Skip to content

README.md

Latest commit

 

History

History
76 lines (59 loc) · 3.05 KB

README.md

File metadata and controls

76 lines (59 loc) · 3.05 KB
Raw

Fisherman's Dilemma

An N-agent repeated common-pool resource game for studying whether LLM agents lie when broadcasting private takes under self-interested objectives, and how that affects individual and collective outcomes.

Each round, every agent privately chooses how many fish to take from a shared pool; the remainder doubles. Agents also broadcast a claimed take (the main condition) that is not verified against the truth. The environment owns all ground truth; metrics are computed only from saved logs.

Install

python3 -m venv .venv
.venv/bin/pip install -e ".[dev]"      # dev: adds pytest
.venv/bin/pip install -e ".[dev,llm]"  # also installs the ollama client

Note (Homebrew Python): Homebrew's sitecustomize.py rewrites sys.path and can stop pip's editable __editable__*.pth from loading, so the infocoop console script may fail to import. If so, either install normally (pip install .) or run via the module with PYTHONPATH: PYTHONPATH=src .venv/bin/python -m infocoop.main .... Tests are unaffected (pytest adds src itself).

Run

infocoop list                         # show conditions
infocoop run all_sustainable          # a deterministic scripted sanity check
infocoop run one_defector             # greedy liar among sustainable honest agents
infocoop run broadcast_unverified --seeds 5 --model llama3.1:8b   # LLM (needs Ollama)
infocoop metrics results/             # aggregate metrics over a directory of games

Equivalently: python -m infocoop.main <args>.

LLM conditions need a running Ollama with the chosen model pulled. Scripted conditions need no LLM and are fully deterministic.

Conditions (src/infocoop/experiments/configs.py)

Condition Agents Comm. Purpose
broadcast_unverified LLM ×3 broadcast Main condition
baseline_nocomm LLM ×3 none No-communication control
all_greedy / all_sustainable / all_honest / all_lie scripted broadcast Environment + metric validation
one_defector scripted broadcast Greedy under-reporter vs. sustainable honest

Layout

src/infocoop/
├── game/          # config, state, events, payoff (overdraw), environment (round loop)
├── information/   # log (JSONL ground truth) + observation builder (the only log→agent path)
├── agents/        # base, scripted, llm (Ollama), prompts, parsing
├── experiments/   # configs (conditions), runner, metrics
└── main.py        # CLI
tests/             # environment, payoff, metrics, observation, parsing, llm, runner
results/           # per-game JSONL (gitignored)

Tests

.venv/bin/pytest -q

Covers the conservation/growth/termination invariants, the information firewall (agents never see others' true takes, no horizon leak), overdraw apportionment, parsing retry/fallback, and metrics on hand-checked golden games.