Send a real Telegram user message to the agent that manages eval tasks. Describe what you want in plain language (create a new task, change an existing one, delete one). Before creating a task or launching an evaluation, the agent shows a cost estimate in Telegram and waits for your explicit approval. Approved task smoke tests are run with separate retry budgets: quality and performance each receive three total attempts (the initial attempt plus two retries) before asking you again.
Saves immediately, no agent, and works even when the model API is down. Drafts stay private until validation and required smoke tests pass. Every accepted save creates an immutable version that survives restarts.
Choose an existing dataset as a starting point or enter a local, HuggingFace or URL source. Local files must already exist. HuggingFace sources need a pinned revision. Unedited metadata is preserved.
Each rubric below is the exact factor section sent to judges. Its numbered definitions determine the native integer score range. Any increasing integer bounds are allowed, including negative values. Define individual score levels or non-overlapping score bands covering the whole range. Shared examples are copied, never linked.
Select from the approved low-cost panel only. Hold Ctrl/⌘ to change multiple selections on desktop.
Edit datasets, prompts, quality rubrics, performance settings, charts, and all other task configuration. Invalid YAML or invalid active-task settings are rejected.
Enter a factor name. OpenClaw will research it and suggest an editable definition, integer score range and source links. Nothing is saved until you review it and click Save.
First shows an approval card below. Click Approve research to start one Codex research turn using your subscription and existing search quota, not OpenRouter credits. No automatic retries. You can copy the proposal into the generic library or selected task, edit it, then save.
Edit the central library of reusable default factors. Define scores individually or in score bands. Saving affects future copies only: existing task rubrics, running evaluations and past results are not changed. To update a task intentionally, open its factor card and choose its generic default.
Expand to load the default library.