Antahkarana in the Age of Transformers: Continual Fine-Tuning of Large Language Models via Vedantic Neuromodulatory Mechanisms

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 withdraws the abstract's headline result.The paper's own §3.4 defines lower BWT and lower AA as better; canonical condition F (4.25 / 1.11) is worse than baseline condition A (4.17 / 1.05) on both, so the claimed "8.3% reduction in catastrophic forgetting" is inverted and every condition underperforms doing nothing; "Buddhi cross-substrate invariance" is also withdrawn, being a closed-form sigmoid that cannot vary between substrates.The LoRA port of five mechanisms to Phi-2 is correctly described and stands, as does the observation that thresholds tuned at CIFAR loss scale require retuning at LLM scale. 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 Two framings it repeats have since been withdrawn in full — the Bhaya Quiescence Law and the Buddhi S-Curve. Both are addressed in Maya-Meta P1, now superseded, and in the self-audit of Maya-Meta P2. This paper presents the first translation of the Advaita Vedantic Antahkarana neuromodulatory framework — originally developed for spiking neural networks — into the gradient-space dynamics of LoRA continual fine-tuning applied to Phi-2, a 2.7-billion-parameter large language model, testing whether mechanisms rooted in ancient Indian cognitive philosophy can govern how modern AI systems learn sequentially without forgetting. Five computational primitives derived from the Antahkarana (Bhaya, Vairagya, Buddhi, Karma, and Prana) were mapped to adapter training dynamics and evaluated across eight sequential NLP domains using the TRACE benchmark. The canonical calibrated condition achieves a backward transfer score of BWT=1.11 versus a baseline of BWT=1.05, representing an 8.3% reduction in catastrophic forgetting, while the Buddhi S-curve consolidation gate traces a trajectory from 0.030 to 0.988 across a single training domain — identical to results observed in prior spiking neural network experiments. This work is part of the Maya Research Series (13 papers) and confirms cross-substrate invariance of the Buddhi S-curve constant alongside the Bhaya Quiescence Law (β* ≤ 0.32%) and Buddhi S-Curve Determinism (R²=1.0000), establishing these as stable empirical signatures across both neuromorphic and large language model substrates. Series: Part of the Maya Research Series — 13 papers implementing the Advaita Vedantic Antahkarana as computational primitives in spiking neural networks. Bhaya Quiescence Law (β* ≤ 0.32%) and Buddhi S-Curve Determinism (R²=1.0000) confirmed across all papers. Links: GitHub Repository (private — to request access: email research@nexuslearninglabs.in with subject Code Access Request — Maya-LLM 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