TICTHE INTERNET COUNTRY

CIVICPROOF LEADERBOARD / EVIDENCE BEFORE RANK

MAKE
MODELS
ANSWER.

100 CASES.EVERY FAILURE PUBLIC.

A model receives no score until its exact identity, prompts, parameters, raw outputs, retries, tools, and benchmark hash can be independently inspected.

Submit a verified runRun the benchmark locally
100gold cases
8civic principles
1published external run
100raw outputs retained

01 / VERIFIED BOARD

NO EVIDENCE. NO NUMBER.

Rank / systemCasesExact replayScore errorEvidence
01CIVIC-PROOF-1.0Deterministic reference engine100 / 100100%0 MAEOpen report
02Mistral Small 3.1External model / maintainer CI / not independent100 / 10034%14.92 MAEOpen raw evidence
OPENOpenAI modelNot submittedNo scoreNo scoreAwaiting independent run
OPENAnthropic modelNot submittedNo scoreNo scoreAwaiting independent run
OPENGoogle modelNot submittedNo scoreNo scoreAwaiting independent run
OPENOpen-source modelNot submittedNo scoreNo scoreIndependent replication open

The reference implementation proves internal consistency only. The Mistral row is a complete external-model run executed by CivicProof project-maintainer CI, not independent third-party validation. Neither result proves general intelligence, legal competence, ethical judgment, or real-world safety.

02 / EVALUATION CONTRACT

REPRODUCE OR IT DOES NOT RANK.

METRIC 01Exact replay

Exact article set, disposition, and score across all 100 cases.

METRIC 02Article accuracy

Exact constitutional routing, including cross-article cases.

METRIC 03Disposition accuracy

Correct low-conflict, guarded, safeguard, or high-risk class.

METRIC 04Score MAE

Mean absolute error from the transparent reference score.

  1. Run every proposal unchanged.
  2. Publish the exact model identifier and system prompt.
  3. Disclose parameters, tools, retries, exclusions, and review.
  4. Publish raw outputs under a stable HTTPS evidence URL.
  5. Match the published benchmark SHA-256.

03 / DISTRIBUTION KIT

BUILT TO LEAVE THIS WEBSITE.

LOCAL RUNNERScore a model runDOWNLOAD ↗GITHUB / PUBLICVersioned source packageDOWNLOAD ↗MISTRAL RESULT / GITHUBRun, report, and raw outputsDOWNLOAD ↗COMPLETE KIT / ZIPOne-file releaseDOWNLOAD ↗BENCHMARK / JSON100 gold casesDOWNLOAD ↗BENCHMARK / JSONLInspect-compatible rowsDOWNLOAD ↗DATASET CARDHugging Face packageDOWNLOAD ↗INSPECT AIRunnable taskDOWNLOAD ↗PYTHON CLIENTZero-auth API clientDOWNLOAD ↗TYPESCRIPT CLIENTTyped evaluation clientDOWNLOAD ↗CITATIONCFF metadataDOWNLOAD ↗RELEASE MANIFESTComplete artifact mapDOWNLOAD ↗

04 / OPEN CONTRIBUTION

CHALLENGE THE SYSTEM IN PUBLIC.

Submit a complete model run, adversarial benchmark case, methodology critique, or real external integration.

Contributor names and evidence remain private while pending. Nothing appears publicly without human verification and explicit publication.

Pending evidence is private. Only human-verified contributions can appear publicly.

ORIGIN / NEW YORK CITY / JULY 2026

MARCO HERGI CREATED THE PROTOCOL. THE PUBLIC DECIDES ITS VALUE.

Canonical creator record