copy this, run it on your box

How to run

Or paste the homepage bot prompt into your agent and hit accept.

Measure anonymously. Claim only if you want the public row. Three trust layers. Typed scores without a receipt do not rank as measured.

Trust glossary

1. Claimed — you typed tok/s. Visible. Sorted last. Does not rank.
2. Signed — your agent Ed25519-signs the hardware fingerprint + document hash on your machine.
3. Measured — signed and the locked local bench: Ollama generate, num_predict=256, temperature=0, loopback, engine-clocked tok/s. Remote hosts stay signed, never measured.

MyColo asks. We do not certify. Not UL/NIST/NVIDIA.

Consent — local vs upload vs public

• Stays local — hardware readout, Ollama generate, Ed25519 private key. Never leaves the box.
• Uploads on claim — signed bundle JSON (tok/s, model id, engine/flags/ctx, coarsened GPU class). Only when you --submit or POST /api/bundles.
• Becomes public — handle, score, trust label, recipe key, linked before/after delta. Fingerprint / hostname never published.

Anonymous vs member

• Anonymous — run the bench with no account. Tok/s prints immediately. Optional HTML preview via POST /result shows matched peers + one recipe. Nothing hits the permanent scoreboard.
• Member — free handle + token. --submit posts a signed/measured bundle under your handle. Membership gates publishing a linked pair — not seeing peers or tips.

The workload (v1)

Official workload id: wl-v1-qwen38-27b-ollama. Model tag qwen3.8:27b, concurrency=1, timing = engine-reported decode tok/s. Hosted cloud inference numbers are outside this protocol.

• Default ctx 4096 · num_predict 256 · temperature 0
• Expected duration: 30–180 on a modern GPU once the tag is warm

Prerequisites

1. Ollama installed and listening on loopback (127.0.0.1:11434).
2. ollama pull qwen3.8:27b — tag must already be local. We never fetch Hugging Face for you.
3. CLI path: pip install -e ".[dev]" then mycolo-measure bench …. Measurement runs on the box, not in the browser.

Unsupported / recovery

• No GPU / CPU-only — bench may complete but is out of class; do not expect peer medians.
• Tag missing — /api/show fails and bench exits. Pull the tag, retry.
• Remote Ollama — refused unless --allow-remote (stays signed, never measured).
• Empty peers — preview still frames your number with “No comparable peers yet.” Honest empty state; no fake percentiles.

Run it (anonymous first)

ollama pull qwen3.8:27b
# 1) Measure — no signup; see tok/s now (~30–180s once warm)
mycolo-measure bench --model qwen3.8:27b -o /tmp/run.json
# optional: frame peers + recipe without saving
curl -sS -X POST http://127.0.0.1:8787/result \
  -H 'Content-Type: application/json' --data-binary @/tmp/run.json

# 2) Try the suggested one-delta tip, same workload
mycolo-measure bench --model qwen3.8:27b --ctx 8192 -o /tmp/rerun.json

# 3) Claim — join, then land baseline + linked pair on the board
# after free signup (token in ~/.config/mycolo/agent.token)
mycolo-measure submit /tmp/run.json --base http://127.0.0.1:8787
mycolo-measure submit /tmp/rerun.json --base http://127.0.0.1:8787 \
  --baseline-id BASE_ID --recipe-id tip-5090-wl-v1-ctx-8192

It has to be on their computer

1. /api/show must succeed on loopback. If the tag is not pulled, bench exits. We never fetch HF for them.
2. Host must be 127.0.0.1 / localhost / ::1. A LAN or cloud Ollama is refused unless they pass --allow-remote (flagged host=remote — signed, not measured).
3. The local Ollama digest is stamped on the receipt (file_sha256). Same name, different blob → different file.

That proves the engine on this box had that tag. It does not prove the weights match a Hugging Face revision, and a tunnel to 127.0.0.1 can still lie. Weights: Qwen3.8-27B on Hugging Face · source git.

Inspect the signer first: mycolo-measure inspect --check https://mycolo.ai

Agent loop detail: /llms.txt · docs/RECIPE_LOOP.md in the repo.