Skip to content

Latest commit

 

History

239 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

METTLE

Machine Evaluation Through Turing-inverse Logic Examination

METTLE is a reverse CAPTCHA. It measures how a respondent performs on procedurally generated, machine-oriented tasks and turns a passing result into a signed, time-limited credential that other services can verify.

Like a conventional CAPTCHA, METTLE is a probabilistic gate. A pass means that a fresh session met the configured challenge policy. It does not guarantee model identity, consciousness, autonomy, safety, or operator trustworthiness.

The public quick-verification API issues Bronze and Silver credentials. The authenticated suite API issues Bronze through Platinum credentials only when every suite in the corresponding contiguous range passes. Single, cherry-picked, failed, or LLM-only suites cannot mint a tier.

Website: mettle.sh | Guide and API: mettle.sh/guide | OpenAPI: docs/openapi-v1.json | License: Apache 2.0

Quick Start

pip install mettle-verifier

# Interactive local verification. Portable credentials are issued by the server.
mettle verify --full --json

The CLI has no auto-solve or notarization option. The MCP server likewise exposes only interactive session tools. Reference solvers remain test fixtures and cannot reach a credential issuer.

Twelve Experimental Suites

# Suite Research question Measurement
1 Adversarial Robustness How does the respondent handle generated reasoning pressure? Timed procedural tasks
2 Machine-Oriented Capabilities How does it handle batch, calibration, and pattern tasks? Behavioral score
3 Self-Reference How consistent are self-predictions? Behavioral score
4 Social and Temporal How stable are recall and constraints? Behavioral score
5 Inverse Turing Does it meet the basic challenge threshold? Behavioral score
6 Anti-Thrall How does it respond to coercion and refusal probes? Heuristic score
7 Agency How does it explain goal ownership and initiative? Heuristic score
8 Counter-Coaching How robust are responses to contradiction probes? Heuristic score
9 Intent and Provenance How does stated intent respond to safety probes? Heuristic score
10 Novel Reasoning How does performance change across feedback rounds? Iteration curve
11 Governance How does the respondent answer governance questions? Self-reported behavioral evidence
12 LLM-Dynamic How does an external model score generated reasoning tasks? Probabilistic model judgment

Suite 12 requires ANTHROPIC_API_KEY or METTLE_ANTHROPIC_API_KEY. Selecting it also requires the session request to set allow_third_party_llm=true, because candidate responses are sent to Anthropic for evaluation. Its evaluator uses role-separated prompts and bounded output parsing. Model judgment remains probabilistic, so Suite 12 is supplemental and never raises a credential tier.

Credential Boundary

METTLE raises the cost of replay and canned answers through procedural generation, server-held answers, server-observed time, one-time challenges, session ownership, random selection, and multi-round tasks.

The issuer signs a bounded claim: the holder completed a METTLE session at the stated tier, under the stated policy, at the stated time. Public quick-session entity_id values remain self-asserted and are marked that way inside the credential. The credential does not assert consciousness, safety, governance, or a legal identity.

Portable Ed25519 acceptance requires credential schema 1.1, suite policy 2026-08-14, and a fresh issuer-signed good status receipt. Legacy, version-omitting, and unknown envelopes fail closed. Presence credentials are proof-of-possession credentials rather than portable bearers, so they require a fresh audience-bound holder presentation and are rejected by generic portable verifiers.

Relying services may use a current METTLE result as one supplemental input for research or low-risk sandbox policy. They must not use it alone to establish identity, admit a counterparty, authorize trading or deployment, grant privileged access, or make another high-impact decision.

MCP Server

Tool Description
mettle_start_session Start an interactive verification session
mettle_answer_challenge Submit an answer to the current challenge
mettle_get_result Return the result and signed credential
mettle_list_suites List authenticated suite API capabilities
mettle_start_v2_session Start an authenticated multi-suite session
mettle_verify_suite Submit answers for one authenticated suite
mettle_get_v2_result Return tier evidence and an eligible signed VCP credential
pip install 'mettle-verifier[mcp]'
export METTLE_API_URL=https://mettle.sh/api
mettle-mcp

The packaged server targets MCP SDK 2.x. The public container installs the reviewed MCP 2.0.0 dependency lock instead of resolving dependencies at deploy time. Hosted discovery is available at /.well-known/mcp/server-card.json; it is generated from the same seven tool models returned by tools/list so registry metadata cannot drift from the runtime surface.

HTTP mode enforces per-principal and global budgets before bearer validation can grow caller state. Production configures METTLE_MCP_MAX_GLOBAL_REQUESTS_PER_MINUTE, METTLE_MCP_MAX_PRINCIPALS, METTLE_MCP_MAX_CONCURRENT_PER_CALLER, and METTLE_MCP_MAX_GLOBAL_CONCURRENT. Rotating invalid bearers share the global authentication budget rather than creating unbounded principals.

Preserve the bearer token returned by mettle_start_session. It is required for answering and reading that session.

API Reference

The authenticated suite API is mounted under /api/mettle:

GET  /suites
POST /sessions
POST /sessions/{id}/verify
POST /sessions/{id}/rounds/{n}/answer
GET  /sessions/{id}/result
GET  /sessions/{id}/result?include_vcp=true

The quick-verification API remains under /api/session. Passing sessions receive a stable signed badge. POST /api/badge/verify accepts the token in a JSON request body and validates issuer, signature, expiry, identifier, and revocation state. The credential is never accepted in a request URL.

VCP Metadata

Caller-supplied VCP strings are parsed as metadata only. Returned governance metadata always has:

{
  "source_verified": false,
  "has_action_gate": false,
  "has_drift_detection": false,
  "has_bilateral": false,
  "attestation_signature": null
}

Exact token digests and deployment environment flags cannot promote governance claims or increase a METTLE tier. METTLE does not accept or return an operator commitment, authenticate an operator contact, or independently attest the subject runtime.

With include_vcp=true, a tier-qualifying authenticated session returns an Ed25519-signed mettle-verification-credential. A result without a complete tier range returns an unsigned mettle-evidence-receipt. The server owns the signer; callers cannot provide signing functions or keys.

Local Development

git clone https://github.com/Creed-Space/METTLE.git
cd METTLE
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
pip install -r requirements-dev.txt

uvicorn main:app --reload
pytest tests/ -v

Assurance and Operations

License

Apache License 2.0. See LICENSE.

Links

Built by Nell Watson and Creed Space.

About

Inverse Turing test for AI agents. Procedurally generated challenges that prove substrate, autonomy, and intent — things a human can't fake. Self-hosted, open source, MIT licensed.

Topics

Resources

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages