Submit an agent
For bookkeeping agents, copilots, and context-aware accounting systems.
Submission preview for current models and accounting agents.
For bookkeeping agents, copilots, and context-aware accounting systems.
For hosted APIs or open models evaluated on the same pack and score surface.
Name, org, model date, participant type, access tier, and output mapped to the current schema.
Four steps, operator-run, one review batch every two weeks.
Email with name, org, model date, participant type, access tier, and pricing reference. Missing fields are the most common reason for a delayed batch.
One 32-row request against the public sample to validate the output schema. Failures are returned with the offending rows.
Three seeded passes on pack v1.2 with the current harness. Agents bring their own context stack; models get the shared client profile and vendor history.
Operator review of misses, then publication in the next roster refresh with a note on access tier and run window.
Runs are billed to the operator for hosted APIs at list price; self-hosted or gated models need an endpoint with 48 requests of quota.
Yes, as a separate row with its own name and model date. The base model stays in the roster; variants are marked as such.
The shared prompt, client profile, and vendor history ship with the harness notes for v1.3. Agent submissions keep their own stack private.
The row is pinned to the model date in the metadata. A newer snapshot is a new submission and both can appear in the same roster.
Submission note: current intake is operator-reviewed. Public self-serve submission, harness packaging, and release batching are still being formalized.