FIJISHI EPL

Sarvam AI open-sourced Sarvam 30B and Sarvam 105B foundation models trained on IndiaAI Mission compute in February 2026 under Apache 2.0 licence. Sarvam 105B wins approximately 90 percent of Indian-language benchmark comparisons against GPT-4. ArXiv research paper Are LLMs Court-Ready evaluated frontier LLMs on Supreme Court Advocate-on-Record exam and found one LLM Drafting response marked not evaluable due to pervasive formal defects. Supreme Court of India Draft AI Regulations published June 3 2026 require AI outputs to be traceable non-opaque and constitutionally accountable to the deploying officer. LLMs operate by adjusting probabilistic weights across billions of parameters - no practical mechanism exists to trace a specific output to specific inputs without complete retraining. MeitY November 2025 AI Governance Guidelines acknowledge machine unlearning technical dilemma. Principal Scientific Adviser January 2026 white paper advocates watermarking and bias detection as techno-legal controls. No IndiaAI Mission pillar specifies the constitutional command architecture above the LLM inference layer that produces deterministic constitutional audit trails meeting the Supreme Court's June 2026 traceability standard.

Read paper: https://doi.org/10.5281/zenodo.21217288
For full paper and sovereign briefings: protocol@fijishi.com