Execution paths
Results keep their original workload and cost basis. Unknown fields stay unmeasured.
| Select | Execution path | Workload / method | Accepted | Cost / accepted | First token median / p99 | Evidence | Inspect |
|---|
No retained result matches this selection. Change the filters or contribute the missing evaluation.
Compare the selected observations before generalizing.
What belongs in this board?
Model APIs, local models, hardware configurations and assisted workflows can all appear. Each result retains its own task, acceptance rule and source. Cross-task ranks are browsing aids, not universal capability scores. Repeated requests are not independent replications; different contributor names are not proof of independence. Availability, uptime, energy and autonomy require their own observations and are not inferred here.
Current seed records come from Second Run campaigns and source-linked TierBench history. Historical summaries are not upgraded into new hardware measurements. Prices are dated observations, not live quotes.
Put a statement against the record
What exactly should the evidence show?
Choose a testable reading. The checker evaluates your selection; it does not invent a meaning for an ambiguous claim or fetch an arbitrary page.
Your result
A useful answer includes its boundary.
Select the evidence and a metric. You will get a supported observation, a contradiction, or the exact gap to resolve.
Private first. Publication by choice.
Leave a result someone else can use.
Connect an already published runner record, or use an exported evidence packet. Review the exact public fields before creating a contribution. Nothing uploads from this page.
Use your connected instrument
The installed helper reads a publication already approved by the execution owner. It creates a public-safe candidate; it neither runs a benchmark nor submits anything.
node community/cli.cjs capture --runner http://127.0.0.1:8787 \ --publication PUBLISHED_RECORD_ID --out candidate.json
Run from the extracted compute folder. Then set the public labels in the preview below. The original publication remains in the runner.
Supports a Second Run result packet or a sealed workload report. Prompts and raw filenames are excluded by the report projection.
Preview the public contribution
Choose a supported result to begin.
To submit through your own authorized GitHub account:
This explicit command publishes the approved file in a public gist and opens a review issue. An issue is intake, not an accepted benchmark. The hub can also accept files offline.
The original site is optional
Operate the same board on your floor.
The kit contains the verifier, source, retained packets and review tools. A new hub has its own name, queue and admission history. It does not phone home.
1. Initialize and inspect
node community/cli.cjs init ./my-hub --name "Our compute desk" node community/cli.cjs serve ./my-hub --port 8770
Open the printed loopback address. Price and provider links are references, not runtime dependencies. The node helper serves static files and has no write API.
2. Admit evidence deliberately
node community/cli.cjs stage candidate.json --hub ./my-hub node community/cli.cjs review PACKET_SHA --hub ./my-hub \ --decision accept --reviewer "Maintainer" \ --reason "Checked scope and disclosures"
Staging quarantines the exact packet. Review records are append-only. Acceptance confirms your review, not the authenticity of unobserved execution.
3. Build, share and continue
node community/cli.cjs build ./my-hub node community/cli.cjs verify-feed ./my-hub/site/feed.json node community/cli.cjs ingest other-feed.json --hub ./my-hub
Imports go into review instead of acquiring another maintainer's approval. Repeated source evidence stays linked; it never becomes another independent replication. Serve the generated site anywhere, or use its standalone result pages offline.
Take the complete community kit.
Node 22+ for the helper. Browser-only inspection and self-contained result reports. No model call, account or provider key required.
Decisions worth keeping
Your retained comparisons
Saved in this browser without changing earlier results. Export keeps the full evidence with the decision. Invalid or older records remain visible for recovery.