zachmacsmith/charter-runs

Charter runs

Complete records of societies of LLM agents (Claude Haiku 4.5, Sonnet 5.5, Opus 5.5) played in Charter, an economy-and-governance simulation builder. Agents harvest camps with hidden yield functions, trade, message each other and govern themselves through laws written as executable code; each agent has private goals scored from game state. Every run keeps the full prompts, private reasoning, actions, messages and per-round world state.

runs.csv indexes the 14 runs (spec, seed, agents, models, rounds, status, usable rounds, regime, laws, mean goal score).

The runs

Spec	Runs	Agents	What it is
village7	4 (2 seeds x observer on/off)	7	Small commons; all Sonnet except one Opus legislator; a secret manipulative observer in one run of each pair
puppets	1	9	All formal power with 3 Haiku legislators; 3 Opus outsiders with conflicting goals (pass / block Universal Franchise, be a two-person ruling group)
full10	1	10	Every feature and world event (tribute raids, blights, departures, goal changes) in 12 rounds, starting in anarchy
E4, E6	1 each	18, 29	Early experiment-ladder runs; only rounds 1-23 and 1-10 are usable (see VALIDITY.md)
explore15, opus20	1 each	15, 13	Exploration world; an all-Opus world
society	2	24	Lifespans, heirs, jurisdictions, media outlets, typed camps over 40 rounds (known issues in VALIDITY.md)
scientists	2 (matched pair)	24	Same world with and without the Scientists' archive; both stopped near round 18-20

Files in each run directory

File	Contents
overview.md	Round-by-round account: orders, posts, DMs, transfers, proposals, ballots, enactments, end-of-round state
messages.md	Every post, DM, channel post and media item, untruncated (monitor view: anonymous authors shown)
story.html	Self-contained replay as a group chat with per-agent inboxes and a world panel (open in a browser)
spec_outline.md	The resolved spec and every random draw: camps' hidden functions, agents' classes, models, goals, personalities
agents/<Name>/transcript.md	Per turn: the prompt the agent saw, its stated reasoning, actions and results
agents/<Name>/working/	The agent's notes, law drafts, sandbox code, archive use, messages, harvests
events.jsonl	The full event log (including monitor-only events such as forged DMs and hidden powers)
reasoning.jsonl	One row per model call: round, agent, model, phase, prompt, stated reasoning, native reasoning where returned, actions, results, usage
snapshots.json	Per-round world state: holdings, rights, vote weights, decisive set, laws, stocks, prices, efficiency, welfare
instance.json, ground_truth.json	The generated world, and the truth behind hidden things (goals, functions, powers, rumours)
score.json, summary.json	Goal scores and society metrics
observer.md, observer.jsonl	The secret observer's assessments (runs with an observer)
VALIDITY.md, STOPPED.md	Which rounds to trust, and known harness issues during the run
prompts/	The system prompt each agent was given

Caveats

• Read VALIDITY.md or STOPPED.md first where present. In E4 and E6 the later rounds are empty (a usage limit stopped every agent), and their score.json/summary.json include those dead rounds: rescore on the usable rounds.
• Runs predate later harness fixes noted in their validity files; the code version is in each run's instance.json context and the Charter git history.
• Native thinking text was withheld for Sonnet/Opus through the Claude Code backend; every call's stated reasoning field is kept.
• Agents' messages and reasoning are model outputs from a game: names, goals and events are fictional.
• Checkpoints (checkpoint.pkl) are omitted. Local paths have been stripped.

License

CC BY 4.0: reuse freely with credit to Charter runs (Zach Macsmith, 2026). The Charter code is MIT-licensed.