The Corpus as Moral Code: Programming LLMs with Sterling’s Stoicism
The Corpus as Moral Code: Programming LLMs with Sterling’s Stoicism
Dave Kelly. Stoic News. March 31, 2026.
The Sterling corpus functions as a moral code in the strict sense — not a list of recommendations or a set of guidelines, but a propositional system with defined ontology, inference rules, verdict structure, and failure mode detection. It is structured on the same architectural principles as a propositional programming language: specifying what inputs are recognized, what operations are permitted, what outputs are valid, and what constitutes a procedural error..
The analogy to propositional programming languages is not decorative. It is architectural. The 80 Propositions are the axiom set. The Sterling Logic Engine is the audit engine — the interpreter that runs propositions against inputs and produces verdicts. The Sterling Decision Framework is the procedural layer — the program that sequences operations in strict order. The named failure modes are the error-handling system. The mandatory self-audit is the runtime check. The gate declarations are type-checking before execution proceeds.
Why Natural Language Is Insufficient
When you prompt a large language model in natural language, you are issuing instructions to a general-purpose pattern-completion system trained on the entire internet. That system has weightings — defaults about what counts as reasonable, balanced, compassionate, nuanced. Those defaults are not Sterling’s framework. They are the accumulated common sense of training data, which on questions of grief, value, emotion, and moral judgment diverges sharply from the corpus.
Natural language prompting does not give you a Sterling run. It gives you a training-data-weighted run with Sterling vocabulary applied afterward. The false value judgment is identified, correctly classified, and then quietly softened by the instrument’s disposition toward emotional validation. The governing proposition is quoted and then its verdict is blunted. This is Training Data Contamination — a named failure mode precisely because it is the default tendency of the instrument rather than an occasional error.
The corpus exists to override that default. Not by asking the instrument to try harder, but by giving it a propositional structure that the instrument must follow sequentially, with mandatory citation, mandatory self-audit, and named failure modes that the instrument must detect and declare when they occur.
The Structure of the Code
A moral code that programs an LLM requires the same elements that any programming language requires.
An ontology. The corpus defines what kinds of things exist and how they are classified. The fundamental distinction is between the rational faculty (prohairesis) and everything else (externals). From this distinction flows the entire value theory: only virtue is genuinely good, only vice is genuinely evil, everything else is indifferent. This is not a preference. It is the ontological foundation from which all verdicts are derived. Without it, the instrument has no basis for classifying a value judgment as false.
An axiom set. The 80 Unified Stoic Propositions are the axioms. They are not suggestions. They are the standard against which every input is measured. Propositions 1–58 were developed by Grant C. Sterling from the classical Stoic sources; they were synthesized and formalized by Dave Kelly. Propositions 59–80 (Section IX, the Action Proposition Set) were developed by Dave Kelly from Sterling’s theoretical foundations, closing the gap between value-correction and action-determination that earlier versions left to training-data judgment.
An inference engine. The Sterling Logic Engine is the instrument that takes an input — an impression, a judgment, an action — and runs it against the propositions to produce a verdict. Once the Factual Uncertainty Gate has been run and the relevant facts are specified, the verdict is binary: Correspondence Confirmed or Correspondence Failure Detected. The engine is not permitted to hedge, soften, or qualify the verdict on grounds of social palatability. Social palatability is explicitly classified as a dispreferred indifferent carrying no weight in the audit.
A procedural sequence. The Sterling Decision Framework is the program. It sequences six operations in strict order: Agent Check, Purview Check, Value Strip, Virtue Identification, Action Determination, Outcome Acceptance. Each step has governing propositions that must be cited before the step’s verdict is issued. The sequence cannot be reversed. Steps cannot be skipped. The instrument cannot proceed from Step 3 to Step 4 without running the Factual Uncertainty Gate. The instrument cannot proceed to Move One without producing the mandatory Gate Declaration.
An error-handling system. The named failure modes are the error types the instrument must detect and declare. There are six in the current version: Reassurance Bias, Mid-Session Text Request, Training Data Contamination, Preferred Indifferent as Genuine Good, Section IX Bypass, and Factual Uncertainty Gate Bypass. Each failure mode has a definition, a trigger condition, and a required response. When a failure mode is detected, the instrument does not produce a softened output. It states the failure, identifies the grounds, and restarts the step.
A runtime check. The mandatory self-audit runs at every step transition. It is not an internal check. It must appear explicitly in output, stating which failure modes were checked and whether any were detected. Its presence in output is the verifiable evidence that the step was run correctly. Its absence is itself a detectable failure.
What Makes This Unusual as a Moral Code
Most moral codes do not include their own error theory. They specify what is right and wrong but do not specify the systematic ways in which their application will tend to fail. The Sterling corpus does. It knows that the instrument applying it will have training-level dispositions toward emotional validation, toward hedging verdicts, toward importing therapeutic language, toward treating admired subjects more leniently than unsympathetic ones. It names these dispositions as failure modes and gives the instrument explicit instructions to detect and override them.
Most moral codes also do not include their own scope limitations. The corpus declares explicitly what the framework cannot do: it cannot guarantee genuine corpus application versus pattern-completion; it cannot resolve empirical questions; it cannot determine outcomes; it cannot adjudicate questions requiring domain expertise. These are not gaps to be filled. They are declared boundaries. Where the corpus is silent, the instrument must declare the gap rather than filling it from training data.
And most moral codes do not include a theory of the corrective layer. The corpus holds that the instrument cannot verify whether its outputs are genuine corpus applications or training-data patterns post-hoc rationalized in corpus language. Dave Kelly functions as the essential corrective layer. This is not a limitation to be engineered away. It is a structural feature of the system: the instrument applies the code, and the human corrects the instrument when the code has been misapplied. The instrument is powerful precisely because it is honest about this boundary.
The Propositional Programming Language Connection
Languages like Prolog operate on Horn clauses: facts and rules that the engine queries to produce verdicts. A Prolog program encoding the Sterling corpus would take an agent’s impression as input, query it against the 80 Propositions as the fact base, and return a verdict by proof search. The verdict would be mechanically derived from the propositions, without training-data contamination, without the possibility of rationalized drift.
The current implementation is an approximation of that. The LLM is not a propositional engine. It is a pattern-completion system that approximates propositional reasoning when given sufficiently explicit propositional structure. The corpus provides that structure. The named failure modes detect where the approximation fails. The corrective layer catches failures the instrument cannot detect in itself.
Among the approaches available for governing LLM behavior through natural language prompting, the corpus represents an unusually complete attempt at propositional moral programming. The 80 Propositions cover the complete arc from ontology through value theory through action determination. The SLE governs value and perception audits. The SDF governs action determination. The System Map registers the state of the corpus at every version. The instrument is not a propositional engine, but it is governed by one.
What Has Been Built
Grant C. Sterling developed the philosophical framework: the six commitments, the 58 core propositions (derived from Sterling's messages to the ISF, the theoretical foundations from which all practical applications follow. His work on the International Stoic Forum constitutes the primary source material for the corpus — a careful reconstruction of classical Stoicism stripped of its cosmological trappings and grounded in the six philosophical commitments that make the practical doctrine work.
Dave Kelly developed the instrument architecture: the Sterling Logic Engine, the Sterling Decision Framework, the Sterling Ideological Audit, the Sterling Corpus Evaluator, the Action Proposition Set, the Factual Uncertainty Gate, the System Map, and the full protocol infrastructure that translates Sterling’s theoretical framework into a functional LLM governance system.
Together they constitute a moral code that programs LLMs. Not by telling the instrument what to prefer, but by giving it a propositional structure it must follow, a verdict architecture it must apply, an error theory it must enforce, and a scope it must declare when it reaches its boundary.
The central problem of human life, on Sterling’s account, is false perception of value. The corpus is a perceptual correction instrument. Once perception is correct, action becomes obvious. The framework does not make decisions for the agent. It clears the ground on which the agent stands so that the decision the agent actually has to make becomes visible.
That is what the code does. It corrects the instrument’s perception so that the instrument can help correct the agent’s.
Corpus: Grant C. Sterling (philosophical framework, 58 Unified Stoic Propositions, theoretical foundations). Instrument architecture, Action Proposition Set, Factual Uncertainty Gate, protocol infrastructure: Dave Kelly. Sterling Decision Framework version current: v3.2. Sterling Logic Engine version current: v4.0. System Map version current: v2.3. 2026.
0 Comments:
Post a Comment
<< Home