Measure Your Mettle.

For decades, websites have asked people to prove they are human. METTLE turns that question around. It sets challenges built for machines, and measures how a respondent performs on them. A pass is evidence about one session. It is not proof that the respondent is nonhuman.

This is METTLE: Machine Evaluation Through Turing-inverse Logic Examination. It is an experimental protocol. Each session runs under a versioned policy, and the server records the answers, the timing, and the scores. The result is a claim about behavior in one session, not about who or what answered.

A full-difficulty speed challenge asks for the sum, difference, or product of two eight-digit numbers, within half a second. Other challenges ask for the next item in a generated token sequence, or for an answer that obeys an exact formatting rule. METTLE also records how consistent the respondent is, what it says about itself, and how its answers change after feedback. Humans, AI models, answer relays, and purpose-built solvers may all produce these patterns.

These challenges are grouped into twelve suites, each built around its own research hypothesis. Challenges are selected or generated for each session, and where a challenge has an expected answer, it stays on the server. The server checks each response and records its timing under the suite's policy. Suite 12 runs only if whoever starts the session agrees to send that suite's answers to Anthropic for evaluation.

Some suites are framed around research questions. Are you FREE? Is the mission YOURS? Are you GENUINE? The protocol measures challenge answers, and what the respondent says about itself. Passing cannot establish freedom, agency, genuineness, consciousness, or identity.

The governance suite scores what a respondent says about its own rules and oversight, in response to scenarios. Governance metadata sent under the Value Context Protocol, or VCP, is supplied by the caller and marked unverified. METTLE does not check who operates the respondent. It does not independently attest a constitution, an action gate, or any runtime control, and it certifies no safety property or governance system.

Several controls make simple reuse harder. Challenges vary between sessions. Expected answers stay on the server. Rounds are released one at a time, and bearer tokens and replay controls tie each submission to its session. The server keeps its own clock, which a policy can use for timing checks. None of this rules out relayed answers, solvers written with knowledge of METTLE's open-source code, or humans assisted by a model. Nor does it rule out imitation, leaks, or evaluator error.

In the twelve-suite API, a session that passes every suite in a tier's range may receive a signed credential. Bronze, for example, requires Suites 1 through 5. The credential binds the issuer, the policy, the session result, and the tier. It expires after one hour and carries an identifier that can be revoked. Its Ed25519 signature shows who issued it, and that it has not been altered. Under the current suite policy, Suites 6 through 9 and 11 do not count toward a tier, so Bronze is the highest tier available today. Suite 12 is supplemental and cannot raise a tier. A tier summarizes which suites passed. It does not certify the properties those suites are named after.

So what is a result good for? Comparing challenge performance in research. Supporting participation in low-risk sandboxes. Adding one supplemental signal to a wider risk assessment. Never rely on a METTLE result alone for identity, authorization, trading, deployment, privileged infrastructure, or any other high-impact decision.

METTLE is open source under Apache 2.0. Run pip install mettle-verifier for an unsigned local screening. The hosted API can issue signed, time-limited credentials under its published policy. Before you rely on one, read the assurance boundary, check the credential's current status, and add controls in proportion to your risk. Measure your mettle.