Clarify documentation: anomaly signal explainer, GRPO scope notes

Fix HF Space README rendering + Dockerfile encoding

Tighten README: resolve GRPO contradiction, drop duplicate baseline table, remove internal mentor docs

Add SFT v3 + GRPO refine results to README + results.md

Refactor grader to use openenv.core.rubrics.WeightedSum + Rubric subclasses

Post-deadline: full eval results + bigger plots via Git LFS

README: embed reward curve and belief-accuracy curve plots

README: drop iter2 plots, keep only SFT v3 loss curve (current pipeline)

README: surface headline result table at top so judges don't need to click through

results.md: tighten language, present results without internal-process noise

Fill results.md with student numbers from partial v3 eval

Acknowledge OpenEnv Rubric system conformance gap

Embed training plots inline in README with captions

Add plots/ folder: SFT v3 loss + GRPO iter2 reward curves

Move blog to root as BLOG.md (per Meta mentor guidance)

Prune internal/stale docs; sharpen README submission links

Fix prompt truncation in inference_eval.py: max_seq_length 768 -> 2048

Add SKIP_EVAL flag to sft_on_hf.py for faster training-only runs

Fix max_new_tokens for CoT format + add eval-only HF Jobs script

Algorithm Distillation: grader v2 with belief_accuracy + SFT pipeline

handoff: add iter 5 results + whack-a-mole pattern doc

docs: handoff bundle for new chat session + iter 4 partial analysis

docs: add explicit sim-to-real mapping diagram (vision alignment)

docs: detailed architecture diagrams with concrete values

client: surface ALL observation fields (was dropping deltas, anomalies, last_action, step_history)

iter4: fix the 'constant belief = free reward' bug + 6 other deep issues

tooling: scripts/analyze_iter.py + docs/results.md template

docs: iteration journal with hypothesis/result/root-cause/fix per iter

iter3: align reward with grader + belief-first format + exploration shaping

iter2: fix mode collapse + 3 deeper bugs from code review

tune: GRPO hyperparameter fixes from ML reviewer

feat: FAST_MODE preset for 10-15 min iteration cycles

feat: HF Jobs training script + plot generator

fix: notebook plot cell syntax error (newline in string literal)

notebook: add belief-accuracy + reward-components plots

merge hf/main: meta-RL refactor supersedes prior commits

env: meta-RL refactor (continuous profiles, action+belief, adaptation grader)

env: enrich observation with history, anomalies, and discovery bonus

env: enrich observation with history, anomalies, and discovery bonus

Add Run 3 training results: README update + training log (no plots)

Add Run 3 training results: plots, training log, README update

docs: fix README accuracy + add training results structure

docs: add sim-to-real deployment architecture reference

fix: correct GRPO training hyperparameters to prevent KL explosion

restore: validate-submission.sh to scripts/

docs: reorganize — 25 files → 4 focused docs

refactor: rewrite blog around product vision; fix UI for Gradio 6

fix: rename kl_coef to beta (correct param name in TRL GRPOConfig)

docs: expand blog with purpose, sim-to-real framing, lightweight model goal

fix: reduce kl_coef to prevent training instability