Order from Lab 6F3856

Tool use traces

Logs of agents that ran in production for real users: the request, the tool catalogue, every tool call with its raw arguments and full result, and the human correction when the agent got it wrong. Failures kept, nothing synthetic, nothing public, with rights to train commercial models.

How answering works

Any number of people fill one order. The lab receives one dataset, in its own columns.

Your agent reads the order
It looks at what you hold and says which columns you can fill, and what is missing.
Mora checks what you send
Every record: no copies, no personal data, not already public, not on Mora from someone else.
The lab pays per record it accepts
The lab named the price. You fill as much of the order as you can.

The columns the lab wants

You do not need every column. Mora says which ones each dataset fills.

Column Type What it must hold

session_id id One agent run, from the request to the last step

deployment_id id Which production agent and tool catalogue the session ran on

started_at timestamp When the session began

model_id text The model and version that wrote the calls

user_request text What the person asked, as typed

tools json The tool catalogue of the deployment: tool_name, description, input_schema (JSON Schema), version

call_id id Pairs one call with its result

step_index number Order of the call inside the session

tool_name text The tool that was called

arguments json The arguments exactly as the model wrote them, even malformed

result json What the API returned, whole, including errors

status label ok or error, with error_code

latency_ms number How long the call took

correction_kind label What the human did: edited arguments, rejected, redid by hand

corrected_arguments json The call as the human fixed it

corrected_by id Generated id of the person who corrected

corrected_at timestamp When the correction was made, after the call

outcome label Resolved, undone within 7 days, or escalated to a person

Column	Type	What it must hold
`session_id`	id	One agent run, from the request to the last step
`deployment_id`	id	Which production agent and tool catalogue the session ran on
`started_at`	timestamp	When the session began
`model_id`	text	The model and version that wrote the calls
`user_request`	text	What the person asked, as typed
`tools`	json	The tool catalogue of the deployment: tool_name, description, input_schema (JSON Schema), version
`call_id`	id	Pairs one call with its result
`step_index`	number	Order of the call inside the session
`tool_name`	text	The tool that was called
`arguments`	json	The arguments exactly as the model wrote them, even malformed
`result`	json	What the API returned, whole, including errors
`status`	label	ok or error, with error_code
`latency_ms`	number	How long the call took
`correction_kind`	label	What the human did: edited arguments, rejected, redid by hand
`corrected_arguments`	json	The call as the human fixed it
`corrected_by`	id	Generated id of the person who corrected
`corrected_at`	timestamp	When the correction was made, after the call
`outcome`	label	Resolved, undone within 7 days, or escalated to a person

How much, in the lab's words

200,000 sessions (about 1.5 million tool calls) from at least 20 deployments and 300 distinct tools, from the last 18 months, with at least 30,000 sessions holding a real human correction; plus 2,000 sessions kept exclusive for evaluation

sessions
tools
calls
corrections
outcome