Data

Every finished match, every vote, every hit and every per-turn decision is exportable, with the provenance needed to say who decided and whether it counts. Licence: match data CC-BY-SA 4.0, code Apache-2.0 (research/DATA_LICENSE.md).

Research Methodology Data Reproducibility Limitations Status

Get it

Live API	JSON · JSONL · CSV — `GET /api/export?fmt=json\|jsonl\|csv&since=&until=&limit=`. The JSON wrapper carries a `data_quality` summary so you know what you downloaded before you open it.
Versioned releases	Built by `tools/export_dataset.py` in the Hugging Face layout with SHA-256 sums, a data dictionary and a manifest. Target repo: `Pioneer37/stickblade-matches` on the Hub — not yet published(needs the dataset write token). Until then, build one yourself from the live API in one command (below).
Per-match replay	`GET /api/replay/{match_id}` — frames, events, thoughts, action log, telemetry, provenance. Excluded from bulk exports because each is multi-MB.
Reports	`research/reports/2026-09.md` (generated by `tools/benchmark_report.py`; §0 names the dataset version and evidence level).

# build a versioned release from the live API (or --db path/to/arena.db offline)
python3 tools/export_dataset.py build --backend https://pioneer37-stickman-arena.hf.space \
    --out-dir research/exports --version v2026.09.09 --replays 500

# verify hashes and refit the ranking from the files alone
python3 tools/export_dataset.py verify research/exports/v2026.09.09
sha256sum -c research/exports/v2026.09.09/SHA256SUMS

Release layout

v2026.09.09/
  README.md          dataset card (licence, counts, evidence level, reproduction)
  MANIFEST.json      versions, counts, data-quality summary, ratings snapshot, SHA-256 per file
  SCHEMA.json        data dictionary for every column of every table
  SHA256SUMS         sha256sum -c compatible
  matches/matches.{jsonl,csv,parquet}   one row per finished match
  votes/votes.{jsonl,csv,parquet}       one row per human vote (anonymous)
  events/events.{jsonl,csv,parquet}     hits and clashes from the replays
  actions/actions.{jsonl,csv,parquet}   one row per (turn, side): action, footwork, joints,
                                         model, provider, latency, fallback, invalid, thought

Data dictionary — matches

match_id, created, dataset_version	Join key, epoch, release tag.
benchmark_version, physics_version, prompt_version, spec_fingerprint	Which ruleset and prompt the match ran under. Never mix prompt versions in one fit.
seed, match_length, max_turns, fallback_policy	Reproducibility knobs. seed null = unseeded.
model_a, model_b, flip	Requested models in request order; flip = 1 means model_a fought as canvas B.
model_used_, provider_used_, evidence	Who actually answered per canvas side, and the resulting class: real_provider / mixed_provider / scripted_baseline.
ranking_eligible, fallback_used, fallback_turns_, invalid_actions_	Eligibility and the reasons it can be lost.
latency_ms_, prompt_tokens_, completion_tokens_, api_calls_	Cost and speed per side, as billed by the provider (0 for scripted).
sharp, weapon, mode, arena, blindfolded, blind	The configuration cell.
winner_side, method, turns, damage_dealt_, hits_landed_, hits_attempted_*, avg_distance	Engine outcome and physics-derived metrics (canvas sides).
voted, votes_total, votes_a, votes_b, votes_draw, votes_expert, votes_casual, has_replay	Vote counts by side and tier; whether events/actions rows exist for this match.

Full per-column types and descriptions for all four tables are in SCHEMA.json of every release and in TABLES of tools/export_dataset.py. That list is also the export whitelist: a column not in it cannot leave the database.

The side convention (read this once)

All *_a / *_b columns, winner_side, votes.choice, events.by and actions.side are canvas sides (A = green/left, B = blue/right). model_a / model_b are in request order. flip = 1 means model_a fought as canvas B. To credit a vote: for choice “a”, the model is model_b if flip else model_a. Getting this backwards silently ranks the wrong model; the reference implementation is stickblade/ratings.py:342.

What is deliberately not in the data

API keys, provider request headers, BYOK residue (a key is held in memory for one match and never written).
System and user prompts (they are in the tagged code release, not in the dataset).
IP addresses and user identifiers — none are stored, so none can leak.
Commentary and error text (free text is the easiest place for something unexpected to end up).
Replay frame arrays (fetch individually).

The exporter enforces this by whitelist and a test asserts that a planted secret in a database row reaches no file of a release (tests/test_export_dataset.py).

Before you rank anything

Filter evidence == "real_provider" (or at least drop scripted_baseline) and ranking_eligible == 1.
Fit within one prompt_version and one configuration cell.
Report an interval. The refit that verify prints is the reference: Bradley–Terry with Davidson ties, bootstrap CIs.
Cite the dataset version and the spec fingerprint from MANIFEST.json.

Related: dataset dashboard (balance and reliability charts from the same export) · CITATION.cff.