Reference

Filing an experiment

Material	Location	Contents
Approach page	`approaches/<name>/README.mdx`	Question, mechanism, plan, decision rule for promotion, evidence status
Study	`approaches/<name>/studies/<slug>.toml`	`id`, `title`, `created`, `updated`, `status` (proposed, active, blocked, complete), `question`, `outcome`, `next_step`, `attribution`, `reports`, `logs`, `experiments`
Registration	`experiments/<uuid>.toml`	`id`, `study`, `status` (registered, running, complete, abandoned), `hypothesis`, `decision_rule`, `red` and `blue` players as registry name plus JSON parameters, `rules`, `games`, `seeds`, `budget`
Record	`records/runs/<run-id>.md`	Registration id, command, revisions, aggregate table, endings, timing, output digest
Report	`reports/<slug>.mdx`	The interpretation, with charts and replays from `content/assets/`
Daily note	`log/YYYY-MM-DD.mdx`	What changed, run IDs, failures, attribution

A registration is frozen when the first game starts. Its status may move to running, complete or abandoned; nothing else changes. A different design is a new file.

Registration example

id = "6d1a1f0e-1c1b-4d29-9e6e-4a3c7e2d9b10"
study = "expectimax/beats-greedy"
status = "registered"
hypothesis = "Depth-2 expectimax over 16 determinizations wins more paired games than greedy under the training rules."
decision_rule = "All 200 games complete. Per-seed contrast (expectimax score as Red plus as Blue, minus one); support if the 95% interval excludes zero above it."
red = { policy = "expectimax", parameters = { samples = 16, depth = 2 } }
blue = { policy = "greedy", parameters = {} }
rules = "training"
games = 200
seeds = "0-99"
budget = { max_ms_per_move = 2000, threads = 32 }

Record example

A record is a Markdown file with one table and one paragraph of provenance. The table has the policy, wins, draws, losses, effective win rate, the 95% Wilson interval, mean plies and mean milliseconds per move, then one row per ending. The paragraph gives the registration link, the exact command, the server and research commit hashes, the machine, and the SHA-256 of the archived JSON.

Assets

Charts and replays embedded in a report are JSON files in content/assets/ following the shapes in ../docs/plan.md, validated by scripts/check_content.py. The arena writes the replay shape directly; charts are produced by a small script kept in scripts/ and named in the asset's caption.