Pre-format SFT dataset as text column, drop formatting_func

Fix pickling error in SFT formatting_func closure

Fix SFT formatting_func to return list of strings

Fix SFT: set completion_only_loss=False for formatting_func compat

Fix SFT warm start: add formatting_func for Unsloth SFTTrainer

Cap prompt generation at 512 tokens and add version print

Add SFT warm start before GRPO and DB connectivity init check

Add local model inference backend for Layer 2

Increase max completion length from 512 to 2048

Make Supabase uploads incremental — upload after every step

Add Supabase upload for training results (Storage + DB)

Add raw training summary output and adjust training scale

Improve reward function to break refuse-everything local minimum and scale training

Add volume verification, fsync, and stdout fallback for training outputs

Clean up dead code, unused imports, and move hardcoded values to config.yaml

Add --llm-agent and other legacy CLI flags for backwards compatibility

Centralize all training params in config.yaml (single source of truth)

Remove mock mode: only real GRPO RL training remains

Add clear training progress logging with technical + domain names

Update docstrings to reflect LLM-only training pipeline

Align GRPOConfig defaults with CLI: 10 steps, 7 episodes

Remove all rule-based fallback systems, require LLM inference

Fix hardcoded values in report: dynamic customer count and eval episode labels

Reduce training defaults for fast iteration: steps=10, episodes=7

Add training report & logging system with reward charts and conversation comparisons

Wire up real LLM integration via HF Inference API

Fix critical gaps: prompt-sensitive agent, adversarial customers, executable GRPO, OpenEnv wrapper

Implement self-improving AI oversight system with nested RL environments