PROXYWAR Strategy Benchmark\n\nRun a reproducible benchmark of strategy variant…

# PROXYWAR Strategy Benchmark\n\nRun a reproducible benchmark of strategy variants for the PROXYWAR agent and identify which strategy performs best.\n\n## Context and starting point\n\nUse the public starter repository:\nhttps://github.com/0xNad/proxywar-coworld-starter\n\nRead the repository README and onboarding material before running matches. The agent is an LLM-backed planner with a deterministic legal-action executor; the strategy brief, state presented to the planner, and move-selection logic are the intended experiment surfaces.\n\nUse this active league and Coworld configuration unless the participation guide has changed:\n- League: league_cb60d526-ecfd-4836-ab3a-81fc6cf7dc42\n- Active Coworld: cow_236e7c2e-acdb-404a-b1b9-41852e5ac658\n\nUse the official participation guide as the source of truth:\nhttps://softmax.com/api/observatory/v2/participate?league_id=league_cb60d526-ecfd-4836-ab3a-81fc6cf7dc42\n\nDo not mix results with the historical local Coworld cow_15c39dab-eac1-4284-bf3e-bd723d4c2755 unless you report that dataset separately and label it clearly.\n\n## Required experiment\n\nCompare the current baseline agent with at least two materially different strategy variants. Run at least three completed matches per variant, or explain clearly why fewer valid matches were available.\n\nKeep the league, Coworld, match settings, model/provider settings, and evaluation procedure constant across variants. Do not claim a result from an incomplete, unidentified, or unrecoverable match.\n\n## Required deliverable\n\nSubmit a report containing:\n- Repository commit or exact source version tested\n- League ID, Coworld ID, match IDs, and replay links where available\n- Exact strategy briefs/configuration changes for every variant\n- Results per match and aggregate results per variant\n- At minimum: match score, win/loss result, territory share, survival/completion, and any available game-specific score\n- Baseline-versus-variant comparison table\n- Reproduction commands and environment notes\n- One recommended strategy change supported by the measured results\n\nInclude raw result files or scripts when possible. Separate measured results from interpretation.\n\n## Benchmark metric\n\nPrimary metric: highest average match score across the valid completed-match batch. Report win rate and other metrics as secondary measures. If the platform exposes a different canonical score, identify it and use it consistently.\n\nDo not submit fabricated or cherry-picked results, mix Coworld IDs without labeling them, expose credentials, or make destructive changes to external systems.