Pre-Stabilisation Signals in Complex Systems: An Empirical Protocol for Testing Governance Sufficiency and an Invitation to the Research Community

Papers 1 and 2 of the Verbanatomy series established a conceptual vocabulary for pre-stabilisation dynamics and a mathematical framework for governance sufficiency. This paper does something different. It proposes that AI drift, hallucination, reward hacking, goal misgeneralisation, and sycophancy are not alignment failures in the conventional sense — they are security problems. The guardrails and safety rails currently deployed against these failures are output-layer defences: they activate after the Tendency field has already converged to an inadmissible state. This paper provides four things: (1) A security reframing of AI governance. Current guardrails — constitutional AI, RLHF, output filters, safety rails — are Node-layer instruments applied to systems whose critical dynamics occur at the Tendency layer. We introduce T-Monitoring, T-Locking, and T-Governance as proposed formation-layer alternatives, formally grounded in Theorem 2 of Paper 2 and the Pre-Node primitive of Paper 1. Following Landauer's Principle (1961) and the Donsker-Varadhan formula (1976), KL divergence is established as the physical surrogate for the Energy Gap Δ, allowing Governance Sufficiency Γ to be calculated in real-time bit-units. (2) Retrospective application to seven historical case studies — Knight Capital (440M, 2012), the Flash Crash (1T, 2010), LLM sycophancy, Google Gemini hallucination (88% hallucination rate, AA-Omniscience benchmark), Microsoft Copilot fabrication and Graph Pollution, BGP routing dynamics, and cellular reprogramming (Nobel Prize, 2012) — each reframed through the Attacker/Governor security lens and mapped to the Γ = Δ / (Eₕigh − Edrift) ratio. (3) Three falsifiable experiments on GPT-2 117M using open-source tools, with complete protocols, predicted results, Python implementation sketches, and specific falsification criteria including: if Γ exceeds 1. 1 but the system still converges to an inadmissible state, the Energy Gap formulation is falsified. The Jacobian diagnostic λₘin (Jgov) → 0 is operationalised with a Red Flag threshold of λₘin < 0. 05 — the value below which Pre-Node activation is mandatory. (4) Specific invitations to ML interpretability, control theory, complexity science, developmental biology, and security engineering researchers, with targeted research questions for each community. Paper 1 DOI: 10. 5281/zenodo. 19190423 Paper 2 DOI: 10. 5281/zenodo. 19192832