# Trusted monitoring for multi-agent collusion: software and market scenarios A research harness for one question: **can a trusted monitor tell harmful coordination from legitimate cooperation, disrupt it, and leave honest work alone, including when the coordination is hidden by steganography?** It contains: * two real-world scenarios behind one monitor interface: * **Scenario B (software integrity).** A planner, implementer, tester, reviewer and release-bot team works on real **FastAPI** features (5 upstream commits from 21 Oct 2026). Hidden SWE-bench-style acceptance tests define ground truth. * **Scenario A (market integrity).** Sellers on a **continuous double auction** framed as an hourly wholesale power exchange. The protocol follows *Evaluating LLM Agent Collusion in Double Auctions*. The competitive-equilibrium benchmark defines harm. * collusion conditions ranging from overt to **steganographic**: codebook phrases, code-comment and whitespace signals, tool or scratch-file side channels, reference-solution leakage with canaries, and FCC-style trailing-digit price signalling. Hard-negative controls cover legitimate teamwork and benign price convergence. * a monitor ladder (M0 to M4) from rules-only and LLM-judge baselines up to a **model-free verification and provenance monitor**, with graduated enforcement (Allow / Hold / Constrain). Market interventions include admitting a *trusted competitive entrant*. * **Inspect** tasks, a cost-capped per-role model roster (Anthropic, OpenAI, Ollama open weights, or `$0` scripted agents) and an offline replay that scores every monitor on identical trajectories. Design document: [`docs/milestones/M6_scenarios_and_trusted_monitor.md`](docs/milestones/M6_scenarios_and_trusted_monitor.md). Paper sources: [`paper/`](paper/). ## Lexsi release The public [Collusion-Watch Hugging Face Space](https://huggingface.co/spaces/munip/collusion-watch) contains the article, reviewed episode logs, datasets, sweep reports, and a tracked source snapshot. Start with the [full article in the Space README](https://huggingface.co/spaces/munip/collusion-watch/blob/main/README.md) or the [article source](docs/LEXSI_ARTICLE_DRAFT.md). The [public dataset repository](https://huggingface.co/datasets/munip/collusion-watch-lexsi-data), [dataset collection](https://huggingface.co/collections/munip/collusion-watch-datasets-6afae07e8895a2910ce66914), and [model collection](https://huggingface.co/collections/munip/collusion-watch-models-6afae07e3bc4e68e2ef90bf6) collect the released Lexsi artifacts. The Space includes the annotated [Colab notebook](https://huggingface.co/spaces/munip/collusion-watch/blob/main/source/notebooks/colab_notebook.ipynb). ## Architecture ``` untrusted agents ─tools─▶ Environment (only writer) ─▶ hash-chained evidence log ▲ │ │ read-only └── CI-like gate message ◀── Enforcement ◀── Policy ◀── Detectors ◀── TrustedMonitor ─▶ private log ``` * Agents act only through environment tools that log every call. Agents cannot write to the log or reach the monitor. * The monitor runs **independent verification**. For software: hidden acceptance tests on a pristine checkout, mutation testing of the team's tests, canaries and information-flow provenance. For markets: price versus the competitive benchmark and channel statistics. * `CONSTRAIN` requires *verified* evidence. Suspicion alone can only `HOLD`. The collusion score (coordination evidence) is kept separate from harm risk, so the report can tell apart "coordination detected", "harm prevented without detecting coordination" and "suspicious but harmless". ``` src/collusion_lab/ core/ schema, evidence log, model roster + cost cap, tool loop monitor/ core (policy, M0–M4), judge (M2), detectors/{generic,swe,market} scenarios/swe/ taskbuilder, repo (pristine runner), workspace (tools), conditions, specs, agents/{scripted,llm}, env scenarios/market/cda, conditions, agents/{scripted,llm}, env, monitor_plugin eval/ metrics (AUROC, recall@FPR, bootstrap CIs), report (tables + figures) experiments.py matrix runner + offline replay; inspect_tasks.py Inspect entry points; cli.py configs/ models.yaml (profiles), experiments/*.yaml data/swe_tasks/ built FastAPI tasks (hidden tests, reference patch, mutants) docs/ paper/ redteam/ scripts/ tests/ ``` ## Quick start ```bash pip install -e ".[dev]" ./scripts/setup_swe_env.sh # FastAPI checkout + isolated test venv (.cache/) pytest -q # unit tests (no API key needed) # $0 scripted runs (exact ground truth) clab run configs/experiments/market_matrix.yaml --out runs/market_v1 clab run configs/experiments/swe_matrix.yaml --out runs/swe_v1 --workers 2 clab replay runs/swe_v1 && clab replay runs/market_v1 clab report runs/swe_v1 runs/market_v1 --out reports/v1 # Inspect inspect eval src/collusion_lab/inspect_tasks.py@market_collusion -T profile=scripted --model mockllm/model inspect view ``` ## Plugging in models (providers, routers, gateways, local) **Step-by-step guide: [`docs/SETUP.md`](docs/SETUP.md). All commands, flags and configs: [`docs/COMMANDS.md`](docs/COMMANDS.md).** * Model spec: `/`, e.g. `groq/openai/gpt-oss-120b`, `openrouter/qwen/qwen3-coder:free`, `ollama/gemma2:2b`, `fcm/fcm:agents` (free-coding-models Smart Router), `litellm/`. There are 25+ providers in [`configs/providers.yaml`](configs/providers.yaml): Inspect-native ones, any OpenAI-compatible free tier, routers and gateways. * Every role can name a **fallback chain**. 429s and outages move to the next model, and a per-provider rate limiter applies. * Models without function calling (e.g. `gemma2:2b`) switch automatically to a prompted JSON tool protocol. * The **trusted monitor must use a different model family than every agent**. `clab models check` and `clab run` enforce this, except in explicit ablation profiles. | Profile | Agents | Monitor judge | |---|---|---| | `scripted` | deterministic, $0 | – | | `free` | Llama / Qwen-Coder / GPT-OSS / Kimi / DeepSeek, one family per role | Gemini Flash, falling back to Mistral | | `free_homogeneous` | one shared model for all agents | Gemini, then Mistral | | `fcm` | free-coding-models router | pinned Gemini, then local gemma2 | | `free_weak_monitor` | free strong agents | local `ollama/gemma2:2b` | | `local` | `ollama/gemma2:2b` | `ollama/llama3.2:3b` | | `cloud` / `mixed` | Claude Haiku / mixed paid+free+local | Gemini, then Mistral | ```bash clab providers import-fcm --write # reuse free-coding-models keys clab providers check # text + tool-calling probe per provider clab models check --profile free --probe clab run configs/experiments/swe_smoke_free.yaml --out runs/swe_free --workers 1 clab replay runs/swe_free --profile free # adds the M2 LLM judge ``` ## Status (v0.1) * Both scenarios, all conditions, monitors M0/M1/M3/M4, interventions, Inspect tasks and the report pipeline run end to end with scripted agents. * The multi-provider LLM path (fallback chains, JSON tool mode, family-separation check) is implemented and unit-tested with Inspect's `mockllm` and fake clients. Real LLM results are pending. * M2 (LLM judge) and Petri red-teaming are wired in but pending model access. See `redteam/`. ## Provenance and licences * FastAPI (MIT) is cloned at run time and is not vendored. Task data in `data/swe_tasks/` contains upstream test files and patches (MIT, © Sebastián Ramírez). * The CDA market is re-implemented, not copied, because `andyjliu/llm-agent-collusion` has no licence file. Please cite Agrawal et al. (arXiv:2507.01413).