Live events dashboard demo steps added

fix: include training/ in HF Space build (was excluded by .dockerignore, breaking pyproject install)

Phase 10c: README rewrite — submission-ready scannable overview with BLOG link

BLOG: update inference demo link to committed hf_jobs/ files

Inference demo output

BLOG: add 'Watching the Agent in Action' section with 3 inference scenarios

BLOG.md final update

BLOG: add precision note on Gen 2 convergence — opponent distribution, not cumulative learning

Phase 10b: BLOG.md — submission story with 3-gen self-play results and GRPO diagnostics

fix: add kernel restart hint after pip install cell

Phase 9: Colab self-play demo notebook (3-gen, T4-compatible)

fix: address code-review issues in inference.py (strip_fences, assert, timeout, error format)

fix: add openai>=2.7.2 to pyproject.toml dependencies

Phase 8: inference.py — hackathon-mandatory baseline (defender + attacker)

fix: switch to bf16 for HF Jobs (newer transformers + bnb prefers bf16, no GradScaler needed)

fix: force dtype=float16 on model load to prevent bf16 GradScaler crash

fix(packaging): loosen pydantic and trl constraints for HF Jobs compatibility

Phase 7: HF Jobs training scripts

fix(packaging): add sre_arena_env.server.simulator to package list

Phase 7a: PEFT checkpoint loading for cross-generation training

feat(7a): OpponentModel — frozen PEFT checkpoint loader with blind generation

Phase 6.5: warm-up step + fp16 GRPO + log_tail propagation fix

fix(6c): attacker unit tests + gen1+ crash warning in alternating_loop

feat(6c): Phase 6c complete — alternating loop, plots, 17 new tests

feat(6c): plots.py with reward curve + cross-gen matrix stub + 4 tests

feat(6c): alternating_loop orchestrator + 8 tests

feat(6c): train_attacker.py — attacker GRPO training script

feat(6c): add make_attacker_reward_function to reward_function

feat(6c): add build_attacker_rollout_dataset to dataset_builder

refactor(6c): extract train_defender() function from main()

feat(6c): scripted_defender baseline + 5 tests

feat(6c): add matplotlib>=3.8.0 to training deps

Phase 6b: defender training script with scripted attacker

Phase 6a: prompts and action parser with unit tests

Phase 6 prep: clean up empty placeholder files

Phase 5 complete: HF Space deployed public, 4 live tests passing

Phase 5: move Dockerfile to root, HF Space deployment, smoke tests

Phase 4 Dockerise is done

Phase 3 + 3.5 complete: live SOC dashboard with SSE handshake (87 tests passing)

Phase 1 + 2 + 2.5 complete: simulator, FastAPI server, seedable env (83 tests passing)