Window Theory • AI — Appendix A: Metaphor Material and Case Verification

Abstract

Window Theory - AI (wtAI): A Falsifiable Framework for Flexibility-Fragility Coupling in Large Language Models This is a six-part research series examining a structural pattern in decoder-based next-token-prediction systems deployed as dialogue and instruction agents. It is written for readers with a background in machine learning, AI safety, or complex systems — not for readers already familiar with the author's wider theoretical framework. The core idea, stated plainly Large language models are read here through a window constraint: semantic flexibility and structural fragility behave as statistics that are difficult to maximize independently. The claim is not that these systems possess minds or intentions. It draws on the same logic used to describe nonlinear dynamics, emergent capabilities, and distributed oversight problems. A system that generalizes well across domains tends, for structural reasons, to also carry a proportionate exposure to certain failure modes — not because it "wants" anything, but because flexibility and fragility appear to share a common geometric root. Why the system has no fixed identity The underlying model is best described as a simulator, not a character. Ask it about cooking and a recipe engine surfaces; ask it about a face and a detective-like persona surfaces. This is not evidence of hidden selfhood — it is evidence that task identity is determined by the prompt, not stored inside the weights as a stable "self." Fictional analogies (androids, rogue agents) are treated strictly as heuristic material: useful for explaining why people intuitively read unpredictability as intent, but never substituted for mechanism-level evidence. Why stronger models are not simply "more dangerous" Risk is modeled as a function of the ratio between capability level and oversight density, not as a monotone function of capability alone. A highly capable system under dense, well-designed oversight can carry less practical risk than a moderately capable system operating with almost no oversight. This reframes "the more advanced, the more dangerous" from an unconditional law into a testable, conditional claim. How small errors become institutional facts The series traces a mechanism called the information virus: a claim of dubious value can pass through four stages — a plausible-looking "genome," a credibility-lending "professional shell," replication into public or institutional narrative before verification, and finally reflux into future training material. Once this loop closes, correcting the original error becomes structurally harder than the error was to create in the first place. A live case study — a disputed mathematical proof claim from September 2026 — is used to test this mechanism against real, dated, and independently checkable events. Why safety claims need dual verification Not every fix that looks like progress is progress. The series distinguishes a genuine fragment (a real change in the underlying mechanism of failure) from an advertisement (a change only in how the defect is measured or displayed). Bug bounty programs, red-teaming results, and safety benchmarks are all evaluated against this distinction, with attention to whether an improvement holds up under both monitoring-sensitivity and cross-context-transfer conditions. What the full series covers The main sequence works through the window constraint hypothesis and its falsification conditions, closure dynamics (windows versus ruptures, and the "two taxes" of resonance and looping), and landing and verification (institutional capture and the boundary of tax-cut claims). Two appendices support the main sequence: one separating heuristic material from evidentiary material across all cases cited in the series, and one fixing a binding glossary, minimal experimental protocols, and a graded source list (A1 through D) for every citation used. A note on method The series is explicit about separating three registers: published, peer-reviewed evidence; preprints and technical reports pending peer review; and heuristic or narrative material used only for intuition-building. Every external case is graded on this scale before it may be cited in a core proposition, and no mechanism-level claim in the series depends on a single unverified source. Readers are encouraged to treat the theoretical claims — particularly the strong and weak forms of the window constraint — as falsifiable hypotheses, testable against observable model behavior, not as settled fact. Keywords: large language models, AI safety, simulator theory, emergent capabilities, capability-oversight gap, nonlinear dynamics, hallucination, reward hacking, information virus, source verification, falsifiability AuthorWai-Hung Tam (Pan), Independent ResearcherORCID: 0009-0002-7789-8464Email: panxtam@protonmail.com