add ablation experiments: truncation baseline + tag removal analysis

readme: full rewrite, human voice, no AI patterns, no dashes

blog: remove all dashes from prose text

blog: remove all --- dividers, fix hero graph URL for LFS

add: before/after thinking comparison graph — untrained vs trained

blog: address reviewer feedback — truncation baseline, ablation gap

blog: casual tone pass, less polished, more hackathon diary

rewrite blog as story format — accessible, human, narrative arc

fix: use HuggingFace resolve URLs for LFS-tracked images

fix: use raw GitHub URLs for blog images so they render on HF

docs: finalize blog post with hero graph and training results

add: real training curves from 200-step GRPO run + improvement panel

fix: broken file path references in JUDGES.md and ENV.md

fix: handle Gradio SSR zero-input calls in run_demo

clean: submission-ready repo

fix: SFT max_seq_length→max_length + 200 episodes (not 150)

tune: LR 5e-6, warmup 0.05, KL 0.04 + EMA smoothing + trend line

train: 200 episodes × 1 epoch = 200 steps (~6.6hrs, done by 10:30 AM)

perf: cut training to 100ep×1epoch (100 steps, ~3.3hrs) to finish before deadline

fix: Red Team tab SSR crash + incremental training curves + storytelling polish

v2.3: right-size training budget + add red-team safeguards + tighter pitch

v2.2: switch to Qwen3-1.7B + unlock full lengths + real calibration logging

hotfix: revert seq/completion length to v1 — Space hit 14GiB OOM

v2: calibrated metacognition as RL + inference-time budget + transfer eval

docs+infra: add submission deliverables (openenv.yaml, colab, eval, README, blog)

fix: install forward pre-hook on lm_head to auto-cast input dtype

fix: cast lm_head and all params to bfloat16 to resolve dtype mismatch

Fix: TRL warnings_issued compatibility bug

STABILITY PUSH: Reduced memory (2048/1024), added Auto-Refresh Timer, and persistent Log File.

FINAL FIX: correct GRPOTrainer API (args/processing_class) + auto-start training on boot

Fix: use env.reset() instead of non-existent _get_observation()

Fix: add 4-bit quantization + LoRA to fallback path (prevents OOM)

Fix: copy data files into code_review_env package

CRITICAL FIX: remove fake fallback eval, exit 1 on crash, proper error reporting

Fix: training crash detection + retry support

Fix Gradio SDK version

🔐 CodeReviewEnv v3: Agentic Security Investigation with Thinking Budget

initial commit