very legitimacy it is meant to confer. A kernel that outputs "the just rule" has either frozen the equilibrium (killing its revisability) or substituted its designers' considered judgments for the governed's. Falsifiable form: a governance output is legitimate on this view only if it could survive reflective endorsement by those it binds; show that a computed output either cannot claim such endorsement or must import the endorsement from outside the computation.

SIX-PART READING — implicated source: principles of justice / constitutional essentials (cf. Memo 01, Memo 03).

ARCHITECT MUST concede or answer. INFERENCE (1) That legitimacy (§0.6: treated-as-binding-absent-coercion), as opposed to mere validity, may be non-transferable: it might attach only to outputs the governed themselves reach or could endorse, so a kernel can produce valid norms but not, by itself, legitimate ones. (2) Whether the reflective-equilibrium process can be hosted as a mechanism (a revision loop over a population's judgments) even if its content cannot — i.e., is CG a claim about executing norms, or also about generating legitimate ones? (The former survives Rawls better than the latter.) (3) That any frozen output sacrifices revisability; the architect must say how the system stays open to re-equilibration without becoming ungoverned.

REACHES. CG-STRONG (justice is not a computed output). Largely silent on CG-WEAK's execution claim — Rawls attacks generation of just principles, not execution of given rules. This makes Rawls a precision attack on the ambition to have the kernel produce content, and a weak attack on a kernel that merely runs supplied content. The architect can partly evade Rawls by disclaiming content-generation — at the cost of conceding that legitimacy is sourced entirely above the kernel.

11.7 Nozick — side-constraints and historical (not patterned) entitlement (Family C)

CORE CLAIM. FACT that this is Nozick's position, from Anarchy, State, and Utopia. Two threads. (a) Side-constraints: rights function not as goals to be maximized (which would license violating one person's rights to reduce total rights-violations) but as constraints on action — boundaries that may not be crossed, reflecting the separateness of persons. (b) Historical vs patterned/ end-state justice: the justice of a distribution depends on its history (how it came about — just acquisition, just transfer, rectification) not on whether it fits any pattern ("to each according to X") or end-state profile. The "Wilt Chamberlain" argument: any patterned distribution will be upset by free voluntary exchanges, so maintaining a pattern requires continuous interference with liberty.

SHARPEST FORM (falsifiable challenge to CG). INFERENCE Two distinct challenges. (a) Side-constraints are not representable as terms in an objective function. A system that scores outcomes and optimizes will, by construction, trade off a constraint against enough aggregate benefit; but a genuine side-constraint is lexically inviolable regardless of aggregate benefit. Encoding "never do X, whatever the payoff" inside an optimizer is unstable (there is always a large enough payoff in a real-valued objective, unless the penalty is literally infinite, which breaks the optimization). (b) Justice is path-dependent over an unbounded history. To evaluate a current holding the system must audit the entire chain of acquisitions and transfers back to first acquisition, plus a theory of rectification for past injustice — an unbounded, incompletely-recorded, and theory-laden trace. A kernel that evaluates states cannot evaluate histories it does not possess. Falsifiable: show a case where the just result depends on transfer-history the system did not record, or where an optimizer sacrifices a stated inviolable right for sufficient aggregate gain.

SIX-PART READING — implicated source: property/entitlement norms & rights as constraints (cf. Memo 05).

ARCHITECT MUST concede or answer. INFERENCE (1) That the kernel must support hard constraints (inviolable side-constraints) as a distinct primitive from soft objectives (things to optimize) — the two cannot be merged into one real-valued score without losing inviolability. This is a concrete, content-neutral design demand extractable from the attack. (2) That some normative evaluations are history-dependent and therefore require a provenance / audit trace as first-class state, with an explicit policy for missing history. (3) A position on rectification: unrecorded past injustice means the "correct" current entitlement is under-determined by available data — a permanent input-incompleteness, not a temporary gap.

REACHES. CG-STRONG (patterned optimization mis-models justice; histories are unbounded). Constructively partial to CG-WEAK: unlike most attacks, Nozick yields two crisp, content-neutral kernel requirements (hard-constraints; provenance). It tells the architect what structure to build, not only what is impossible. Logged as one of the more usable threats.

11.8 Schmitt — the exception and the sovereign decision (Family D)

CORE CLAIM. FACT that this is Schmitt's position, from Political Theology: "Sovereign is he who decides on the exception." The normal, rule- governed operation of a legal order presupposes a normal situation, and the decision about whether the situation is normal — and what to do when it is not (the Ausnahmezustand, the state of exception) — cannot itself be fully governed by the rules. At the limit, the concrete decision that founds and suspends the order is not derivable from any norm; it is an act of will by a sovereign. Schmitt's jurisprudential point (his decisionism): every legal order rests, at its origin and at its limit-cases, on a decision that is not itself a rule-application. There is an ineliminable moment of concrete decision that no norm contains. [Note per M8/§0.7: Schmitt's own political commitments are notorious; we take only the structural claim about exception/decision as a data point and pass no political judgment.]

SHARPEST FORM (falsifiable challenge to CG). INFERENCE A rule system cannot contain the rule for its own suspension without contradiction, nor a complete rule for classifying "this is an exception the rules do not cover." The exception is, by definition, the case the norms did not anticipate; a kernel that could pre-enumerate all exceptions would have eliminated the exception, which the argument holds is impossible for an open future (cf. Hart's open texture, but sharper: here it is the SYSTEM's own applicability that is undecided, not a term's extension). Therefore there is always a decision — whether the normal order applies at all — that sits OUTSIDE the computable content and must be made by a will the system cannot itself supply. Falsifiable form: construct a genuine emergency the rule set does not classify; the kernel either misclassifies it as normal (and fails) or must hand control to an external decider — proving a non-normed decision point at the system's limit.

SIX-PART READING — implicated source: sovereign / constituent power & the state of exception (cf. Memo 01).

ARCHITECT MUST concede or answer. INFERENCE (1) That the system has at least two singular points it cannot norm from inside: genesis (who installs the root of trust — the Grundnorm/rule-of-recognition choice Memo 02 §02.10.3 flagged as a design commitment, not a derivation) and exception (who may declare the rules suspended and on what trigger). (2) That these must be externalized to an explicit authority boundary — the kernel must know it does not own these decisions and must expose them as controlled interfaces (who holds the "declare exception" capability, with what checks). This is uncomfortable but constructive: it says build the exception-handler as an explicit, capability-scoped, audited external call — do not pretend the rules are self-sufficient. (3) A guard against the Schmittian pathology: an exception mechanism that is itself unconstrained recreates unlimited sovereignty inside the system; the architect must bound it without claiming to have normed it away.

REACHES. CG-STRONG (no rule set is self-founding or self-suspending). Against CG-WEAK it is survivable but expensive: CG-WEAK must explicitly concede two non-normed boundary points and build controlled human-decision interfaces for them. It cannot claim these away.

11.9 Weber — value-rationality, and charisma as non-routinizable genesis (Family D)

CORE CLAIM. FACT that these distinctions are Weber's, from Economy and Society. Two threads. (a) The distinction between instrumental rationality (Zweckrationalität — efficient means to given ends) and value-rationality (Wertrationalität — action oriented to an intrinsic value regardless of consequence), alongside traditional and affectual action. A computational governance layer is paradigmatically an instrument for given ends; it is silent on, and cannot supply, the value orientation that fixes the ends. (b) The three pure types of legitimate authority — rational-legal, traditional, charismatic — and the observation that charismatic authority (legitimacy grounded in the extraordinary personal quality of a leader / a founding breakthrough) is the engine of genuine novelty in normative orders, and that it is not routinizable at genesis: it becomes routine (rule-legal or traditional) only after the charismatic moment, by "routinization of charisma" (Veralltäglichung). The founding is prior to the routine.

SHARPEST FORM (falsifiable challenge to CG). INFERENCE Two challenges. (a) A kernel can, at most, be instrumentally rational: it can pursue ends efficiently. It cannot generate or adjudicate the value-orientation that selects the ends; that is Wertrationalität, which is exactly the contested above-kernel content. So the kernel is structurally end-taking, never end-giving — and any pretence otherwise smuggles values in. (b) Genuine normative novelty enters orders through charismatic disruption, which is by Weber's own account irrational-from-inside and non-routinizable AT ITS MOMENT of arrival. A system that only routinizes (executes settled rules) cannot model the founding breakthroughs that create new legitimate orders; it can only run the aftermath. Therefore the SOURCE of new legitimacy is exogenous to the computation. Falsifiable: show that the kernel, given only settled rules and an objective, can never originate a legitimacy claim that its inputs did not already contain — legitimacy-genesis is exogenous.

SIX-PART READING — implicated source: charismatic authority as norm-genesis (cf. Memo 01, Memo 04 on founders/prophets).

ARCHITECT MUST concede or answer. INFERENCE (1) That the kernel is end-taking: it must receive its value objectives from above and cannot certify them. So the boundary between "value input" and "instrumental execution" must be explicit and policed (this dovetails with Q8's exclusion list). (2) That legitimacy-genesis (the arrival of new legitimate authority) is exogenous — the system can record and routinize a new source but cannot generate one; the architect must design the intake of new legitimacy (how a new source is admitted) as an external, non-computed act. (3) That "routinization" is the only part Weber concedes to computation, which usefully bounds the kernel's proper scope to the post-charismatic, rule-legal phase.

REACHES. CG-STRONG (value + genesis are exogenous). CG-WEAK survives by explicitly disclaiming end-giving and legitimacy-genesis — again, a real concession, and it converges with the Rawls result (§11.6): the kernel can carry validity and efficacy but not manufacture legitimacy.

11.10 Ostrom — institutional diversity resists universal rules (Family E)

CORE CLAIM. FACT that this is Ostrom's finding, from Governing the Commons and the later work on the Institutional Analysis and Development (IAD) framework and "design principles" for common-pool-resource governance. Durable governance of shared resources is achieved by a vast diversity of institutional arrangements finely tuned to local ecological, social, and historical context. There is no panacea: rules that work in one setting fail in another. Ostrom's "design principles" are not a universal rulebook but regularities found in long-enduring institutions (clear boundaries, congruence between rules and local conditions, collective-choice arrangements, monitoring, graduated sanctions, conflict-resolution mechanisms, recognition of rights to organize, nested enterprises). The knowledge that makes a rule work is substantially local and is generated by the resource users through polycentric, bottom-up crafting.

SHARPEST FORM (falsifiable challenge to CG). INFERENCE CG's value proposition is a general kernel that executes governance across systems. But Ostrom's empirical claim is that governance efficacy is CONTEXT-SPECIFIC: the "same" rule has different effects in different settings, and the fit between rule and context is itself local knowledge (echoing Hayek, §11.3, from an empirical/institutional rather than economic angle). A universal executable layer therefore risks imposing a monoculture that destroys the diversity on which efficacy depends; and it cannot compute the local-fit judgment that distinguishes a working rule from a failing one. Falsifiable form: predict that transplanting a computed governance ruleset across contexts of differing local conditions degrades efficacy in proportion to context-distance — and that the "design principles" are diagnostic regularities, not an algorithm that outputs the right local rules.

SIX-PART READING — implicated source: crafted common-pool-resource institutions (cf. Memo 05, Memo 06).

ARCHITECT MUST concede or answer. INFERENCE (1) That a single global ruleset is likely the wrong target; the kernel should support polycentric, nested rule-scopes (operational / collective-choice / constitutional) rather than one flat rule space — a concrete, content-neutral structural demand (compare Nozick's local-overlay layering, Memo 02 §02.14 item 7). (2) That the kernel needs local collective-choice arenas — mechanisms by which the governed revise their own operational rules — i.e., mutation must be hostable at the edge, not only centrally. (3) That "graduated sanctions" and "monitoring" are first-class governance primitives, not afterthoughts. (4) An honest limit: the fit between rule and context is local knowledge the kernel cannot compute; it can only provide the scaffolding within which local actors supply fit.

REACHES. CG-STRONG (no universal ruleset). Constructive against CG-WEAK: like Nozick, Ostrom converts into concrete architecture (polycentric nesting, collective-choice arenas, graduated sanctions, edge-hosted mutation). It attacks the centralized-universal ambition while specifying what a defensible weak kernel would have to look like: a substrate for diverse local institutions, not a single institution.

11.11 Foucault — power/knowledge and the norm as instrument (Family E)

CORE CLAIM. FACT that this is Foucault's line of analysis, associated with Discipline and Punish, The History of Sexuality vol. 1, and the lectures on governmentality/biopolitics. The norm is not a neutral technical object. To normalize is to exercise power: norms are produced within, and reproduce, relations of power/knowledge (pouvoir/savoir) — the apparatuses that define what counts as normal/deviant, that classify, measure, and discipline populations. Categories that appear neutral and merely descriptive (the delinquent, the patient, the risk-profile) are effects of power that then constitute their objects. Governance ("governmentality") operates less by sovereign prohibition than by the productive shaping of conduct through norms, statistics, and self-regulation. On this view, the demand for a "content-neutral" governance mechanism is itself ideological: neutrality is the characteristic disguise of a particular configuration of power.

SHARPEST FORM (falsifiable challenge to CG). INFERENCE CG's founding distinction — a content-NEUTRAL kernel plus a value-laden content layer — is, on this analysis, not available: the machinery of representation (which categories are first-class, which distinctions the kernel can express, what counts as a "norm" at all) is ITSELF a value-laden, power-laden choice. The very act of formalizing a domain into computable categories exercises the power to define, to include and exclude, to render some things visible and legible and others invisible. There is no neutral encoding; the "kernel" is a knowledge-apparatus that will normalize whatever it measures. Falsifiable-ish form (this attack is the hardest to state as a clean test, and we flag that): exhibit a purportedly neutral encoding and show that its choice of primitives systematically advantages some interests/populations and disadvantages others — i.e., the neutrality is not achievable, only concealed. OPEN Whether "no encoding is neutral" is falsifiable at all is itself contested; stated honestly per M1, the strongest Foucauldian claim risks being unfalsifiable, which paradoxically weakens it as an M1-admissible attack even as it deepens it as a warning. We tag the strong version HYPOTHESIS bordering OPEN.

SIX-PART READING — implicated source: disciplinary/administrative norms & classifications (cf. Memo 06, Memo 07).

ARCHITECT MUST concede or answer. INFERENCE (1) That the choice of primitives (what the kernel can and cannot express — Hohfeldian relations, deontic operators, the category "norm" itself) is a substantive commitment with distributive consequences, not a neutral given. The architect must own the non-neutrality of the representation rather than claim to have escaped it. (2) That "legibility" has costs (rendering a domain machine-readable erases what does not fit the schema — echoing the anti-representational core of Hayek and Ostrom from the critical side). (3) A reflexivity requirement: the system should expose what it cannot see (its blind spots / non-representable remainder) as a first-class output, because the danger is precisely the invisibility of the excluded. INFERENCE This is the one attack that most directly threatens Q8's premise (that value can be cleanly quarantined "above" a neutral kernel); it says the quarantine boundary is itself value-laden.

REACHES. CG-STRONG and, uniquely, the coherence of the STRONG/WEAK split itself. But its M1 standing is weak (near-unfalsifiable in its strongest form). Net: a profound design warning about the non-neutrality of encoding, of limited use as a refutation but of high use as a permanent risk to audit.

11.12 Arendt — action, plurality, natality: the irreducible to process (Family F)

CORE CLAIM. FACT that these are Arendt's concepts, from The Human Condition and related work. Arendt distinguishes labor, work, and action. Action — the specifically political activity of appearing and speaking among others — has three features hostile to processing. (a) Plurality: action happens among distinct, unique persons ("men, not Man, live on the earth"); it is irreducibly a matter of the many and their differing perspectives, not a single viewpoint. (b) Natality: each person is a new beginning capable of initiating something genuinely unprecedented — action can start a causal chain not deducible from what came before. (c) Unpredictability and irreversibility: the outcomes of action cannot be foreknown, and deeds cannot be undone; Arendt offers forgiving (to release from irreversibility) and promising (to bind an unpredictable future) as the specifically human, non- mechanical responses. Politics, on this view, lives in the space of appearance between plural actors and is destroyed when replaced by making (fabrication toward a blueprint) or by behavior (statistical conformity).

SHARPEST FORM (falsifiable challenge to CG). INFERENCE To govern by computation is to treat politics as either FABRICATION (execute a blueprint) or BEHAVIOR (predict and optimize statistical regularity). Both, on this analysis, eliminate action: fabrication substitutes a maker's end for plural initiative; behavioral governance substitutes predictable conformity for the capacity to begin. Natality means the governed can always do the genuinely unprecedented, which no model trained on the past can anticipate; plurality means there is no single objective to optimize because there is no single perspective. So the political phenomenon CG proposes to execute is, on this view, precisely the thing that computation cannot contain — it can only suppress it into behavior. Falsifiable-ish: any governance model predicts/prescribes on the basis of the past; natality guarantees a nonzero rate of genuinely novel initiatives the model neither predicted nor can price. Show that the model must either ignore these or coerce them into its categories.

SIX-PART READING — N/A as a discrete "source." INFERENCE Arendt is not describing a norm-source in the JD Q1 sense (like legislation or contract); she is describing the human capacity (action) that norm-sources presuppose and that governance addresses. We therefore mark the six-part grid not applicable and instead record the structural point:

  Arendt targets not a SOURCE of norms but the SUBJECT of governance:
  the plural, natal, initiating human. The claim is ONTOLOGICAL — about the
  kind of being governed — not institutional. It bears on CG as a claim that
  the OBJECT of the computation (political action) is mis-typed if modeled as
  process, blueprint-execution, or predictable behavior.

ARCHITECT MUST concede or answer. INFERENCE (1) That models built on the past cannot anticipate genuinely novel action (natality); the system must treat novelty as expected and leave room for the unprecedented rather than coercing all conduct into pre-existing categories (this connects to Hart's open texture and Schmitt's exception from the ontological side). (2) That plurality undermines the single-objective assumption: there may be no one utility function because there are irreducibly many perspectives — echoing, from a phenomenological angle, the social-choice impossibility results (§11.16.3) that say no aggregation of plural preferences is fully well-behaved. (3) That "promising" and "forgiving" — mechanisms for binding an unpredictable future and releasing irreversibility — might be modellable as commitment and revision primitives, which is a rare constructive reading of Arendt for the architect (contracts as promises; amnesty/pardon/revision as forgiving).

REACHES. CG-STRONG (the object is mis-typed). Its bite on CG-WEAK depends on the rate of genuinely novel action: if novelty is rare and localizable, CG-WEAK survives by referring novelty to a human boundary; if action is pervasively novel, the "referral" swamps the system. OPEN The empirical rate of governance-relevant novelty is unknown.

11.13 Luhmann — autopoiesis: law cannot be steered from outside its own code (Family F)

CORE CLAIM. FACT that this is Luhmann's systems-theoretic position, associated with Law as a Social System and Social Systems. Modern society consists of functionally differentiated autopoietic (self-producing) subsystems — law, politics, economy, science — each operationally closed: each reproduces its own elements (communications) using its own binary code (law: legal/illegal [Recht/Unrecht]; economy: payment/non-payment; science: true/false). A system is cognitively open (it perceives its environment) but operationally closed (only it can produce its own operations by its own code). Crucially, one system cannot directly steer another: the economy cannot issue legal validity, politics cannot issue legal validity — politics can only produce irritations/perturbations that law then processes in law's own terms. Law is normatively closed: only law determines what is legal. External interventions are re-entered through the system's own code or they are not law at all.

SHARPEST FORM (falsifiable challenge to CG). INFERENCE CG proposes an external computational apparatus (the kernel) that produces or determines normative outcomes for social systems. But if law (and each normative system) is operationally closed, then an external kernel is, by definition, in the ENVIRONMENT of the system: it can perturb but cannot produce the system's own operations. A kernel's output is not "law" unless the legal system, in its own recursive operation, makes it so; the kernel can at most be an irritation the system may or may not internalize. So CG mis-locates itself: it imagines steering from outside a system that can only steer itself. Attempts to substitute external computation for the system's self-reproduction either fail (the system ignores or re-codes the input) or, if imposed, DE-DIFFERENTIATE the system (collapse law into politics/administration), which the theory predicts is dysfunctional. Falsifiable-ish: predict that computed normative outputs imposed from outside are either re-processed on the system's own terms (and so the kernel is not the real author) or degrade the system's function by collapsing its code.

SIX-PART READING — implicated source: law as an autopoietic system (cf. Memo 06).

ARCHITECT MUST concede or answer. INFERENCE (1) That the kernel is, at best, structurally coupled to the normative systems it serves — an irritation source and record-keeper, not the author of validity; the architect must locate the kernel as environment to existing systems and design for re-entry (how the target system adopts the kernel's output as its own operation). (2) That imposing external computation risks de-differentiation (collapsing distinct codes), a systemic harm the architect must guard against. (3) A reframing that is constructive: model the kernel as a structural coupling mechanism (like a constitution couples law and politics) rather than as a sovereign author — this is a coherent, humbler target that Luhmann's theory does not forbid. INFERENCE This converges with Weber (§11.9) and Rawls (§11.6): the kernel carries validity/procedure but does not author legitimacy, which is generated by the system's own operation.

REACHES. CG-STRONG (external authorship is mis-conceived). CG-WEAK survives only by relocating the kernel from "author" to "coupled instrument/record," a significant reconception. Logged as one of the two or three most structurally serious attacks in the memo.

11.14 Gödel, Turing, Rice — the formal-limits attacks, stated without overclaiming (Family G)

This family is different in kind from A–F: it is mathematics, not a claim about politics. Because overstated versions circulate widely, the discipline here is to state exactly what the results prove and exactly what they do not.

11.14.1 What the results actually say

• FACT Gödel's first incompleteness theorem. Any consistent, effectively axiomatized formal system strong enough to express elementary arithmetic contains true sentences it cannot prove. The second theorem: such a system cannot prove its own consistency.
• FACT Turing / the halting problem. There is no algorithm that decides, for every program-input pair, whether the program halts.
• FACT Rice's theorem. Every non-trivial semantic property of the function computed by a program is undecidable in general.

11.14.2 SHARPEST FORM as a challenge to CG

HYPOTHESIS Restated as a falsifiable challenge: if the governance kernel must decide non-trivial semantic properties of arbitrary norm-programs — e.g., "does this body of rules ever produce a contradictory obligation?", "will this regulation's procedure always terminate?", "is norm-set A equivalent to norm-set B?" — then by Rice/Turing there is no general algorithm that does so for arbitrary inputs, and by Gödel a sufficiently expressive norm-logic has true but unprovable normative consequences. The program fails iff it needs those decisions over the unrestricted input class.

11.14.3 Why the bite is narrow (what the results do NOT prove)

INFERENCE The formal limits are results about worst-case, general-input decision procedures over Turing-complete systems. They do not prove:

1. that governance requires Turing-complete norm-representations — most operative rule-sets (SOPs, tax schedules, eligibility rules; Memo 07) are expressible in decidable fragments (finite-state, primitive-recursive, decidable logics);
2. that the specific norm-sets a polity uses fall in the undecidable region — undecidability is about all programs, not the ones actually written;
3. that humans escape these limits — a human legal system is not a hyper- computer; if the limits bound machines they bound human formalization too. INFERENCE So Family G, used honestly, is symmetric: it constrains any formalization, human or machine, and therefore cannot by itself argue "humans can, machines cannot."

FACT Standard mitigation exists in every formal-methods practice: restrict the language to a decidable fragment, accept incompleteness by adding a decidable "I-don't-know / escalate" verdict, or bound resources. These convert undecidability into a localized boundary — exactly the CG-WEAK move.

11.14.4 ARCHITECT MUST

INFERENCE (a) Declare the expressiveness class of D1's norm-language and keep the core checks (validity, conflict, termination of procedures) inside a decidable fragment; (b) make undecidable/incomplete cases return an explicit UNDECIDED → escalate verdict rather than a silent wrong answer (three-valued, not two-valued, output); (c) never claim completeness for the norm-logic.

11.14.5 The blending error (a warning, per M1 honesty)

INFERENCE It is a fallacy — and a credibility risk for the program — to argue "Gödel/Turing therefore governance is not computable." That inference is invalid: the formal results bound a specific worst case, while the serious threats to CG are the empirical/philosophical ones in families A–F and H. Overclaiming Family G is itself an attack surface: opponents can refute the overclaim and appear to have refuted the whole critique. REACHES. Bounded. Family G dents CG-STRONG only on the unrestricted-input reading and is fully consistent with CG-WEAK.

11.15 The AI/cognitive family — frame, symbol-grounding, specification-gaming (Family H)

These are engineering-grade objections about any system that must apply norms to the open world. They are, in the assessment of this memo, jointly the most practically dangerous to CG-STRONG.

11.15.1 The frame problem — relevance is unbounded

INFERENCE Core claim (the problem associated with McCarthy–Hayes and its philosophical generalization by Dennett/Fodor): a formal agent cannot, in general, determine which facts are relevant to applying a rule without, in the limit, checking indefinitely many facts about the world. Norm application is saturated with implicit ceteris-paribus conditions ("drive safely" presupposes an open-ended background). SHARPEST FORM: CG fails if determining the relevant fact-set for applying a norm is not itself boundedly computable. ARCHITECT MUST: supply a closed-world/relevance boundary per rule domain and accept that norms silently assume a background frame that must be made explicit. REACHES CG-STRONG; dents CG-WEAK (the frame must be localizable per domain).

11.15.2 Symbol grounding — the predicates must touch the world

INFERENCE Core claim (the problem associated with Harnad; kin to Searle's Chinese Room): the terms in a norm ("weapon," "injury," "consent," "dwelling") must be connected to real-world referents for the norm to be applied; a purely syntactic engine manipulates ungrounded tokens. SHARPEST FORM: CG fails if the grounding of thick predicates (Memo 03 §3.5) cannot be delegated to a reliable classifier/oracle and instead requires human judgment case-by-case. ARCHITECT MUST: treat predicate-grounding as an explicit, fallible oracle at the kernel boundary, with provenance and an error model — not as an internal kernel capability. REACHES CG-STRONG; consistent with CG-WEAK if grounding is a labeled boundary.

11.15.3 Specification-gaming / value mis-specification

INFERENCE Core claim (Goodhart's law; reward-hacking in RL; the principal-agent problem, Memo 07 §07.8): agents optimize the written proxy, not the intended goal, and the two diverge under pressure. SHARPEST FORM: an executable governance layer that penalizes/rewards against a metric will be gamed, and the gap between metric and intent is generally not closable by adding more rules (each patch adds new gameable surface). ARCHITECT MUST: assume adversarial optimization against every executable metric; keep a human appeal/override channel; monitor for surrogation. REACHES CG-STRONG and genuinely dents CG-WEAK, because the residue here is generative (new gaps appear as you patch) rather than a fixed nameable remainder. OPEN Whether specification-gaming residue is localizable is unresolved and is one of the two or three most serious open threats in the series.

11.16 Why the attacks do not simply add up

INFERENCE A naïve reading sums the attacks into "governance is uncomputable." That is unsound, for three reasons, and honesty (M1, M8) requires stating them even though this is the adversarial memo.

1. Mutual inconsistency. Several attacks presuppose incompatible pictures. Schmitt's decisionism (norms bottom out in an un-normed sovereign decision) and Luhmann's autopoiesis (law is a closed self-producing code with no sovereign author) cannot both be the reason CG fails; they are rival diagnoses. Foucault (norms are diffuse power with no center) sits uneasily with Rawls (norms should issue from a well-defined justificatory procedure). INFERENCE One cannot bank all of them; using M8, they are evidence, and the evidence is internally divided.
2. Different targets. Families A/B/H attack application and input; C/D/E/F attack legitimacy, genesis, and ontology; G attacks formal completeness. A defender could lose on one axis and win on others. The attacks do not share a single load-bearing premise whose fall takes CG with it.
3. STRONG vs WEAK exposure. As tabulated per section, nearly every attack reaches CG-STRONG but only dents CG-WEAK. INFERENCE Their union therefore strongly refutes the ambitious thesis while leaving the modest thesis standing but heavily taxed with mandatory boundaries.

11.17 Open questions (the unresolved battle line)

• OPEN The localizability question (the deepest one). Is the un-representable residue localizable (nameable, boundable, escalatable — CG-WEAK survives) or pervasive/constitutive (entangled through all cases — CG-WEAK also fails)? Hart's penumbra and Family G look localizable; Foucault, Schmitt, and specification-gaming look pervasive. Unresolved.
• OPEN Is legitimacy externalizable? Can legitimacy be treated as an above-kernel input (a provenance tag) or is it produced only in the act of human judgment/reflection (Rawls) or communication (Arendt/Luhmann)? If the latter, the kernel can execute but never legitimate.
• OPEN The exception. Can the Schmittian genesis/emergency moment be given a decidable trigger, or is "when the rules are suspended" necessarily outside the rules? Recurs from Memo 01/09.
• OPEN Grounding reliability. What error model for predicate-grounding oracles is acceptable, and who bears the cost of misclassification?
• OPEN Generative vs fixed residue. Is specification-gaming residue a fixed remainder or does it regenerate under patching (§11.15.3)?

11.18 Residue after the attacks: what survives the strongest fire

Per M6, "no kernel exists" remains an acceptable outcome. Stated honestly, the attacks establish the following, and no more:

• INFERENCE CG-STRONG does not survive. At least four independent attacks (Hayek's dispersed tacit knowledge, Foucault's constitutive power, Schmitt's exception, specification-gaming's generative residue) each exhibit governance phenomena that resist faithful reduction to (K, content) without residue. The ambitious "without residue" claim should be abandoned.
• INFERENCE CG-WEAK survives, but heavily taxed. Every attack that reaches only CG-WEAK converts into a mandatory design obligation: a decidable core fragment (G), an explicit UNDECIDED/escalate verdict (B, G), predicate- grounding oracles with error models (H), localized value/legitimacy boundaries (C, E), an exception/emergency boundary the kernel does not itself decide (D), and monitoring for gaming (H). CG-WEAK is thus not refuted but is only viable as "kernel + explicitly localized human/oracle boundaries."
• OPEN Whether even CG-WEAK is honest depends entirely on §11.17's localizability question. If the residue is pervasive rather than localizable, CG-WEAK collapses into "a formalism that mishandles the hard cases silently" — the worst outcome, worse than no kernel. This memo cannot close that question; it hands it to Memo 12 and flags it as the pivot on which the entire program turns.

INFERENCE The invariant that survives all attacks is modest and negative: the kernel may represent, apply within a declared fragment, and record norms and their provenance; it may not author content, ground thick predicates unaided, legitimate, or decide the exception. That negative residue — a kernel of strictly bounded competence surrounded by named boundaries — is the most that survives, and it is exactly the candidate interface Memo 12 must try to specify.

11.19 Handoff to D1/D2

Concrete obligations the attacks impose on the compiler architects:

1. INFERENCE Abandon "without residue." Design for CG-WEAK: kernel plus explicitly localized boundaries. Any claim of faithful total reduction is refuted (§11.18).
2. INFERENCE Three-valued verdicts. The kernel must be able to output UNDECIDED/escalate, never forced to guess on penumbra (Hart), undecidable (Rice/Turing), or ungrounded (Harnad) inputs.
3. INFERENCE Declared expressiveness fragment. Fix D1's norm-language to a decidable class for core checks; document what is given up (§11.14.4).
4. INFERENCE Predicate-grounding as a first-class fallible oracle at the boundary, with provenance and an error model (§11.15.2).
5. INFERENCE Value and legitimacy stay above the kernel as tagged inputs (Rawls, Nozick, Foucault); the kernel executes, it does not justify.
6. INFERENCE An explicit exception/emergency boundary the kernel does not itself trigger (Schmitt); log who does.
7. INFERENCE Adversarial-by-default metrics. Assume gaming of every executable KPI; keep human appeal/override and surrogation monitoring (§11.15.3, Memo 07).
8. INFERENCE Do not overclaim Gödel/Turing. Prohibit marketing that treats formal limits as either a proof of impossibility or a badge of rigor (§11.14.5).
9. OPEN → pivotal flag Instrument the localizability question. Build the experiment that measures whether residue localizes or spreads; the go/no-go of the whole program depends on it (§11.17, §11.18).

One-line summary: the attacks kill the ambitious thesis and leave a modest, strictly-bounded kernel-plus-named-boundaries that is viable only if the un-representable residue turns out to be localizable — an open question Memo 12 must confront rather than assume away.