This repository now tracks only the dissertation experiment path that matters for the final analysis:

• the 200k MCTS dissertation matrix
• the telemetry full matrix
• the Isolation Forest sanitize-gated ablation
• derived report and visualization data for those result families

All retained experiments use:

• environments: HalfCheetah-v4, Walker2d-v4, Hopper-v4
• poison types: reward_poisoning, action_perturbation, observation_corruption
• schedules: random_sparse, bursty
• seeds: 0-4
• horizon: 200000 environment steps

The MCTS matrices compare clean, attack_none, and attack_defended. The Isolation Forest comparison is a sanitize-gated no-lookahead ablation, not a passive detector-only baseline.

The machine-readable source of truth is experiments/canonical_experiments.json.

• results/dissertation/mcts_poison_runs_200k/
• results/dissertation/telemetry_runs/
• results/dissertation/iforest_parameterized_runs/window200_iforest_baseline_iforest_steps200000_window200_thresholddefault_20260507T214909/
• results/dissertation/visual_data_200k/
• results/dissertation/window200_artifacts/
• results/dissertation/iforest_window200_artifacts/

• src/adsl/: SAC training, corruption, detection, MCTS control, telemetry, and Isolation Forest pipelines
• scripts/run_dissertation_campaign.py: 200k MCTS matrix runner
• scripts/run_parameterized_campaign.py: parameterized MCTS or Isolation Forest campaign runner
• scripts/run_window200_worker_pool.py: coordinated window-200 MCTS worker pool
• scripts/run_iforest_detector_baseline.py: Isolation Forest sanitize-gated ablation runner
• scripts/analyze_dissertation_results.py: MCTS report builder
• scripts/export_visual_datasets.py: visual dataset exporter
• scripts/build_window200_report.py: window-200 MCTS report builder
• scripts/build_iforest_window200_report.py: Isolation Forest comparison report builder

• RUNBOOK.md
• docs/REPRODUCIBILITY.md
• docs/DATA_DICTIONARY.md
• docs/METHODOLOGY_DRAFT.md
• docs/README_WINDOW200.md
• docs/ARCHITECTURE.md
• docs/DISSERTATION_READY.md