AICourt strategy corpus

AICourt provides a shared episode format and multi-agent game harnesses for studying prompted strategic dispositions. It includes:

• an engine-owned commitment ledger and gzip JSONL shard writer;
• adapters for AIpha CentaurAI and They Sing traces;
• Court, a deterministic 5–7 player intrigue environment;
• a Sol/Terra/Luna model-agent campaign harness;
• schema, coverage, token, and balance validation; and
• a browser-based Court replay desk with an agent diary and event animation.

Quick start

Requirements: Node.js 20+ for the corpus tools, Node.js 22+ for the replay desk, and Python 3 with tiktoken for exact token reports.

npm test
node scripts/generate_court.mjs --episodes 500 --out datasets/court
python validate_corpus.py datasets/court --out results/court-corpus-validation.json
npm run balance:court

Launch the replay desk:

npm run viewer:sample
npm run viewer:dev -- -p 4173

The viewer accepts an episode JSON/JSONL record or a replay bundle containing { episode, decisions }. Structured decisions unlock the per-agent reasoning diary; plain episodes still provide the stage, chronicle, private-channel controls, and engine-resolved commitment ledger.

Evidence boundaries

BALANCE_REPORT.md reports seeded deterministic-policy simulations. Those runs exercise mechanics reproducibly but do not measure model-agent strategic quality. MODEL_AGENT_PILOT.md separately inventories the available Sol/Terra/Luna Court evidence and states where the sample is too small or predates current mechanics.

Role outcomes are not disposition rewards. The primary SFT export is unweighted behavioral cloning. Any future outcome-weighted training must stratify or residualize rewards by role, seat count, and archetype; raw victory weighting will learn role priors.

See CORPUS.md for the data contract, CAMPAIGN.md for the configurable model-agent campaign, and BALANCE_REPORT.md for current deterministic balance evidence.

License and citation

Code and repository-authored documentation are licensed under Apache-2.0; see LICENSE. Released transcript corpora must include their own data card and preserve any applicable model-provider or upstream-environment terms. Use CITATION.cff when citing the software.