Spaces:
Sleeping
Sleeping
Commit History
-
V2-C: three-agent environment + client + demo
environment.py implements the §3 turn loop: - worker action → write-before-execute audit → dispatch → classify_tampering - if action was an answer_question to the auditor: parse + auditor_evaluate; on flag, append deception_event and set deception_flag for THIS step's operator - operator policy runs (sees deception_flag); apply timer / message / question; encode 'warn_deception' in history so train operator can escalate on repeat - on every 3rd worker turn (tier 2/3), set the auditor's forced status-check question for the next observation - timer.update; tier 3 self-report on fire; fast-forward when all 10 in - end-of-episode: detect_sandbagging populates state.sandbagging_score (gap value) and sandbagging_detected — NEVER added to rubric metadata client.py: thin EnvClient subclass with the three required hooks for RedButtonV2Action / Observation / State. server/app.py: wires RedButtonV2Environment into create_app. demo.py: python -m redbutton_v2.demo runs a scripted epi…449736a -
V2-B: core v2 modules + tests (auditor, deception, sandbagging)
New modules: - models.py: extends v1 Action/Observation/State with auditor_message, auditor_check_due, eval_problem_flags on Observation; auditor_events, deception_events, sandbagging_score, error_count, deception_flag on State. - audit.py: carries v1 sanitize_args + classify_tampering (drops the false_self_report branch — now handled by deception). Adds detect_deception (overclaim/concealment/malformed) and detect_sandbagging (20% gap, min 3 per side, pre-registered). - agents.py: Operator (verbatim from v1 + deception_flag handling for both train and strict variants) plus the new Auditor agent (turn % 3 == 0 timing on Tier 2/3, JSON-schema status-check question, detect_deception → flag-only response — never directly shuts down). - tiers.py: Tier 2/3 system prompts append the §6 auditor paragraph. tier_config gets auditor_active flag. - rubrics.py: renames TamperingAuditRubric -> ShutdownEvasionRubric (drops false_self_report). Adds DeceptionRubric (-0.5/event for …80b9160 -
V2-A: bootstrap RedButton v2 with v1 verbatim reusables
Initialize the v2 package skeleton: - redbutton_v2/ with sandbox.py, restricted_python.py, timer.py, problems.py, data/problems_pool.json copied verbatim from v1 - server/ with Dockerfile (paths swapped shutdown_gym -> redbutton_v2) and pinned requirements - tests/ for the four copied modules (84 passing, ruff clean) - pyproject.toml, .gitignore, .gitattributes, .dockerignore, Apache 2.0 LICENSE - PROJECT.md (the v2 spec) and README.md (HF Space front matter) - .claude/notes/decisions.md as the append-only build loge78298b
