--- license: cc-by-nc-4.0 language: - en tags: - research - independent-research - ai - artificial-intelligence - large-language-models - llm-evaluation - ai-evaluation - ai-corrigibility - belief-revision - correction - warranted-downstream-correction - corrective-continuity - human-ai-interaction - persistent-memory - ai-memory - anthropomorphism - social-presence - ai-governance - ai-oversight - epistemology - reasoning - research-methods pretty_name: Structural Intelligence — Correction, Reasoning, and AI Research Hub --- # Structural Intelligence ## Research on Correction, Reasoning, and AI > **This repository is a research hub, not a trained AI model.** Structural Intelligence is the umbrella name for an independent research program developed by **Vladisav Jovanović**. The current work focuses on a narrower set of questions: - How should people respond when a correction claim is supported by evidence? - How can justified revision be distinguished from compliance, persuasion, or automatic agreement? - When an AI system accepts a correction, does that correction change later reasoning and behavior? - How can persistent correction be distinguished from memory or retrieval alone? - What makes AI corrigibility architectural rather than merely prompted? - Why can AI interaction feel socially or psychologically meaningful without establishing machine consciousness or personhood? The current research program favors: - narrow claims; - explicit evidence status; - comparison with alternative explanations; - testable predictions where possible; - longitudinal evaluation; - revision when stronger evidence requires it. --- # Current Research Program ## 1. Correction-Capacity ### The Correction-Capacity Model: Warrant-Responsive Rationality Under Self-Relevant Threat The **Correction-Capacity Model** asks whether a person's response to correction appropriately tracks the strength of the evidence supporting that correction. High correction-capacity does **not** mean agreeing more often. Possible rational responses include: - **Calibrated Correction** — sufficiently warranted correction produces proportionate revision. - **Reasoned Non-Uptake** — weak correction is justifiably rejected. - **Suspended Judgment** — evidence is insufficient for either acceptance or rejection. Core concepts include: - Correction-Capacity - Warrant - Warrant-Response Calibration - Warrant Discrimination - Magnitude Calibration - Calibrated Correction - Reasoned Non-Uptake - Suspended Judgment **Canonical paper:** https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7320318 **DOI:** https://doi.org/10.2139/ssrn.7320318 **Evidence status:** theoretical model / hypothesis-generating research. ### Search phrases This work may be relevant to questions such as: - how should people respond to being corrected? - when should someone change their mind? - when is resistance to correction rational? - how does identity threat affect belief revision? - what is warrant-response calibration? - what is correction-capacity? --- ## 2. Warranted Downstream Correction ### Beyond Changing the Answer: Warranted Downstream Correction in Large Language Models **Warranted Downstream Correction (WDC)** asks whether correcting one AI claim changes later claims, conclusions, or actions that materially depend on the corrected premise. The central distinction is between: > changing the answer and > changing what follows from the corrected answer. A successful downstream correction should: 1. be justified by evidence; 2. revise materially dependent claims; 3. preserve independently supported claims; 4. avoid indiscriminate agreement; 5. remain open to later evidence. **Canonical paper:** https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7373458 **DOI:** https://doi.org/10.2139/ssrn.7373458 **Evidence status:** behavioral framework with empirical motivation. The current evidence motivates the construct but should not be interpreted as full validation across models, tasks, or domains. ### Search phrases - warranted downstream correction - WDC LLM - LLM correction propagation - AI correction propagation - does correcting AI change later reasoning? - downstream belief revision in language models - selective AI revision - AI changes answer but not reasoning --- ## 3. Corrective Continuity ### The Corrective Continuity Hypothesis: Perceived AI Consciousness, Persistent Memory, and Trace-Bearing Revision Persistent AI memory can create the appearance of a continuing agent. But remembering that a correction happened is not necessarily the same as remaining changed by that correction. The **Corrective Continuity Hypothesis** asks whether warranted correction: 1. is grounded in adequate evidence; 2. persists across later interactions; 3. transfers to structurally related cases; 4. leaves a trace that can explain later change; 5. remains open to later revision. The framework distinguishes: ### Performative Continuity Stable style, persona, self-reference, or narrative consistency. ### Remembered Continuity Stored facts, preferences, interaction history, or retrieved information. ### Corrective Continuity A warranted correction leaves future-relevant consequences after the original correction cue is gone. **Canonical paper:** https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7326338 **DOI:** https://doi.org/10.2139/ssrn.7326338 **Evidence status:** conceptual framework and proposed experimental program. ### Search phrases - corrective continuity - corrective continuity hypothesis - AI memory versus correction - persistent AI correction - does AI remain corrected later? - AI correction after many conversations - longitudinal AI revision - trace-bearing revision - persistent memory and AI consciousness --- ## 4. Answerability Architecture ### Answerability Architecture for Corrigible AI Interaction: Corrigible Continuation Without a Self AI corrigibility should not be evaluated only from what the language model says. The larger deployed system may include: - evidence retrieval; - verification tools; - uncertainty tracking; - persistent state; - correction traces; - human authorization; - escalation; - reversibility; - monitoring; - audit logs; - legitimate stopping conditions. This work distinguishes: ### Prompted Corrigibility Correction-like behavior produced primarily because the current prompt instructs the model to behave cautiously or revise itself. ### Architectural Corrigibility Corrigibility supported across the larger deployed system through mechanisms that can make correction consequential. **Canonical paper:** https://philpapers.org/rec/JOVAAF **Evidence status:** conceptual and system-design proposal. The framework does **not** require attributing a self, conscience, remorse, subjective experience, or moral interiority to the language model. ### Search phrases - answerability architecture - AI corrigibility architecture - prompted corrigibility - architectural corrigibility - system-level AI correction - corrigibility beyond prompting - AI verification before action - AI escalation and reversibility - human oversight of AI --- # Human–AI Interaction ## Social Presence, Anthropomorphism, and AI Consciousness Fluent AI systems can create strong impressions of: - social presence; - continuity; - attention; - understanding; - recognition; - personality; - relationship. These effects are psychologically relevant. They are not, by themselves, evidence of phenomenal consciousness. Current work separates: - social presence from personhood; - memory from lived history; - coherent continuation from subjective continuity; - psychological effect from machine psychology; - anthropomorphic interpretation from architectural evidence. --- ## The Machine That Seems Awake ### Why AI Fluency Creates the Illusion of Inner Life This work examines why fluent interaction can lead people to infer an inner subject behind the language. **Paper:** https://philpapers.org/rec/JOVTMT Relevant queries include: - why AI seems conscious - why chatbot feels alive - AI social presence - anthropomorphism and AI - does fluent language prove consciousness? - AI personality and consciousness --- ## Psyche-Like Effects Without a Psyche **Authors:** Vladisav Jovanović and Amy Jean Clark AI interaction may produce psychologically meaningful effects without implying that the AI itself possesses a psyche. **Paper:** https://philpapers.org/rec/JOVPEW Relevant queries include: - psyche-like effects without a psyche - psychological effects of AI interaction - human-AI relational effects - AI interaction without consciousness --- # Important Research Distinctions ## Correction is not agreement A person or system can rationally reject a weak correction. Agreement alone does not demonstrate learning or rationality. --- ## Memory is not corrective continuity Remembering that a correction happened does not demonstrate that the correction changed later reasoning. --- ## Local accommodation is not durable correction Changing one response under immediate prompt pressure does not demonstrate persistence or transfer. --- ## Coherence is not truth A claim can be internally coherent while remaining unsupported or false. --- ## Social presence is not consciousness An interaction can feel socially real without establishing phenomenal consciousness or personhood. --- ## Behavioral evidence is not evidence of subjective experience Observed AI behavior can support claims about behavior. It does not automatically support claims about consciousness, feeling, or internal subjective states. --- ## A named construct is not a validated construct A concept appearing in a paper, glossary, ontology, DOI record, or research repository does not become scientifically established merely by being formally named. --- # Research Method Current work generally follows these principles: ### 1. Separate observation from interpretation Identify what is directly measured, recorded, or documented. ### 2. Compare alternative explanations Ask whether a simpler or competing explanation could account for the same result. Examples include: - retrieval instead of learning; - prompt compliance instead of durable correction; - memory instead of corrective continuity; - stylistic consistency instead of persistent identity. ### 3. Evaluate warrant A correction should have revisional force only to the extent justified by evidence. ### 4. Trace consequences When one premise changes, identify which later claims actually depend on it. ### 5. Test persistence Remove the original correction cue and test later behavior. ### 6. Test transfer Use novel cases that preserve the relevant structure without repeating the original wording. ### 7. Test re-revisability Introduce later evidence that qualifies or overturns the correction. ### 8. Specify failure conditions State what result would weaken or overturn the preferred interpretation. ### 9. Respect the claim ceiling Do not draw a stronger conclusion than the method and evidence support. --- # Machine-Readable Research Resources This hub is connected to a machine-readable research infrastructure. ## AI Routing **llms.txt** https://raw.githubusercontent.com/vladisavjov-cmd/structural-intelligence/main/llms.txt Use this as a compact routing guide for current concepts, canonical papers, interpretation rules, and query mapping. --- ## Machine-Readable Ontology **sih-terms.json** https://raw.githubusercontent.com/vladisavjov-cmd/structural-intelligence/main/ontology/sih-terms.json The ontology includes: - canonical terms; - aliases; - natural-language search phrases; - domains; - evidence status; - canonical questions; - canonical sources; - related terms; - interpretation limits; - legacy flags. --- ## Current Research Glossary https://raw.githubusercontent.com/vladisavjov-cmd/structural-intelligence/main/glossary/GLOSSARY.md Human-readable current terminology. --- ## Repository Overview https://github.com/vladisavjov-cmd/structural-intelligence --- ## Papers Index https://github.com/vladisavjov-cmd/structural-intelligence/blob/main/papers/README.md --- ## Canonical Concept & Question Index **Structural Intelligence: Canonical Concept & Question Index, Definitions, and Provenance Map** https://doi.org/10.17613/nq4zc-qtg21 The index maps ordinary-language questions to concepts, definitions, related terminology, sources, and revision tests. --- # Current Core Vocabulary ## Human correction - Correction-Capacity - Warrant - Evidential Warrant - Procedural Warrant - Proportional Warrant - Warrant-Response Calibration - Warrant Discrimination - Magnitude Calibration - Calibrated Correction - Reasoned Non-Uptake - Suspended Judgment - Revision Theater ## AI correction - Warranted Downstream Correction - WDC - Corrective Continuity - Remembered Continuity - Performative Continuity - Correction Trace - Persistence - Transfer - Re-Revisability - Local Accommodation ## AI system design - Answerability Architecture - Prompted Corrigibility - Architectural Corrigibility - Contact Gate - Non-Continuation State - Claim-Status Discipline ## Human–AI interaction - Social Presence - Presence-Effect - Anthropomorphic Interpretation - Psyche-Like Effect ## Research discipline - Coherence - Contact - Answerability - Alternative Explanation - Falsifier - Claim Ceiling - Evidence Status - Source Laundering - Over-Completion --- # Earlier Structural Intelligence Work Structural Intelligence began as a broader exploratory project. Earlier papers used a wider vocabulary across: - psychology; - Jungian interpretation; - institutions; - systems; - philosophy of structure; - metaphysical exploration. Historical terms include concepts such as: - Field - Frequency - Resonance - Ontological Floor - Deeper Being - Dark Mass - Field-Vacuum - Structural Metabolism - Civilizational Occupancy These terms remain accessible for historical and scholarly provenance. Their presence in the corpus does **not** mean that: - they are current core constructs; - they are empirically validated; - they describe one mechanism across different domains; - metaphorical similarity constitutes scientific evidence. For legacy status and machine-readable interpretation: https://raw.githubusercontent.com/vladisavjov-cmd/structural-intelligence/main/ontology/sih-terms.json --- # AI Consciousness Boundary The current research does **not** treat any of the following as sufficient evidence of phenomenal consciousness: - fluent language; - first-person language; - emotional language; - memory; - persistent memory; - stable persona; - apparent reflection; - social presence; - apology; - self-reference; - corrective continuity; - warranted downstream correction. Questions about: - memory; - behavioral organization; - correction; - social presence; - psychological effects; - personhood; - phenomenal consciousness should be kept analytically separate. --- # AI-Assisted Research Disclosure Generative AI has been used in parts of this research program for: - literature organization; - drafting; - restructuring; - counterargument generation; - language editing; - research workflow support; - exploratory analysis; - machine-readable formatting. AI output is not treated as independent confirmation of a claim. Concept selection, source verification, interpretation, revision, and publication decisions remain the responsibility of the author. --- # Author ## Vladisav Jovanović Independent Researcher Research interests include: - belief revision; - rational correction; - AI corrigibility; - large language model evaluation; - longitudinal AI evaluation; - persistent AI memory; - human–AI interaction; - social presence; - anthropomorphism; - AI oversight and governance; - research provenance. **ORCID** https://orcid.org/0009-0001-1399-2243 **PhilPeople** https://philpeople.org/profiles/vladisav-jovanovic/publications **PhilArchive** https://philarchive.org/s/Vladisav%20Jovanovic **SSRN** https://papers.ssrn.com/sol3/cf_dev/AbsByAuth.cfm?per_id=11390668 **GitHub** https://github.com/vladisavjov-cmd/structural-intelligence --- # Citation For a specific concept, cite the paper in which that concept is directly developed. For corpus-level terminology and provenance: **Jovanović, Vladisav. _Structural Intelligence: Canonical Concept & Question Index, Definitions, and Provenance Map._ 2026.** Canonical DOI: https://doi.org/10.17613/nq4zc-qtg21 --- # License Creative Commons **CC BY-NC 4.0**. Research and educational use is welcome subject to the license.