Research participation

Hansard Research is a separate participant application and study website. Installing Hansard does not enroll you in a study or send a research submission. The participant application and researcher platform share a private source repository. The public Hansard study kit defines the file format, counts, and encryption. Its public source does not make the participant application’s full source public.

Consent and enrollment

A study invitation identifies the study website and its consent form. The participant application shows what the study collects before enrollment. Enrollment happens only after agreement. For Prolific recruitment, the study website also receives the identifiers needed for payment. Those identifiers stay outside the study package.

The consent form must explain the measures, survey questions, retention, withdrawal rules, and who holds the study key. Studies that request conversation labeling must also explain the separate labeling request and the model provider’s retention terms.

What the study package contains

The participant application counts history on your computer. Before sending, it shows the study package as a table and as raw JSON. The package contains:

• Counts by calendar month: active days, conversations, messages sent, messages to agents, and the number of tools used.
• Counts by month and tool, using fixed product names such as ChatGPT, Claude Code, or Codex.
• Token counts as ranges, such as 1M to 10M, instead of exact values.
• For studies that request them, agent hours, actions, and interruptions by month and agent, and conversations by month and model maker.
• The survey answers you give.
• A random participant code, the number of months, tools, and rows removed, and the versions of the rules that built the file.

The count tables accept only months, numbers, and fixed vocabulary values. They exclude message text, prompts, replies, and tool output. They also exclude titles, file names, folder names, project names, websites, account names, email addresses, Prolific IDs, custom model names, and dates finer than a month.

Survey text is an explicit exception. It contains what you type. Enrollment details and payment identifiers go to the study website separately.

Conversation labeling

A study can request labels for a sample of conversations. Sampled records start included. Before starting, you can inspect the model instructions and every sampled record, then remove records you do not agree to send.

Starting relay labeling sends the included redacted conversation excerpts through the study server to its model provider. Each record can contain message text and tool activity, up to 16,000 characters. Redaction cannot guarantee removal of every secret. The application offers previews and an explicit start action. It does not require separate approval for each record.

When the study permits a local model, labeling can run on your computer. Only label counts enter the final study package. Relay labeling sends excerpts separately, before the final package review and send.

Review and send

You can remove months, tools, or individual month-and-tool rows before sending. Removing an item recomputes every total that includes it, including label counts. The application seals the reviewed package for the study’s key. A change to the reviewed package requires another review before saving or sending.

Participants without available history can explicitly continue without history after a completed count. Unreadable history or stale counts must not become an empty study package.

Withdrawal and local copies

Erasing the local study copy removes its imported history, saved packages, and imported export files. The participant application keeps submission receipts. Local erasure does not withdraw a submission or delete model-provider records. Use the study’s withdrawal process for data already sent. The study’s consent form explains retention and withdrawal limits.

See What leaves your machine for Hansard’s other network features and their controls.