turnset
Request access

Real human sessions with agents, licensed with the rights attached.

Turnset is present while the work happens. Every trace is recorded as a person directs an agent through work they needed done, under a permission granted before the session started. No scraping, no telemetry reads, no zipped-up history folders.

01 / The difference

Recorded live, not reconstructed after the fact.

Most workflow data on offer is an export. Someone finishes the work, finds the log file, uploads it and ticks a box. The left column is what that can tell you. The right column is what being there at the time can tell you.

An exported log A Turnset trace

Consent collected after the work, when a file gets uploaded

Consent recorded before the session existedThe contributor starts each capture session on purpose. The trace carries the terms version in force at that moment. Nothing arrives by default.

No way to tell a person from a script

Evidence a person was at the keyboardTurn timing, the corrections in the transcript, and for the top tier a repository and a test run. An uploaded file is a file.

Clean one-shot examples that resolve on the first try

The failures and the recovery after themWhat broke, what the error said, and how the person redirected the agent. This is the part that cannot be synthesized, and public data has almost none of it.

Text with no evidence the work succeeded

Outcome signals from the session itselfTest runs, exit codes, retries, commits, and whether the task resolved.

Rights asserted by whoever uploaded the file

A rights boundary enforced at capturePersonal, open-source and sandbox work only, stated before they start and checked against the trace.

Scrubbing applied to a file that already left the machine

Redaction before the data exists anywhere elseSecrets are removed on the contributor's own machine. Each trace carries the scrub version that ran and what it removed.

Why the human half matters. Model output is abundant and getting cheaper. What stays scarce is a competent person deciding what to ask for next, judging a wrong answer, and steering an agent through a task that did not go straight. That decision sequence is what a trace preserves.

02 / What you buy

Sold per trace, by tier.

A trace enters the tier its artefacts qualify it for. You pay for the tier you need, and only for the traces you license.

Conversation
A person directing an agent through a real task, with their own turns marked. For evaluation sets, SFT, and tuning user simulators on how people actually talk to agents.
Trajectory
Tool calls with their results, a checkable step inside the trace, and human steering after the agent started working. For RL, for labelled failures, and for preference pairs drawn from real corrections.
Reproducible
Repository identified, state recorded at start and exit, the diff, and tests that ran after the last edit. For benchmarks and environments seeded from work that happened.

Packages carry the provenance record for every trace, the scrub version that ran, and the tier evidence. Raw captures are never delivered, and there is no browser over individual contributors.

03 / Provenance

Written down while it happens.

A reconstruction is a claim rather than a record. Being in the capture path is what lets us hand you the record.

  • Consent is recorded per session, not assumed once. Contributors accept the terms, then choose which sessions to capture. Each trace carries the terms version from the moment of capture.
  • Rights are asserted at capture, not inferred later. Personal, open-source and sandbox work only. Employer code, client code and anything under NDA are out. Stated before they start, not certified afterwards.
  • The processing history is recorded, not asserted. Collector version, scrub version and what it removed, the host allowlist in force, and timestamps for capture, window close and finalization. A chain you can audit, per trace.
  • Removal runs forward, never backward. Every trace sits visible to its contributor for 24 hours and can be excluded free, so packages only contain what passed. Later withdrawal drops it from all future packages. Delivered packages are not recalled.
  • Secrets are removed on the contributor's machine. Tokens, keys and identifying paths are scrubbed before the data exists anywhere else. Privacy-preserved, not fully anonymized: traces this rich cannot support the stronger claim.
04 / Questions

What buyers ask first

Can a contributor recall data after we license it?
Not from a delivered package. Every trace sits with its contributor for 24 hours before it can be licensed, so packages contain only what passed that window. A later withdrawal removes the trace from future packages and nothing else.
How do you know a person did the work?
The session was started by the contributor through our collector, on their machine, under terms accepted first. Turn timing, the corrections in the transcript, and for the reproducible tier a repository and a test run all come from the wire, not from a form.
What about secrets and personal information?
Keys, tokens and identifying paths are removed on the contributor's machine before anything is written to disk. Each trace records which scrub version ran and what it removed. We describe the result as privacy-preserved; traces this rich cannot honestly be called anonymized.
Whose code is in a trace?
Personal, open-source and sandbox work only. Employer code, client code and anything under NDA are outside the pilot's scope, stated to the contributor before they start and checked against the trace before payment.
How is it priced?
Per trace, by tier. You pay for the traces you license. Tell us the coverage you need and we quote against it. Datasets are licensed under a written agreement, not click-through terms.

Tell us what you need it to cover.

Supply is being built cohort by cohort, so the shape of what we collect is open to influence. A language, a class of agent, a kind of task, or a rights posture your legal team can sign: say it early rather than late.