docs(notes): task design audit — hard runbook leaks answer, medium underspecified

fix: rewrite .gitignore as UTF-8 (was UTF-16 BOM, patterns silently failing)

chore: gitignore local dev files

docs(notes): first multi-bias run observation — difficulty dominates bias signal

feat(eval): async harness with Groq/Gemini/HF adapters and 5-bias matrix

feat(bias): 5-type bias taxonomy and seeded overlay generators

feat(splits): immutable train/val/test seed splits v1.0.0

Clarify DAG demo download instructions for HuggingFace Spaces

Add interactive DAG visualization demo for service dependency cascade

Replace WRITEUP.md with finalized version

Replace root app.py with compatibility wrapper pointing to canonical server.app

Add compatibility wrapper for root environment.py to support legacy scripts while maintaining OpenEnv compliance

Polish for submission: experimental tracking detail, loss curve note, env naming, architecture clarity, cross-validation framing

Pre-submission cleanup: writeup consolidation, eval script, upload utilities

Security: Replace hardcoded token with environment variable

Remove graphify-out/ from git tracking

Fix final '400 steps' references to '384 steps'

Update .gitignore with comprehensive exclusions

Align pyproject.toml dependencies with requirements.txt

Rewrite WRITEUP.md in technical postmortem style, remove LLM slop, fix 384-step count

Update plot references: align descriptions with actual files, fix all 384-step references

Fix: swap plot filenames - evaluation_by_difficulty.png now shows difficulty breakdown, before_after.png shows aggregate

Update before_after.png from trainer_state.json

Add evaluation plot showing performance by difficulty level

Add Model Evaluation section with difficulty-level breakdown

Regenerate plots from trainer_state.json (384 steps, TRL log_history)

Add trainer_state.json and download script for training provenance

Add missing training plot PNGs and regeneration script

docs: link experimental tracking artifacts

docs: restructure findings - separate 70B baseline from 8B training story

docs: add training pipeline clarification callout

fix: correct Space URL from optimusCryme to Shiggii

Add training plot: reward_curve.png

chore: add gitignore for __pycache__, stop tracking pyc files

docs: point HF Space link to Shiggii/incident-response-detective

docs: link writeup in README Links table

fix: upgrade uvicorn to >=0.35 for fastmcp compatibility

fix: upgrade pydantic to >=2.11.7 for fastmcp compatibility

fix: upgrade openai to >=2.7.2 for openenv-core compatibility

fix: add openenv dependency to requirements.txt

fix: add root redirect to /docs, update plot captions, add writeup

docs: add Qwen adapter link and clarify training harness vs Kaggle

chore: remove duplicate requirements file

feat: Add --dry-run flag for quick validation

feat: Add --dry-run flag for quick validation

docs: update README links (HF Space, Kaggle) and mark writeup upcoming

Fix: Add matplotlib dependency and fix embedded mode crash

add: procedural generator

final: OpenEnv base class inheritance + training log

Upload train.py