Complete curriculum list: Add multi_cascade and adversarial tiers to examples

Final link fixes for Judges' Guide

Final final final polish: Add Mermaid architecture diagram

Final blog post rewrite: Professional student perspective

Final blog post polish: Fix kubectl reference and enhance storytelling

Final README polish for submission requirements

Fix critical ModuleNotFoundError by removing absolute cloud_sre_v2 imports

Bulletproof Colab Notebook: Catch all exceptions, add rm -rf CloudSRE, and pass --no-wandb to prevent infinite hanging

Fix framework details: custom Unsloth GRPO, not vLLM

Fix Colab notebook structure to be perfectly clean for judges

Final structural cleanup, Colab refactor, and massive Blog Post overhaul

Remove private WandB links from Deliverables table

Add advanced training metrics (Leg 2) plots and resolution rates

Update reward_curve.png with high-resolution, fully annotated plot

Add both WandB project links to deliverables

Track PNG files with Git LFS for HuggingFace

Final submission: add reward curve, Colab notebook, Adversarial Designer integration, updated docs

feat: Mount interactive Gradio dashboard to root URL

docs: highlight 72B dual-LLM architecture

add: BLOG_POST.md — narrative walkthrough for judges

fix: pin torch<2.11 and torchvision<0.26 for unsloth compat

fix: install torch+torchvision in Cell 2 for HF Space compatibility

fix: auto-detect platform (Colab/HF/local) — no more /content crashes

fix: orchestrator logging and fallocate bugs

feat: narrative training output — alerts, phases, per-step rewards

feat: show best rollout command trace in training output

fix: install unsloth_zoo explicitly (was missing with --no-deps)

fix: add COLAB FRIEND / HF FRIEND labels to every cell

fix: total_mem → total_memory + clear friend assignment box

feat: bulletproof Colab training notebook — copy-paste relay format

feat: narrative README — Acts 1-4 storytelling, 25/25 OS-level faults, Docker Compose, rich transcripts

feat: docker-compose + eliminate ALL HTTP flags

feat: WandB integration + rich inference transcripts

chore: remove binary file blocking HF push

feat: Hybrid OS-level + HTTP fault injection (70B Killer v2)

docs: Cursor context file + GRPO training evidence (100% warmup, 93% single_fault)

docs: update README to reflect 16 real microservices

feat: 16 real services — 10 new microservices (search, gateway, scheduler, storage, metrics, email, billing, config, dns, loadbalancer) all running as real OS processes

docs: major README upgrade — reward tables, API docs, training evidence, real vs simulated comparison

feat: GRPO training + evaluation scripts — curriculum across 5 tiers with group-relative advantages

fix: suppress noisy warnings in REINFORCE output — clean Colab logs for judges

docs: add JUDGES_START_HERE.md — 60-second judge fast-path

results: REINFORCE 50ep — 86% resolution, reward -1.14→+0.77 avg +0.46

fix: recompute log probs with forward pass for gradient flow

results: training evidence

feat: REINFORCE training with actual gradient updates

fix: skip LoRA init if SFT checkpoint already has adapters

feat: Colab notebook, blog, README deliverables, auto-plot training curves

feat: balanced 100 SFT demos + latency injection + reward normalization

cleanup: remove test scripts