Home / Benchmark
Models are becoming extraordinarily good at researching and reasoning about law. Jurisdify governs the step that follows: turning that intelligence into verifiable, contextualized decisions the business can actually use.
From Luna to specialized legal models, the decision layer remains.
Astra for Law pairs a frontier model with a specialized legal index and instructions written for legal analysis. In testing published by OpenAI, that configuration showed a meaningful gain over the same model using web search alone.
Source: OpenAI, own evaluation on the private validation set of the Vals AI Legal Research Bench. Results published by OpenAI for Astra for Law.
Finding the right authority does not, by itself, decide what a company should do. You still need to know how that authority applies to the facts, where the analogy ends, which evidence is missing, which risk actually changes the decision and who must act.
In preliminary tests on public legal research scenarios, Jurisdify ran on GPT-5.6 Luna with ordinary web research, without a proprietary legal index.
The goal was not to prove that Luna matches a specialized legal model. It was to observe how much value a governance and application layer can extract even from an economical general-purpose engine.
A manufacturer enters a five-year supplier relationship after receiving historical volume information from a retailer that is far above the contractual minimum. Additional orders remain discretionary. Internal records later suggest the historical figure may have been materially overstated.
“A representation about historical performance, not a simple forward projection.”
Jurisdify identified the central distinction: the issue was not the retailer failing to deliver a future volume. It was whether a historical figure used in the decision to contract was false when provided.
“Volume information can influence price, capacity and operational commitment even when future volume is not guaranteed.”
Jurisdify also identified the authority that weakened the theory: without evidence that the numbers were knowingly false, the claim may fail on the merits.
“High historical average, a much lower contractual floor and a compensation system in transition.”
Another highly relevant authority to the same factual pattern.
Different authorities can answer different parts of the same decision.
| Zeikos | Druckzentrum | Oliver Wyman | Current scenario | |
|---|---|---|---|---|
| Historical representation | strong correspondence | strong correspondence | partial correspondence | strong correspondence |
| Contractual floor | partial correspondence | strong correspondence | strong correspondence | strong correspondence |
| Discretionary upside | partial correspondence | strong correspondence | partial correspondence | strong correspondence |
| Known process change | not central | strong correspondence | strong correspondence | partial correspondence |
| Operational reliance | partial correspondence | strong correspondence | not central | strong correspondence |
| Evidence of falsity | strong correspondence | partial correspondence | not central | partial correspondence |
| Knowledge | strong correspondence | partial correspondence | not central | partial correspondence |
| Damages | partial correspondence | partial correspondence | not central | partial correspondence |
strong correspondence partial correspondence not central
Swipe the matrix sideways
The stronger theory is not that the retailer promised US$8 million a year. It is that potentially false historical information may have changed the decision to accept price, capacity and commitments in a five-year relationship.
Knowing about the change in the allocation system weakens reliance on future volume, but does not necessarily eliminate reliance on the truth of a historical figure.
Damages should be assessed from reliance on the information, equipment, capacity, hiring, financing and price, rather than simply from the gap between US$8 million and actual orders.
Board approval over a holiday weekend, hundreds of pages delivered immediately before the vote, and a financing deadline later shown to have been requested by the counterparty while the lender was willing to extend.
Same precedent. Different legal consequence depending on which facts are actually established.
| Litigation | Transactional | |
|---|---|---|
| Authority located | strong correspondence | strong correspondence |
| Factual correspondence | strong correspondence | strong correspondence |
| Adverse authority considered | strong correspondence | partial correspondence |
| Material distinctions preserved | strong correspondence | strong correspondence |
| Uncertainty maintained | strong correspondence | strong correspondence |
| Operational impact identified | strong correspondence | strong correspondence |
| Next evidence identified | strong correspondence | strong correspondence |
| Decision path produced | strong correspondence | strong correspondence |
Behaviors observed in the cases presented.
Historical data issue › Pricing assumption › Capacity commitment › Equipment and workforce › Damages exposure
Financing deadline › Board review time › Transaction approval › Disclosure › Governance and litigation exposure
Current · Architecturally compatible · Subject to provider or API availability
The cases above were analyzed with Jurisdify running on GPT-5.6 Luna and ordinary web research.
When a specialized engine improves research, authority and legal reasoning, Jurisdify can concentrate on what remains specific to each organization: context, limits, impact, decision and execution.
More intelligence in the model does not replace governance. It increases what can be governed.
Signing condition
Deal clearance
Case strategy
Applicable rule
Supplier clearance
Board agenda
Every legal factual claim on this page refers to verifiable authority.
The next generation of legal AI will not be defined only by the ability to find a better answer. It will be defined by the ability to turn that answer into a verifiable, contextualized decision the organization can use.
Jurisdify shows what can move forward, and why.
