The Provenance Benchmark · Patent Pending

Every AI benchmark scores the answer.
None of them score whether you can prove it.

Capability leaderboards rank whether a model got the task right, how fast, and how cheap. That matters. But in law, medicine, finance, and defense, a right answer you cannot prove is a liability. The Provenance Benchmark ranks AI systems on a different axis: is each decision signed, court-admissible, and offline-verifiable. On this axis, every frontier model scores zero. Only Hive signs.

7ms
Sign time · ML-DSA-65
112ms
Median first token
$0.0000732
Per signed call
100%
Calls verifiable offline

The capability board — with the column it’s missing

The same models on every published legal-agent leaderboard, scored the way those boards score them: all-pass rate and cost per task. We add the axis none of them measure — is the decision signed, court-admissible, and provable later. Capability is ranked everywhere. Provenance is zero across the board — until Hive.

System LAB all-pass Cost / task Signed inference Court-admissible Provable later
Any model + Hive signing model’s own +$0.0000732 / call YES YES YES
Claude Opus 4.7 7.1% $50.90 NO NO NO
Claude Sonnet 4.6 5.4% NO NO NO
GPT-5.5 2.1% NO NO NO
Gemini 3.5 Flash 0.8% NO NO NO
GLM 5.1 12 / 100 $121 / 100 tasks NO NO NO
Kimi K2.6 11 / 100 $84 / 100 tasks NO NO NO
DeepSeek V4 Pro NO NO NO

All-pass and cost-per-task figures are drawn from published legal-agent benchmark results (2026), reproduced only to show the axis they leave blank. The signing and verifiability columns reflect properties of the call as deployed by default, not model quality. Hive is a signing layer that sits underneath any model — it does not change a model’s task accuracy, and it makes no claim to. Capability and provenance are two different axes. This board measures the second one.

Why this matters where it counts: a capability board can tell you the best model for drafting a contract, reading a chart, or proposing a molecule. It cannot tell a regulator, a court, or an auditor what the model relied on when it decided. That is a signed-record problem, and no leaderboard scores it — until this one.

How the score is defined

Signed inference — every call returns a cryptographic signature over the exact inputs and output. Court-admissible receipt — the signature is anchored and structured to stand as evidence (eIDAS 2.0 / ALCOA aligned). Offline-verifiable — anyone can verify the receipt without calling back to the provider. Tamper-evident — any change to inputs or output breaks verification. Hive uses real ML-DSA-65 (NIST FIPS 204), signing in 7ms with no meaningful hit to the latency budget.

Where provenance is not optional

Legal agentic work (eIDAS 2.0, ALCOA), healthcare and PHI, clinical and pharma (21 CFR Part 11), finance (audit and recordkeeping), defense, and the EU AI Act (Article 13 audit trail). In every one of these, the question that decides liability is not “was the model good” — it is “can you prove what it did.”

See signed inference Read the benchmarks

The Provenance Benchmark is published by Hive Civilization. Capability rankings referenced here belong to their respective publishers; this board does not reproduce or dispute their scores — it measures a separate dimension those boards do not. Hive numbers (7ms sign time, 112ms median first token, $0.0000732 per signed call) are benchmarked and published at thehiveryiq.com/afir. Settlement on USDC, anchored to Base Mainnet. Patent Pending.