A Constitution for the Context Window

Large language models do not encounter users, projects, institutions, or environments directly. Each provider invocation presents the model with a bounded construction assembled from governing policy, retrieved evidence, runtime observations, persistent memory, tools and capability descriptions, selected history, the current user request, and vendor or base-system instruction. These source classes form PROMTHUS: Policy, Retrieval, Observations, Memory, Tools, History, User request, System instruction. PROMTHUS is an inventory, not an authority hierarchy or serialization order. Its sequence serves a mnemonic and Promethean construction; it does not imply that Policy is universally supreme or that later classes possess lower standing. PROMTHUS names the principal materials from which an invocation context is compiled. To promthe an invocation is to subject those materials to constitutional context compilation: governed discovery, authentication, semantic finding, independently derived consequence analysis, contest, adjudication, authorized transformation, budget closure, provider realization, effect authorization, and evidentiary sealing. This paper proposes constitutional context compilation as a typed authorization, adjudication, and transformation layer between context sources and inference. The compiler governs what may count as context, what may govern inference, what may be disclosed, how admitted context may be transformed, what may cross a provider boundary, which consequences may follow, and what must remain auditable afterward. The architecture defines the minimum contract required of a Constitutional Kernel capable of governing context compilation. Such a kernel must provide an explicit and non-extensible bootstrap basis, identity-bearing root law, governed discovery obligations, authorized amendment, court jurisdiction, non-self-ratification, typed judgment, structurally triggered contest, review and appeal, refusal semantics, independent effect authorization, and witness. The paper distinguishes a presupposed Grundnorm Basis from a positive, executable Root Constitution written in Deterministic Semantic Compiler (DSC) form. DSC coding means that law is expressed as explicit, canonically serializable, content-identified rules that execute deterministically over admitted structured predicates, rather than as prose whose application depends on interpretation at the moment of use. The Grundnorm Basis identifies the genesis custody and initial standing that any finite constitutional order must presuppose. It is fixed at genesis and cannot later be enlarged by declaration. The Root Constitution makes operational successor law deterministic over admitted structured predicates, inspectable, content-identified, testable, auditable, and mutable only through authorized amendment. Context-specific law derives from that root. Discovery profiles, calibration profiles, conflict-detection profiles, independence profiles, domain canonicalization profiles, consequence tiers, court procedures, provider-adapter contracts, and effect boundaries all receive standing from law outside the discretion they constrain. Any profile whose declarations bound the scope of scrutiny must be identity-bearing, evidence-class-typed, and adjudicated before use. Identity itself is layered to avoid a bootstrap circularity. A primitive, exclusion-free Constitutional Canonical Form (CCaF), bound into the genesis recognition machinery, identifies constitutional objects; adjudicated domain profiles govern application material. Every identity is permanently typed by the layer under which it was computed, and relations across layers are typed correspondences rather than collapsible equivalences. Completeness of governance is made checkable rather than asserted. A discretion census over the declared schema and transition registry enumerates every mechanism able to affect identity, authority, admission, scope, scrutiny, conflict detection, transformation, realization, effect permission, or evidentiary claims, and each censused locus must bind to an audit row or a proven reduction to one. Classifiers do not become lawmakers; their outputs become receipted findings subject to calibration, conservative consequence computation, independent contest, judicial review, appeal, and refusal. The context compiler executes admitted judgments. The Context Intermediate Representation is the identity-bearing product of that governed process, not its source of authority. The architecture distinguishes six principal failures: context loss, containment loss, authority loss, representation loss, realization loss, and receipt loss. It defines three layers of context identity over two canonicalization layers, strict partial-order precedence, inter-court conflict law, scrutiny-profile adjudication, cross-layer correspondence, governing projections, delegation contracts, autonomous-loop review, availability protections against stay flooding, privacy-preserving receipts, two-phase budget closure, staged conformance, and measurable governance cost. Its general constitutional proposition is: Every mechanism that determines whether another mechanism is scrutinized must itself be governed by authority external to the discretion it constrains. The paper audits the architecture against that proposition by enumerating its discretionary loci, the power held at each locus, the external governor, the evidentiary record, and the failure emitted when governance is absent. The discretion census supplies the enumeration procedure, so the audit's completeness claim rests on a declared schema rather than on architectural omniscience. The architecture does not prove truth, model understanding, obedience, classifier completeness, contest independence beyond available evidence, profile completeness beyond available evidence, or remote-provider honesty. This paper specifies a narrower conditional assurance claim: a conforming realization may establish, relative to its declared capability surface and evidence, that context was assembled under explicit machine-governance rules, adjudicated through authorized procedures, realized according to an admitted adapter contract, and bound to evidence sufficient for later inspection. The paper itself is a proposal and proof-obligation map; it supplies no implementation-conformance verdict. A constitutional context layer could unify memory, retrieval, personalization, policy, capability exposure, delegation, and long-running automation within one systems discipline. It would also make durable computational agency portable across replaceable inference providers---where the user or another legitimate custodian actually holds the registry, amendment authority, and receipts on which such portability depends.