"Why is my AI hallucinating — Could I be the cause?"

Humans and AI systems both tend to fill gaps with assumptions. When an AI “hallucinates,” it is often 'not inventing freely', —it is completing patterns left ambiguous by human input. A reasonable question follows: “Why is my AI hallucinating — and could I be the cause?” A FUNt / Physmatics series addition.Public Tooling Overview: Words → Math → Simulation Version: v1.0 — January 10, 2026 Michael K. Nowlin Author’s Statement- When an AI “hallucinates,” it is often not inventing freely,it is completing patterns left ambiguous by human input. A reasonable question follows: “Why is my AI hallucinating — and could I be the cause?” This project emerged from confronting that question directly during the mathematical audit of: FUNt Core Physics - https://doi.org/10.5281/zenodo.17497241 The Core Problem: Drift Natural language is powerful, but it is subjective and fluid. Words allow meaning to shift without notice. Mathematics, by contrast, is humanity’s shared precision language. A mathematical statement is explicit: it declares symbols, relations, domains, and boundaries. When those elements are missing, both humans and AI tend to supply them implicitly. This leads to a specific failure mode: AI-assisted research drift — when code runs successfully but no longer corresponds to the original claim. The problem is not computation. The problem is under-specified statements. Why Two Tools Were Needed ? During the development of FUNt Core Physics, many claims were written in words and then tested computationally. A consistent pattern emerged: When a claim was not grounded as an explicit mathematical object(variables + relation/operator + domain/boundary) both humans and AI systems filled in gaps with assumptions, producing simulations that were executable, but not necessarily faithful to the intended idea. To prevent this, the audit process had to be split into two distinct layers, each with a single responsibility. This separation produced two public tools (colab notebooks attached as part of this record): HUG — to force clarity in statements AUDIT — to execute only what has been stated The Three-Layer Architecture (No Loops) The workflow is intentionally linear: HUG:Words → explicit mathematical statement (with declared domains and boundaries) AUDIT:Mathematical statement → simulation outputs (numbers, diagnostics, plots) Results / Interpretation:Outputs → optional real-world meaning (kept outside the notebooks) This structure prevents circular reasoning. The notebook does not interpret results, and the statement tool does not invent mathematics. HUG — What It Does and Does Not Do HUG (Hydrogenic Universe Gate) is a clarity tool, not a judge. What HUG does Helps convert an idea written in words into an explicit mathematical statement Forces hidden assumptions to surface Ensures a claim is executable before simulation Encourages use of AI as a coach, not an oracle What HUG does not do HUG does not invent theory HUG does not decide truth HUG does not interpret results HUG returns one of three statuses: Rooted — the statement is explicit and executable Go Deeper — the statement exists but needs refinement Not Evaluated — the statement cannot yet be written as math Minimal Admissibility Template (Ledger-Ready) N. Title Status: Rooted | Go Deeper | Not Evaluated Claim: mathematical statement Domain / Boundary / Range: where it applies Reason: optional clarification A Simple HUG Example (Non-Physics) Idea (words): “This update rule tends to lock into a small set of angles.” HUG-resolved question (math-only): Given a recurrence on the circle, does the post–burn-in trajectory concentrate on a low-period orbit? One admissible form: Let θ₀ ∈ [0, 2π). Define θₙ₊₁ = F(θₙ). After burn-in B, define P = #distinct{θₙ : n ≥ B} within tolerance ε. Question: for the declared parameters, is P ≤ Pₘₐₓ? HUG’s role is to force F, B, ε, and Pₘₐₓ to be explicit so the claim can be executed without interpretation. AUDIT — What It Does and Does Not Do AUDIT is a frozen, repeatable mathematical instrument. What AUDIT does Executes an explicitly declared recurrence Reports numerical and qualitative diagnostics Produces reproducible outputs (including JSON for ledgers) What AUDIT does not do It does not claim physical truth It does not interpret outputs It does not validate or invalidate FUNt as a whole Each AUDIT notebook version is intentionally fixed. Subject Under Audit (Math-Only) The current audit notebook evaluates the recurrence: \theta_{n+1} = \psi\big(R_{\beta_n}(\theta_n)\big), \qquad \beta_n = (0, \pi/3, 0, \pi/3, \dots) with: R_\beta(\theta) = W(2\beta - \theta), \quad W(x) = x \bmod 2\pi \in [0, 2\pi). The specific operator \psi is explicitly defined in code and treated as locked for that notebook version,(e.g., “Option A — POLY3 projector”). Why This Audits FUNt (Without Claiming Physics) FUNt Core Physics introduces operators and update rules that, even without interpretation, define mathematical dynamics. Before any physical meaning is argued, those dynamics must be: executable, inspectable, and repeatable by independent parties. That is the purpose of AUDIT. The audit does not “prove FUNt.” It establishes a transparent computational record of what the declared mathematics does. How to Run (Moms / Dads Version) Open the notebook in Google Colab (or Jupyter) Press Run all Read the printed summaries and view the plots To test sensitivity, change only declared run settings and rerun If you disagree with the outputs, rerun with different parameters or implement the same recurrence independently. Transparency is the goal. Scope Boundaries Interpretation is intentionally excluded from the notebooks. Any mapping of operators (for example, ψ → physical constructs) belongs in a separate interpretive document. This governance choice prevents narrative contamination of the simulation record. Public Deliverables and Permanence Six versioned artifacts are provided: HUG (notebook) + README — statement admissibility and AI-assisted clarification AUDIT (notebook) + README — frozen execution and diagnostics Bridge / Results document — contextual discussion and selected outcomes FUNt Mathematical Audit Notebook Energy Ladder (φ-indexed) RESULT Zenodo uploads are treated as permanent records of FUNt / Pjysmatics. Closing; The guiding principle is simple: Reduce ambiguity before running code. HUG reduces ambiguity in statements. AUDIT reduces ambiguity in execution. Together, they provide a clean, public audit trail that others can reproduce without reliance on trust, authority, or hidden assumptions.