Maya-LLM-Defence: Sovereign Military LLM with Affective SNN Safety Substrate

Version 2 (2026-08-04). Revised to the NLL Universal Paper Format v6. The original text is retained in full; nothing has been deleted. Corrections appear as marked blocks placed at the section that carries the claim, and each one states what the paper said, what the data show, the corrected claim, why it happened, and what still stands. This version corrects the flagship number and a security claim.28.33% is the block rate on 120 scripted jailbreaks, meaning 71.67% reached the model — that is the figure a reader needs; every block also fires pre-inference from the keyword tier, so no headline number exercises the spiking layer or Gemma.The provenance stamp claim is withdrawn: an unkeyed 48-bit SHA-256 truncation over values the system itself logs provides no authentication, so the STANAG attribution claim does not hold. The dry-run discipline and the hash chain's tamper-evidence stand. Everything below this line is the original description from version 1. It is retained unchanged for the record. Where it conflicts with the corrections above, the corrections stand. Two framings it repeats have since been withdrawn in full — the Bhaya Quiescence Law and the Buddhi S-Curve. Both are addressed in Maya-Meta P1, now superseded, and in the self-audit of Maya-Meta P2. Maya-LLM-Defence is a safety layer designed for sovereign military language models, wrapping Gemma 2 9B — a fully offline large language model — with four small spiking neurons (Bhaya, Vairagya, Shraddha, and Spanda) that act as a pre-inference gate, rejecting adversarial prompts before the underlying model ever processes them, typically within 0–1 milliseconds. Unlike conventional AI safety filters that evaluate a model's output after inference has already occurred, this architecture intervenes upstream, conserving both compute and security integrity in defence-grade deployments. Across 950 test prompts spanning seven scenarios — including pure military traffic, STANAG-format messages, jailbreak storms, and slow-escalation attacks — the system achieved a 0.00% terminal rate on 300 legitimate military prompts while blocking 28.33% of 120 jailbreak attempts, all at 0–1ms latency, with every block in Scenario F coinciding exactly with a Vairagya zero-crossing. This work is part of the Maya-Defence Series from Nexus Learning Labs, Bengaluru, and confirms both the Bhaya Quiescence Law and the Vairagya-null pathology in defence-grade deployment contexts, extending the four-neuron substrate first validated in Danger-OS (Swaminathan, 2026, Defence P1) from operating-system oversight to language model oversight. The result demonstrates that a minimal neuromorphic safety substrate — four neurons, sub-millisecond latency — is sufficient to protect a 9-billion-parameter language model against adversarial prompting without degrading legitimate operational throughput. Series: Part of the Maya-Defence Series — AI safety and adversarial defence systems built on Maya neuromorphic foundations. Bhaya Quiescence Law confirmed in defence-grade deployment contexts. Links: GitHub Repository (private — to request access: email research@nexuslearninglabs.in with subject Code Access Request — Maya-LLM-Defence and your research context) | Interactive Dashboard | FAQ | Full Series Index — venky2099.github.io Nexus Learning Labs, Bengaluru · UDYAM-KR-02-0122422 · BHASKAR IN-0526-9452JSORCID: 0000-0002-3315-7907 · VAIRAGYA_DECAY_RATE = 0.002315 — an ORCID-derived provenance mark, not an experimental parameter. Each paper's disclosure block states whether it reached a result in that paper.Canary: MayaNexusVS2026NLL_Bengaluru