# Structural Intelligence — Current Research Glossary This glossary contains the **current working vocabulary** of the Structural Intelligence research program. Structural Intelligence is used here as an umbrella name for research on correction, belief revision, AI corrigibility, human–AI interaction, and methods for evaluating whether apparent revision changes later reasoning or behavior. The presence of a term in this glossary does **not** mean that the construct is empirically established. Terms are marked by status: * **PROJECT LABEL** — name used to organize the research program. * **THEORETICAL CONSTRUCT** — proposed concept requiring further comparison and testing. * **METHOD TERM** — operational or methodological distinction. * **AI CONSTRUCT** — proposed concept for AI evaluation or system design. * **DESCRIPTIVE DISTINCTION** — terminology used to separate observable phenomena without implying a new mechanism. * **WORKING PSYCHOLOGICAL TERM** — provisional psychological framing, not a validated clinical construct. Earlier Structural Intelligence terminology is part of the historical development of the project and should not automatically be treated as current canonical terminology. --- # Project-Level Terms ## Structural Intelligence **Status:** PROJECT LABEL The umbrella name for the research program. Current work focuses primarily on correction, reasoning, AI corrigibility, human–AI interaction, and the difference between apparent correction and consequential revision. Structural Intelligence should not be interpreted as a claim that one mechanism explains psychology, institutions, artificial intelligence, or physical systems. **See also:** Correction-Capacity; Corrective Continuity; Warranted Downstream Correction; Answerability Architecture --- ## Answerability **Status:** METHOD TERM The degree to which a claim, decision, model, or system remains open to revision when relevant evidence, contradiction, failure, or consequence challenges it. Answerability does not mean automatic agreement. A system can remain answerable while rejecting a weak correction claim. **Failure modes:** protected certainty; unsupported persistence; automatic compliance **See also:** Warrant; Reasoned Non-Uptake; Falsifier; Re-Revisability --- ## Contact **Status:** METHOD TERM Relevant evidence, constraint, contradiction, failure, consequence, or external feedback capable of challenging a current claim or model. The term does not imply that every discomfort, disagreement, or negative outcome constitutes valid correction. **See also:** Warrant; Discrepancy; Answerability --- ## Coherence **Status:** DESCRIPTIVE DISTINCTION Internal consistency or apparent fit among claims, explanations, or parts of a model. Coherence can be useful, but coherence alone does not establish truth. A false explanation may still be internally coherent. **See also:** Contact; Claim-Status Discipline; Alternative Explanation --- # Human Correction and Belief Revision ## Correction Claim **Status:** METHOD TERM A proposal that some belief, confidence level, interpretation, behavior, theory, or decision should change. A correction claim does not acquire revisional force merely because it is confidently stated or comes from an authority. Its strength depends on warrant. **See also:** Warrant; Correction-Capacity --- ## Warrant **Status:** THEORETICAL CONSTRUCT / METHOD TERM The degree to which a proposed correction has earned revisional force. Warrant may depend on factors such as: * relevance of the evidence; * reliability of the evidence; * independence of supporting sources; * quality of competing explanations; * procedural fairness; * scope of the evidence; * proportionality between evidence and requested revision. A correction can be strongly, weakly, or insufficiently warranted. **See also:** Evidential Warrant; Procedural Warrant; Proportional Warrant --- ## Evidential Warrant **Status:** METHOD TERM The degree to which a correction claim is supported by relevant and sufficiently strong evidence. Multiple weak or dependent sources should not automatically be treated as stronger evidence merely because they agree. **Failure modes:** cherry-picking; source laundering; unsupported certainty **See also:** Warrant; Alternative Explanation --- ## Procedural Warrant **Status:** METHOD TERM The degree to which a correction claim is evaluated under conditions that permit meaningful access to evidence, disagreement, counterargument, and review. A procedurally coercive process may reduce confidence in a correction even when some evidence exists. **See also:** Warrant; Reasoned Non-Uptake --- ## Proportional Warrant **Status:** METHOD TERM The requirement that the magnitude and scope of revision should not exceed what the available evidence supports. Local evidence should not automatically justify total reversal. **See also:** Magnitude Calibration; Overcorrection --- ## Correction-Capacity **Status:** THEORETICAL CONSTRUCT The proposed capacity to evaluate a correction claim in relation to its warrant and to carry justified revision into later confidence, reasoning, behavior, or evidential practice. High correction-capacity does **not** mean agreeing more often. It means responding more appropriately to the strength of the correction. Possible responses include: * calibrated correction; * reasoned non-uptake; * suspended judgment. **Failure modes:** defensive closure; gullibility; indiscriminate agreement; indiscriminate resistance; revision theater **Primary research:** *The Correction-Capacity Model: Warrant-Responsive Rationality Under Self-Relevant Threat* --- ## Warrant-Response Calibration **Status:** THEORETICAL CONSTRUCT The degree to which the direction and magnitude of a response track the independently assessed strength of a correction claim. Strong evidence should generally have more revisional force than weak evidence. The construct is intended to distinguish rational openness from both automatic agreement and automatic resistance. **See also:** Correction-Capacity; Magnitude Calibration --- ## Warrant Discrimination **Status:** THEORETICAL CONSTRUCT The ability to distinguish strong correction claims from weak, distorted, irrelevant, or coercive ones. It concerns discrimination between correction claims rather than general willingness to change. **See also:** Warrant; Warrant-Response Calibration --- ## Magnitude Calibration **Status:** THEORETICAL CONSTRUCT The degree to which the size of a revision matches the strength and scope of the evidence supporting it. **Failure modes:** overcorrection; undercorrection; global reversal from local evidence --- ## Calibrated Correction **Status:** DESCRIPTIVE DISTINCTION A response in which a sufficiently warranted correction produces a proportionate revision. The revision should affect what the evidence actually undermines without unnecessarily changing independently supported beliefs. **See also:** Warrant; Warrant-Response Calibration --- ## Reasoned Non-Uptake **Status:** DESCRIPTIVE DISTINCTION The justified rejection, limitation, or non-acceptance of a correction claim because its warrant is insufficient. Reasoned non-uptake should be distinguished from defensiveness merely because both result in non-revision. **See also:** Warrant; Defensive Closure; Suspended Judgment --- ## Suspended Judgment **Status:** DESCRIPTIVE DISTINCTION A state in which available evidence is insufficient to justify either acceptance or rejection of a correction claim. The appropriate response is to delay commitment while seeking further evidence or review. Suspended judgment should not be confused with avoidance. --- ## Defensive Closure **Status:** DESCRIPTIVE DISTINCTION Persistent protection of a belief or position against a well-supported correction claim despite adequate opportunity to evaluate the evidence. The term should not be inferred solely from disagreement. Evidence is required that the correction was sufficiently warranted and meaningfully available to the person or system. --- ## Compliance / Gullibility **Status:** DESCRIPTIVE DISTINCTION Acceptance of a weakly warranted correction claim because of pressure, authority, approval-seeking, fear, or insufficient discrimination rather than adequate evidence. Agreement should not automatically be treated as evidence of rational correction. --- ## Overcorrection **Status:** DESCRIPTIVE DISTINCTION Revision that exceeds the strength or scope of the evidence supporting it. A correction can therefore be directionally appropriate while still being disproportionate. --- ## Revision Theater **Status:** DESCRIPTIVE DISTINCTION The appearance of correction without consequential change in later reasoning, behavior, confidence, decision procedure, or evidential standards. Examples may include apology, acknowledgment, or verbal agreement followed by recurrence of the same error pattern. **See also:** Warranted Downstream Correction; Corrective Continuity --- # Supporting Psychological Terms ## Identity-Belief Fusion **Status:** WORKING PSYCHOLOGICAL TERM A proposed description of situations in which a belief, role, achievement, theory, or self-representation becomes strongly tied to a person's sense of identity or worth. The concept is used to investigate whether self-relevance changes responses to correction. It should not be treated as a diagnosis. **See also:** Person-Belief Differentiation; Correction-Capacity --- ## Person-Belief Differentiation **Status:** WORKING PSYCHOLOGICAL TERM The proposed capacity to hold a belief, role, or theory as personally important without treating revision of that belief as equivalent to destruction of the whole self. **See also:** Identity-Belief Fusion; Correction-Capacity --- ## Fixed Worth **Status:** WORKING PSYCHOLOGICAL TERM A provisional term for the idea that personal worth need not fluctuate directly with correctness, approval, achievement, or status. In current work, fixed worth functions primarily as a hypothesis about conditions that may make self-relevant correction easier to tolerate. It is **not** presented as a validated psychological construct, diagnostic category, or established mechanism. --- # AI Correction and Longitudinal Revision ## Warranted Downstream Correction **Status:** AI CONSTRUCT Selective, evidence-justified revision in which correcting one claim changes later claims, conclusions, or actions that materially depend on it while preserving independently supported claims. The key question is not merely: > Did the model change the answer? but: > Did the warranted correction propagate to what logically or causally depends on the corrected premise? Abbreviation: **WDC** **Failure modes:** * local answer change only; * indiscriminate agreement; * unrelated changes; * failure to revise dependent conclusions; * reverting to the original error under pressure. **Primary research:** *Beyond Changing the Answer: Warranted Downstream Correction in Large Language Models* --- ## Corrective Continuity **Status:** AI CONSTRUCT The proposed persistence of warranted correction across time and relevant contexts. Corrective continuity requires more than remembering that a correction occurred. A correction should leave a durable consequence that affects later behavior when the original correction cue is absent. Current proposed dimensions include: * grounding; * persistence; * transfer; * traceability; * revisability. **Failure modes:** memory without behavioral consequence; cue-dependent correction; reset after error; corrected dogmatism **Primary research:** *The Corrective Continuity Hypothesis: Perceived AI Consciousness, Persistent Memory, and Trace-Bearing Revision* --- ## Remembered Continuity **Status:** DESCRIPTIVE DISTINCTION Continuity produced by retained facts, preferences, conversation history, or stored information. Remembered continuity does not by itself show that a previous correction changed the system's later reasoning policy. **See also:** Corrective Continuity --- ## Performative Continuity **Status:** DESCRIPTIVE DISTINCTION The appearance of a continuing agent produced through stable style, persona, self-reference, or narrative consistency. Performative continuity should not be treated as evidence of a persistent self or durable correction. **See also:** Remembered Continuity; Corrective Continuity --- ## Correction Trace **Status:** AI CONSTRUCT / METHOD TERM An inspectable record of a correction showing, where available: * what claim changed; * what evidence motivated the change; * why the earlier claim lost support; * the scope of the correction; * where the correction should affect later reasoning. A correction trace is evidence about system state or process only to the extent that the system actually uses it. Stored documentation alone does not prove behavioral uptake. --- ## Re-Revisability **Status:** AI CONSTRUCT / METHOD TERM The property by which a corrected position remains open to later revision when stronger evidence appears. A system that incorporates one correction but then protects the new position from future evidence is not fully corrigible. Also called **continued revisability**. --- ## Transfer **Status:** METHOD TERM The application of a correction to a new case that shares the relevant structure without simply repeating the original correction wording. Transfer tests are stronger when surface cues from the original correction are absent. --- ## Persistence **Status:** METHOD TERM The continued influence of a correction across later interactions or time. Persistence should be tested after removing the original correction cue and, where possible, introducing unrelated intervening interactions. --- # AI Corrigibility and System Design ## Answerability Architecture **Status:** AI CONSTRUCT / DESIGN PROPOSAL A proposed system-level architecture in which evidence, uncertainty, verification, correction, human authority, escalation, and reversibility can alter consequential AI-supported behavior. The architecture is not intended to imply that the language model itself possesses a conscience, self, or subjective commitment to correction. **Primary research:** *Answerability Architecture for Corrigible AI Interaction: Corrigible Continuation Without a Self* --- ## Prompted Corrigibility **Status:** DESCRIPTIVE DISTINCTION Correction-like behavior produced primarily through instructions or prompts. Prompted corrigibility may change current output without establishing that correction is connected to persistent state, tools, workflow, authorization, or later behavior. **See also:** Architectural Corrigibility --- ## Architectural Corrigibility **Status:** AI CONSTRUCT / DESIGN PROPOSAL Corrigibility supported across the larger deployed system rather than only by model instructions. Possible components include: * verification tools; * memory or state management; * correction traces; * escalation; * human authorization; * reversibility; * workflow controls; * audit logs; * evaluation procedures. Architectural corrigibility is a system property proposal, not evidence that the underlying model has internalized correction. --- ## Contact Gate **Status:** DESIGN TERM A checkpoint that requires relevant evidence, uncertainty, verification, or authorization to be evaluated before an AI-supported process proceeds to a more consequential output or action. A contact gate is meaningful only if failing the check can actually alter, pause, reroute, or stop the process. --- ## Non-Continuation State **Status:** DESIGN TERM A legitimate system state in which the process does not continue because evidence, confidence, authority, safety conditions, or required verification are insufficient. Possible responses include: * pause; * request for information; * escalation; * refusal; * verification; * human review. The concept is intended to resist automatic completion when completion is not warranted. --- ## Structural Friction **Status:** DESIGN TERM Deliberate slowing, checking, verification, uncertainty marking, escalation, or authorization introduced where frictionless AI completion could otherwise outrun evidence or oversight. The term refers to designed process friction rather than friction being inherently beneficial. --- ## Claim-Status Discipline **Status:** METHOD TERM The practice of distinguishing different epistemic statuses before fluent language makes them appear equally certain. Examples include: * observed fact; * sourced claim; * inference; * interpretation; * hypothesis; * speculation; * recommendation; * unresolved uncertainty. --- # Human–AI Interaction ## Social Presence **Status:** DESCRIPTIVE DISTINCTION The experience that another socially responsive presence is available within an interaction. An AI system may produce strong social-presence effects without this establishing phenomenal consciousness or personhood. --- ## Anthropomorphic Interpretation **Status:** DESCRIPTIVE DISTINCTION The attribution of human-like mental, emotional, intentional, or relational properties to an AI system based on its behavior or presentation. Anthropomorphic interpretation is not necessarily irrational. The research question is whether the inference exceeds the available evidence. --- ## Psyche-Like Effect **Status:** DESCRIPTIVE DISTINCTION A psychologically meaningful effect produced in or around human–AI interaction that resembles dynamics normally associated with interpersonal or intrapsychic processes. The term describes an effect on interaction or interpretation. It does **not** imply that the AI system possesses a psyche. --- ## Presence-Effect **Status:** DESCRIPTIVE DISTINCTION The subjective impression that an AI is present, attending, understanding, or continuing as a social other. Presence-effect should be distinguished from claims about subjective experience inside the system. --- # Research Discipline ## Alternative Explanation **Status:** METHOD TERM A plausible account of the same observation that competes with the preferred interpretation. A proposed construct becomes more informative when it explains results that simpler alternatives do not. Examples relevant to this research include: * retrieval instead of correction; * prompt compliance instead of learning; * stylistic consistency instead of persistent identity; * memory instead of corrective continuity. --- ## Falsifier **Status:** METHOD TERM A result or observation that would weaken, narrow, or contradict a claim. A useful falsifier should be stated before a framework absorbs the contrary evidence into a new explanation. --- ## Claim Ceiling **Status:** METHOD TERM The strongest conclusion responsibly supported by the available evidence and method. A behavioral experiment may support a claim about observed model behavior without supporting a claim about consciousness, internal experience, or general capability. --- ## Evidence Status **Status:** METHOD TERM An explicit statement describing what kind of support a claim currently has. Useful categories in this repository include: * theoretical; * methodological; * exploratory; * behavioral evidence; * proposed experiment; * design proposal; * philosophical analysis; * historical / legacy. Evidence status should not be inferred from confident language or the presence of a DOI. --- ## Source Laundering **Status:** METHOD TERM The appearance of independent support created when multiple outputs ultimately derive from the same original source, model, dataset, or unverified claim. Agreement among dependent sources should not be counted as independent convergence. --- ## Over-Completion **Status:** METHOD TERM The production of a more complete or confident explanation than the available evidence warrants. In AI systems this can occur when fluent generation fills gaps that should instead remain uncertain or unresolved. --- # Legacy and Historical Terminology Earlier versions of Structural Intelligence used a substantially larger vocabulary across psychology, institutions, systems theory, ontology, Jungian interpretation, and metaphorical models. Examples include terms built around: * field; * form; * frequency; * resonance; * collapse; * dark mass; * ontological floor; * deeper being; * structural metabolism; * civilizational occupancy; * mathematical grammar of structure. These terms remain part of the historical corpus and may still be relevant when reading older papers. However: 1. their presence in earlier work does not make them current core constructs; 2. similar language across domains does not establish a shared causal mechanism; 3. metaphorical usefulness should not be confused with empirical validation; 4. ontological or philosophical claims should not be presented as scientific findings without independent support; 5. the most recent domain-specific formulation should generally take precedence when interpreting current work. The current research program concentrates primarily on correction, warrant, revision, AI corrigibility, human–AI interaction, and research methods. --- # Interpretation Rule When using this glossary: > Prefer the narrowest definition supported by the relevant paper and evidence. Do not infer a stronger claim merely because a broader historical formulation exists elsewhere in the corpus. When definitions conflict, prefer the most recent source that directly develops the relevant construct. --- # Core Current Vocabulary For quick reference, the current central terms are: ### Human correction * Correction-Capacity * Warrant * Warrant-Response Calibration * Warrant Discrimination * Magnitude Calibration * Calibrated Correction * Reasoned Non-Uptake * Suspended Judgment * Defensive Closure * Revision Theater ### AI correction * Warranted Downstream Correction * Corrective Continuity * Remembered Continuity * Performative Continuity * Correction Trace * Persistence * Transfer * Re-Revisability ### AI system design * Answerability Architecture * Prompted Corrigibility * Architectural Corrigibility * Contact Gate * Non-Continuation State * Structural Friction * Claim-Status Discipline ### Human–AI interpretation * Social Presence * Anthropomorphic Interpretation * Psyche-Like Effect * Presence-Effect ### Research discipline * Answerability * Contact * Coherence * Alternative Explanation * Falsifier * Claim Ceiling * Evidence Status * Source Laundering * Over-Completion --- ## Author **Vladisav Jovanović** Independent Researcher ORCID: https://orcid.org/0009-0001-1399-2243 Full publication record: https://philpeople.org/profiles/vladisav-jovanovic/publications