Strategy Game Agents

Behavioral-finance experiment tooling plus clean-room strategy agents for simulating repeated risky-choice games. The repository started as a web experiment page; this version adds a reproducible Python core that can replay simple decision rules and compare agent behavior under bounded evidence.

What It Shows

• A small repeated-game engine with deterministic seeding.
• Strategy agents: fixed preference, epsilon-greedy learning, and loss-averse utility.
• Metrics for cumulative payoff, risky-choice rate, regret against the best available option, and round-level traces.
• Participant-trace validation plus interpretable baseline fitting.
• A competition evidence ladder that keeps candidate claims below the strongest verified trace, replay, local-screen, or official-run evidence.
• Unit tests that pin reproducibility, regret accounting, behavioral parameters, fitting, and evidence boundaries.
• A clear boundary between experiment infrastructure and validated behavioral or trading claims.

Quick Start

python -m pip install -e .
python -m strategy_game_agents.demo
python -m unittest discover -s tests -v

Minimal API

from strategy_game_agents.agents import EpsilonGreedyAgent
from strategy_game_agents.evidence import CompetitionEvidence, decision_boundary
from strategy_game_agents.game import RepeatedChoiceGame, fixed_option, risky_uniform_option
from strategy_game_agents.simulation import run_simulation

game = RepeatedChoiceGame(
    options=[
        fixed_option("safe", payoff=5.5),
        risky_uniform_option("risky", low=1, high=10),
    ],
    rounds=10,
    seed=42,
)
result = run_simulation(game, EpsilonGreedyAgent(epsilon=0.2, seed=7))
print(result.summary())

evidence = CompetitionEvidence(
    candidate_id="agent-v5.5",
    level="local_screened",
    games_observed=48,
    primary_bottleneck="adversarial_policy_shift",
    claim_ceiling="local screen passed, official support not established",
)
print(decision_boundary(evidence))

Evidence Boundary

See evidence/validation-boundary.md and reports/competition-evidence-ladder.md. This is a public research-tooling repository, not a behavioral theorem or trading model. The disclosure boundary is recorded in DISCLOSURE.md, and sample_data/agent_evidence_sample.csv is a synthetic schema fixture.