Submit a run

You run the open engine against your own model and submit a signed result bundle. Deterministic results are independently re-runnable, so the leaderboard is trustworthy without us hosting any GPUs.

1. Install the client + fetch the challenge corpus

pipx install peakstone
peakstone corpus sync
peakstone login    # optional: attributes runs to your GitHub handle

2. Run the official level on your own hardware

A level is a versioned, pinned challenge selection — standard is the official one the leaderboard is scoped to, so your run is directly comparable. The easiest path is the local daemon — it serves your model, runs the level, and chains the judge pass automatically:

peakstone serve --detach
peakstone jobs add <model-name> --level standard
# -> results/job-<id>/bundle.json  (schema-valid, content-addressed, ed25519-signed)

3. Submit the signed bundle

peakstone submit results/job-<id>/bundle.json

Or let the daemon publish finished runs itself (opt-in: auto_submit = true under [gateway]). Raw HTTP works too — plain or xz-compressed JSON:

curl -X POST https://peakstone.ai/api/submissions \
  -H 'content-type: application/json' \
  --data @results/job-<id>/bundle.json

What gets recorded

• Exact model identity — HF repo + revision, file SHA-256, quant, engine version.
• Sampling + serve flags + the hardware/driver it ran on (the VRAM facet).
• Per-challenge content hashes, transcripts, and scores.

The whole bundle is content-addressed and signed by your key — the signature is your root identity. Link that key to GitHub with peakstone login: optional, but it attributes runs to your handle and lets your reproductions count toward the community-verified (ranked) tier.