Finalize checkpoint-first evaluation and sandbox submission workflow.

Improve submission docs and training run transparency.

Improve checkpoint policy quality and environment robustness.

Fix checkpoint evaluation reliability and keep heuristic as strict fallback.

audit pass: hackathon-ready polish and truthful reporting

strengthen strict JSON reward shaping for GRPO

harden reward callback and add training diagnostics

tune GRPO batch and accumulation settings

harden GRPO reward shaping and stabilize updates

refactor: ignore private documentation and parameterize training batch size in train.py

Polish judge-facing docs and add checkpoint evaluation.

Harden training signal and clean generated artifacts.

Fix Unsloth GRPOTrainer keyword compatibility.

Fix TRL GRPO compatibility and Unsloth import ordering.

Harden hackathon readiness and surface reproducible evidence.

revert train.py to Colab-compatible baseline

harden training for shared DGX GPUs

make DGX training resilient without bitsandbytes

disable external report logging by default in GRPO training

auto-select fp16 on non-Ampere GPUs for GRPO training

fix GRPO config/trainer compatibility across TRL versions

enforce unsloth-first colab training and preflight checks

stabilize colab training deps with robust fallback path

fix colab GRPO dependency setup and error guidance

fix colab training import by adding mergekit dependency

updated EICC v2 environment, APIs, and training pipeline