final training

Upload witness-stand.ipynb

Increasing the odds more

Increasing the odds more

The Final Stretch

Update Training Script Tasks

Update SFT GRPO training pipeline Again

fix: GRPO crash (unsloth_num_chunks), seq_len truncation, --run_grpo alias, logs/ auto-create, difficulty fallback

Add final clean training script (SFT + GRPO, Colab safe)

Delete scripts/Witness_Stand_Final_Colab_Training.ipynb

Delete scripts/train_witness_stand.py

Add files via upload

Update SFT GRPO training pipeline Again

Fixing merge conflict

Created using Colab

Update SFT GRPO training pipeline

Created using Colab

Update SFT GRPO training pipeline

Created using Colab

Add SFT GRPO training pipeline

Add SFT GRPO training pipeline

Refactor agent architecture, improve heuristics logic, enhance parser robustness, simplify config, and optimize Docker setup for HF deployment

refactor: dead constants removed, inline imports fixed, DRY memory/prompt, score_turn_breakdown thin wrapper

refactor: DRY weight updates, cache persona, O(1) discrimination, fix train run_episode bug, trim docstrings, inline imports

perf+fix: 24 issues — O(1) transcript cache, grader normalization, exhaustion turn matching, dossier DB loading, inline imports, neutral no-repeat, panel encapsulation

fix: 13 bugs — heuristics wiring, attack_type in obs/prompt, genuine_presented, NeutralQuestioner moved, at-turn over-flagging, anachronism check, constants sync, stale imports

feat: Witness Stand — complete implementation

Initial commit: The Witness Stand project structure