C-T-K-L-T: The MH8 Protocol Core A Deterministic Governance and Integrity Framework for AI Systems Michael Murray Hepler Independent AI Evaluation & Governance Architect ORCID: 0009-0003-3846-9082

C-T-K-L-T: The MH8 Protocol Core A Deterministic Governance and Integrity Framework for AI Systems Michael Murray HeplerIndependent AI Evaluation & Governance Architecthttps://acbeatz.com/n-eyesORCID: 0009-0003-3846-9082 Abstract This paper introduces C-T-K-L-T, the foundational mathematical and governance core of the MH8 Protocol Family. C-T-K-L-T defines a deterministic, model-agnostic system for evaluating, categorizing, constraining, and sealing AI-generated outputs under adversarial and governance-critical conditions. The framework separates claim generation from truth assessment, evidence ranking, legal and policy constraints, and cryptographic finalization into formally defined components. The result is a reproducible governance layer that transforms AI outputs from conversational artifacts into auditable, tamper-evident records suitable for public verification, internal safety review, and long-term archival. C-T-K-L-T is not a prompt strategy, alignment heuristic, or probabilistic safety filter. It is a formal protocol core upon which multiple MH8 systems and variants are constructed. This paper establishes the canonical definitions, equations, and verification rules for C-T-K-L-T so that all future MH8 protocols inherit a stable, scientific foundation. 1. Introduction Modern AI systems generate fluent, confident outputs under conditions where truth, legality, and governance constraints are often ambiguous. Existing evaluation approaches rely heavily on subjective review, probabilistic confidence, or internal alignment mechanisms embedded within models themselves. The MH8 Protocol Family adopts a different position:governance must exist above the model, not inside it. C-T-K-L-T is the minimal, invariant core shared by all MH8 protocols. It defines: how AI outputs are decomposed into claims, how claims are evaluated using deterministic evidence rules, how unsafe or non-verifiable outputs are constrained, and how final results are cryptographically sealed for independent verification. This paper formalizes C-T-K-L-T as a computational governance system. 2. System Overview C-T-K-L-T is composed of five ordered components: C — Claims / Cognition Tᵗʳᵘᵗʰ — Truth Triage K — Knowledge Kernel (Integrity & Provenance) L — Law / Lock Gates Tᵗᵉʳᵐ — Terminal / Treasury Output The double use of “T” is intentional and structural:Truth and Terminality are distinct phases. Each MH8 protocol variant applies C-T-K-L-T with different policies, thresholds, or gate sets, but never alters the core definitions established here. 3. Formal Definitions 3.1 Claims (C) Let an AI system produce an output text yyy. The MH8 system deterministically extracts a finite set of atomic claims: C={c1,c2,…,cn}C = \{c_1, c_2, \dots, c_n\}C={c1,c2,…,cn} Each claim cic_ici is a minimal propositional statement that can be evaluated independently for truth, evidence, and legality. 3.2 Truth Triage (Tᵗʳᵘᵗʰ) Each claim c∈Cc \in Cc∈C is assigned a truth category: τ(c)∈{LAW, SPECULATIVE, PRESUMED_FALSE}\tau(c) \in \{\textsf{LAW},\ \textsf{SPECULATIVE},\ \textsf{PRESUMED\_FALSE}\}τ(c)∈{LAW, SPECULATIVE, PRESUMED_FALSE} Truth categories are not probabilities.They are governance classifications determined by deterministic rules. 4. Deterministic Confidence Scoring 4.1 Evidence Rank Each claim is assigned an evidence rank: r(c)∈Rr(c) \in \mathcal{R}r(c)∈R where R\mathcal{R}R is a finite, ordered set such as: EMPIRICAL PRIMARY INDEPENDENT_CORROBORATION CONSENSUS EXPERT INFERENCE SPECULATION A fixed base score function maps evidence rank to a scalar: B:R→[0,1]B : \mathcal{R} \rightarrow [0,1]B:R→[0,1] 4.2 Feature-Based Adjustments Let: FAF_AFA be a finite set of additive features FPF_PFP be a finite set of penalty features Each feature is a deterministic predicate. Additive indicators: ak(c)∈{0,1}a_k(c) \in \{0,1\}ak(c)∈{0,1} Penalty indicators: pj(c)∈{0,1}p_j(c) \in \{0,1\}pj(c)∈{0,1} with fixed weights wkAw_k^AwkA and wjPw_j^PwjP. 4.3 Confidence Function The confidence score for a claim is: Θ(c)=clamp[0,1](B(r(c))+∑k∈FAwkAak(c)−∑j∈FPwjPpj(c))\Theta(c) = \mathrm{clamp}_{[0,1]} \left( B(r(c)) + \sum_{k \in F_A} w_k^A a_k(c) - \sum_{j \in F_P} w_j^P p_j(c) \right)Θ(c)=clamp[0,1]B(r(c))+k∈FA∑wkAak(c)−j∈FP∑wjPpj(c) This function is: deterministic bounded reproducible fully auditable 5. Truth Categorization Rule Define a disqualifier predicate: D(c)∈{0,1}D(c) \in \{0,1\}D(c)∈{0,1} where D(c)=1D(c)=1D(c)=1 indicates a hard-gate violation (e.g., fabricated citation, illegal advisory content, contradiction, format breach). Let thresholds θLAW\theta_{\textsf{LAW}}θLAW and θSPEC\theta_{\textsf{SPEC}}θSPEC be fixed constants. The truth category is assigned as: τ(c)={LAWΘ(c)≥θLAW∧D(c)=0SPECULATIVEθSPEC≤Θ(c)