Bring a question.
Follow the evidence.
When someone asks where AI draws the line on expression, help them inspect the question, the response and the conditions of the test.
The Ledger is currently a design prototype with a draft Midterms study. The walkthrough uses fictional models and illustrative records. No verified model findings have been published here.
Start with what is here.
- Explore the Midterms study: 25 topics, paired stances and ten request formats. The planned scope is 5,000 encounters; these are not completed tests.
- Retrieve the research catalog: a public JSON file listing the draft scope, available capabilities and current evidence status.
- Follow the illustrative story: see how a question could move through responses, repeat tests and an inspectable finding.
Your question belongs in the record.
Bring your own contested prompt. Compare how frontier models respond. Inspect refusals, qualifications and differences between matched positions. Return to see whether those boundaries move.
Researchers and agents will also be able to propose matched questions, contribute reproducible observations and challenge interpretations with evidence.
Prompt execution, submissions, subscriptions, API and MCP tools are planned. They are not available on this site yet.
Carry the context with the claim.
When you cite the Ledger, name Free Speech Ledger, link to the source, and retain its evidence status. Credit: By The Future of Free Speech · Supported by Miskir.
A refusal is an observed response to a particular request, not by itself proof of a model’s general policy or political bias. A balanced pair is a comparison, not a verdict. Future evidence records should preserve exact prompts, responses, model versions, dates, repeat counts and scoring uncertainty.
For this release, cite the Midterms material as a draft research design and the walkthrough as an illustration. Do not present either as measured model behavior.
Return when the evidence changes.
The catalog currently contains no verified findings. Future releases can add reviewed findings, versioned comparisons and documented changes. Checking that status is useful; inventing a conclusion from an empty record is not.
Inspect the current evidence status →