← Back to Home

1. Introduction: The Problem of Ungoverned Output

The central problem in deployed AI systems is not accuracy but accountability. A language model that produces a correct answer without structural justification for its correctness is, from an engineering standpoint, indistinguishable from one that produces an incorrect answer with equal confidence. Both lack what we term semantic validity — the structural completeness that distinguishes genuine knowledge from sophisticated pattern matching.

Current mitigation strategies — retrieval-augmented generation (RAG), chain-of-thought prompting, reinforcement learning from human feedback (RLHF) — address symptoms rather than architecture. They improve output quality on average while leaving the fundamental structural deficit intact. A model with RAG still cannot determine whether its grounded retrieval is relevant to the claim being constructed, because it lacks a structural definition of relevance.

This paper presents a constraint-based architecture that addresses the structural deficit directly. The architecture identifies six necessary relational constraints that any semantically valid claim must satisfy. These constraints are not design choices but geometric inevitabilities arising from the topology of knowledge claims. We demonstrate that enforcing completeness across all six constraints prior to output emission eliminates structurally deficient outputs by construction rather than by statistical correction.

2. Background and Prior Art

Existing approaches to AI output governance fall into three categories: post-hoc filtering, in-context steering, and retrieval grounding. Post-hoc filtering systems such as content classifiers and fact-checking modules evaluate outputs after generation, creating a reactive rather than preventive architecture. In-context steering methods such as chain-of-thought and system prompting modify the generation process but do not constrain it — the model remains free to produce structurally incomplete outputs that happen to follow the prompted format. Retrieval-augmented generation anchors outputs to external documents but provides no mechanism for evaluating whether the retrieved content is structurally sufficient for the claim being constructed.

None of these approaches answer the foundational question: what constitutes a complete claim? Without a formal definition of completeness, all governance is partial. Our contribution is precisely this formal definition, derived not from statistical analysis of language corpora but from ontological analysis of the structure of knowledge itself.

3. The Six-Constraint Framework

The framework rests on a single structural observation: any knowledge claim necessarily involves exactly four components and six binary relations between them. The four components are the Claimant (who makes the claim), the Subject (what the claim is about), the Grounds (the evidence or basis), and the Purpose (the reason for making the claim). These four components are connected by six pairwise relations, each of which imposes a constraint on semantic validity.

Constraint	Relation	Question	Failure Mode
Referential	Claimant — Subject	What is being referred to?	Ambiguous reference
Contextual	Claimant — Grounds	Where does this stand?	Decontextualized assertion
Premissive	Subject — Grounds	Why does this hold?	Unsupported premise
Inferential	Grounds — Purpose	How does this follow?	Invalid inference
Constraining	Subject — Purpose	What does this exclude?	Unbounded claim
Teleological	Claimant — Purpose	For what end?	Purposeless assertion

Table 1. The six constraints, their relational basis, diagnostic questions, and characteristic failure modes when absent.

4. System Architecture

The system consists of four components operating in sequence. The Hexis Extraction Engine decomposes raw input into six constraint channels, assigning each element of the input to one or more of the six relational constraints. The Metadata Tagging Module encodes each extracted element with constraint identifiers, preserving provenance and relational context. The Validation Gate evaluates completeness by confirming that all six constraint channels contain structurally adequate content. The Governed Reconstitution Engine reassembles validated elements into output, ensuring that the emitted claim preserves the structural relationships established during extraction.

The critical architectural decision is the placement of the validation gate before output emission rather than after. This is not a filter — it is a gate. The distinction is meaningful: a filter removes bad outputs from a stream; a gate prevents incomplete outputs from being constructed in the first place. The system cannot hallucinate because a hallucination, by definition, is a claim that lacks structural grounding in one or more constraint channels, and such claims cannot pass the gate.

5. The Tetrahedral Structure

The six constraints correspond to the six edges of a tetrahedron. The four vertices represent the four necessary components of a knowledge claim: Claimant, Subject, Grounds, and Purpose. Each edge connecting two vertices represents one binary relation — one constraint — that must be satisfied.

The constraints organize into three dyads — pairs of complementary constraints that form a complete relational axis. The RC dyad (Referential + Contextual) governs grounding: establishing what is being discussed and where it stands. The PI dyad (Premissive + Inferential) governs reasoning: establishing why a claim holds and how it follows. The KT dyad (Constraining + Teleological) governs governance: establishing what the claim excludes and for what end it is made.

7. Empirical Results

We evaluated the architecture across four content domains: geopolitical analysis, medical case summaries, legal contract review, and technical documentation. For each domain, a corpus of 200 claims was extracted from representative documents and processed through the six-constraint validation pipeline. Each claim received a binary classification — valid (all six constraints satisfied) or invalid (one or more constraints unsatisfied) — along with a diagnostic report identifying which constraints were satisfied and which were not.

Domain	Claims	System Valid	System Invalid	Expert Agreement
Geopolitical Analysis	200	127	73	>95%
Medical Case Summaries	200	142	58	>95%
Legal Contract Review	200	118	82	>95%
Technical Documentation	200	156	44	>95%

Table 2. Validation results across four domains. “Expert Agreement” denotes observed agreement rate between the system’s categorical (valid/invalid) determination and independent human expert assessment of the same claims.

The observed agreement rate exceeding 95% across all domains indicates that the six-constraint framework captures the structural features that human experts use — whether explicitly or implicitly — when evaluating claim completeness. Notably, the failure mode distribution was non-uniform: the Constraining constraint (what does the claim exclude?) was the most frequently absent, accounting for 34% of all constraint failures. This aligns with the well-documented tendency of language models to produce maximally inclusive — and therefore semantically unbounded — assertions.

10. Conclusion

We have presented a six-constraint semantic validity architecture that addresses the structural deficit in current AI output governance. The architecture is derived from ontological analysis of the necessary components of knowledge claims rather than from statistical analysis of language patterns. It provides a deterministic validation gate that prevents structurally incomplete claims from reaching the output stage. Empirical evaluation across four domains demonstrates observed human-expert agreement exceeding 95%, confirming that the six constraints capture the structural features operative in human epistemic judgment.

The architecture is compatible with existing transformer-based language models and does not require retraining. It operates as a governance layer that can be applied to any system that generates natural language claims. Future work will extend the framework to multi-claim discourse structures, real-time processing at production scale, and domain-specific constraint weighting.