Run an unsaved action_ai_research column against ONE row and answer what its cell would hold: before the column exists, or with the edited prompt of one that does (the drawer's per-row Preview). The same gate, model resolution and output coercion as the paid run, so a preview never promises what the run refuses. PRICED AS ONE ROW: the platform's charge comes back in credits_charged; a row the gate skips (will_run false, skip_reason run_condition_not_met or empty_context) costs nothing. Writes nothing: no cell, no run, no ledger receipt. Every call asks the model again with a fresh key, so a repeat may answer differently and is charged again. Read value, reasoning and confidence; error_code names a refusal before any charge (model_not_in_catalog, web_search_not_supported), a failed call, or an answer that could not be coerced to payload.output (the model ran and was paid). Show the user the sample and its cost; the whole column is create_playtable_run after estimate_playtable_run.

Contract:

• MCP tool preview_playtable_field_run, registry package mcp.playtables/playtable_fields, mount playtables.columns.
• Operation action, response envelope action.
• Flags: none.