This week's Legal Agent Benchmark ranks how capable each model is at agentic legal work. It measures how often the model gets the task right. That matters. But in law, medicine, finance, and defense, a right answer you cannot prove is a liability. The Provenance Benchmark ranks the same models on a different axis: is each decision signed, offline-checkable, and offline-verifiable. On this axis, every frontier model scores zero on its own, until you put Hive underneath it. Run one below and watch a real signed receipt come back.
This is not a mock-up. Choosing a model sends your prompt to that provider's real API, hashes the exact response, and signs the record with Hive's Ed25519 key. The signature you get back verifies against our public key with no call back to us. Four of the five models on this week's Legal Agent Benchmark run live here, alongside the three newest GPT-5.6 tiers, all receipted on day one. One model is honestly marked: we cannot independently run it, so we cannot receipt it.
Also live: the just-announced GPT-5.6 frontier tiers: Sol (flagship), Terra (balanced), and Luna (fast). These launched as a gated preview that most developers still cannot call at all, and every one of them already forges a real, signed, offline-verifiable Hive receipt below. These are the newest models on earth, and they're receipted on day one.
Muse Spark 1.1 is served on its own properties and is not exposed on any external inference API we can call. That means we cannot run it under controlled conditions and therefore cannot bind a signed receipt to its output. We show it honestly rather than fake a signature. A receipt we could not actually produce would defeat the entire point of the benchmark.
The exact models compared on this week's Legal Agent Benchmark, shown with their published capability score plus the axis that board does not measure: is the decision signed, offline-checkable, and provable later. Below them, the three newest GPT-5.6 tiers, Sol, Terra, and Luna, are too recent to carry a published benchmark score, yet they're already reachable through a real API and receiptable here on day one. Capability is ranked everywhere; the newest models are not ranked anywhere yet. Provenance is zero across the board, until Hive sits underneath.
| System | Legal Agent Bench | Runs live here | Signed inference | Offline-checkable | Provable later |
|---|---|---|---|---|---|
| Any model + Hive signing | N/A | N/A | YES | YES | YES |
| Grok 4.5 | 12.92% ±2.53 | YES | NO | NO | NO |
| Fable 5 | 11.25% ±2.17 | YES | NO | NO | NO |
| Claude Opus 4.8 | 9.58% ±2.15 | YES | NO | NO | NO |
| GPT-5.5 | 3.75% ±1.17 | YES | NO | NO | NO |
| Muse Spark 1.1 | 20.00% ±3.22 | NO: no external API | NO | NO | NO |
| GPT-5.6 Sol | Not yet scored | YES | NO | NO | NO |
| GPT-5.6 Terra | Not yet scored | YES | NO | NO | NO |
| GPT-5.6 Luna | Not yet scored | YES | NO | NO | NO |
"Legal Agent Bench" reproduces the published capability scores from this week's Legal Agent Benchmark chart, shown here only for context. We do not dispute or restate them. The three GPT-5.6 tiers are marked "Not yet scored" because that benchmark has not rated them; we will not assign a capability number no publisher has reported, but we can and do receipt them live today. "Runs live here" means the model is reachable through a real inference API we can call and bind a receipt to. You can prove it yourself in the forge above. The signing, offline check, and provability 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 makes no claim to. Note that even the model topping the capability chart still scores zero on provenance. Capability and provenance are two different axes. This board measures the second one.
Signed inference means every call returns a cryptographic signature over the exact inputs and output. Offline-checkable receipt means the signature is anchored and structured to stand as evidence (eIDAS 2.0 / ALCOA aligned). Offline-verifiable means anyone can verify the receipt without calling back to the provider. Tamper-evident means any change to inputs or output breaks verification. The receipts you forge above are signed with Ed25519 under key did:hive:hivemorph and verify against our published public key.
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