simcraft-experiment

Files (1)

• SKILL.md

SKILL.md (first lines)

# simcraft: experiments with the designer

**Roles.** The designer asks the question and judges the answer; you design the runs, execute them, and report evidence. Say what you will vary and how many runs before running.

## Where the truth is

- Panel (`engine.toml`): `[run]` seed and length, `[params]` overrides, `[switches]` rules on/off, `[agent]` seats. See `docs/architecture.md` (https://github.com/cemphlvn/simcraft/blob/main/docs/architecture.md).
- Scripted players: `agents/market.py`, `agents/mercy_dungeon.py`. Example of a sweep and how results are written up: `docs/emergence.md` (game 3 strategy table, game 4 culture sweep).

## Method

1. **Question → variables.** One question per experiment ("does crunch help?"). Pick the params/switches to vary and what to measure (scores, event counts from `step`, props from `observe`).
2. **Panels.** Write one panel per setting (e.g. under a scratch folder) and run with `--config`:
   ```bash
   cargo build -q --release -p sim-agent
   ./target/release/simcraft-agent games/<name> --config /tmp/panel_a.toml
   ```
   Drive it over stdin/stdout JSON (one request per line, one reply per line). Step in large batches (`{"cmd":"step","n":240}`) and count `events` by `name`.
3. **Seeds.** At least 3 seeds per setting; report the mean and say when seeds disagree.
4. **Report** a small table (setting → outcome) and 2–3 plain sentences: what changed, what did not, what no rule states directly (the emergent part). Distinguish evidence from your interpretation.
5. **Keep it** if the designer wants: add the table to `docs/emergence.md` under the game.

Determinism means a surprising result can always be replayed: note the seed and panel.

From cemphlvn/simcraft, MIT.

More from cemphlvn/simcraft