Differential Validation
Independent judges, grounded critics and task-specific checks scrutinise high-impact outputs against evidence, improving accuracy and making uncertainty visible when human review is required.
Differential Validation
Independent judges, grounded critics and task-specific checks scrutinise high-impact outputs against evidence, improving accuracy and making uncertainty visible when human review is required.
Differential Validation
Independent judges, grounded critics and task-specific checks scrutinise high-impact outputs against evidence, improving accuracy and making uncertainty visible when human review is required.
THE VALIDATION LAYER: HOW IT WORKS
Two ways to earn trust
AI models can sound certain even when they are wrong. To validate high-impact AI outputs, Luminance separates generation from validation through a multi-model approach by asking the model that produces an answer to grade its own work, but that assessment is treated as one input only.
Luminance uses two complementary approaches, depending on the type of work being performed.
Panel of Judges tests whether independent models reach the same conclusion.
Grounded Critics check whether an answer is supported by the underlying contract and meets the requirements of the task.
Both lead to the same clear outcome: verified or required review.
1. Panel of Judges: Independent models test the conclusion
For structured tasks such as identifying clauses or extracting key information, multiple models assess the same question independently.
Luminance compares their conclusions and applies a clear agreement rule. When the required agreement is reached, the answer is verified. When it is not, the answer is flagged for review.
One model’s answer is not accepted on its own. It must be supported by the wider panel.
2. Grounded Critics: Separate critics test the evidence
For more open-ended answers, a separate model checks the response against the source contract and criteria specific to the task.
It asks whether the answer is supported by the evidence, whether anything important is missing, and whether the response meets the requirements of the feature.
Every required check must pass before the answer is verified.
The answer is not judged by how convincing it sounds, but by whether the contract supports it.
The model produces the answer. Luminance determines whether it has earned trust.
Grounded critics and feature-specific checks, the Panel of Judges challenges high-impact outputs through independent agreement, source evidence and criteria tailored to the legal task. This makes clear what has been verified, what requires human review and why.
THE VALIDATION LAYER: HOW IT WORKS
Two ways to earn trust
AI models can sound certain even when they are wrong. To validate high-impact AI outputs, Luminance separates generation from validation through a multi-model approach by asking the model that produces an answer to grade its own work, but that assessment is treated as one input only.
Luminance uses two complementary approaches, depending on the type of work being performed.
Panel of Judges tests whether independent models reach the same conclusion.
Grounded Critics check whether an answer is supported by the underlying contract and meets the requirements of the task.
Both lead to the same clear outcome: verified or required review.
1. Panel of Judges: Independent models test the conclusion
For structured tasks such as identifying clauses or extracting key information, multiple models assess the same question independently.
Luminance compares their conclusions and applies a clear agreement rule. When the required agreement is reached, the answer is verified. When it is not, the answer is flagged for review.
One model’s answer is not accepted on its own. It must be supported by the wider panel.
2. Grounded Critics: Separate critics test the evidence
For more open-ended answers, a separate model checks the response against the source contract and criteria specific to the task.
It asks whether the answer is supported by the evidence, whether anything important is missing, and whether the response meets the requirements of the feature.
Every required check must pass before the answer is verified.
The answer is not judged by how convincing it sounds, but by whether the contract supports it.
The model produces the answer. Luminance determines whether it has earned trust.
Grounded critics and feature-specific checks, the Panel of Judges challenges high-impact outputs through independent agreement, source evidence and criteria tailored to the legal task. This makes clear what has been verified, what requires human review and why.
THE VALIDATION LAYER: HOW IT WORKS
Two ways to earn trust
AI models can sound certain even when they are wrong. To validate high-impact AI outputs, Luminance separates generation from validation through a multi-model approach by asking the model that produces an answer to grade its own work, but that assessment is treated as one input only.
Luminance uses two complementary approaches, depending on the type of work being performed.
Panel of Judges tests whether independent models reach the same conclusion.
Grounded Critics check whether an answer is supported by the underlying contract and meets the requirements of the task.
Both lead to the same clear outcome: verified or required review.
1. Panel of Judges: Independent models test the conclusion
For structured tasks such as identifying clauses or extracting key information, multiple models assess the same question independently.
Luminance compares their conclusions and applies a clear agreement rule. When the required agreement is reached, the answer is verified. When it is not, the answer is flagged for review.
One model’s answer is not accepted on its own. It must be supported by the wider panel.
2. Grounded Critics: Separate critics test the evidence
For more open-ended answers, a separate model checks the response against the source contract and criteria specific to the task.
It asks whether the answer is supported by the evidence, whether anything important is missing, and whether the response meets the requirements of the feature.
Every required check must pass before the answer is verified.
The answer is not judged by how convincing it sounds, but by whether the contract supports it.
The model produces the answer. Luminance determines whether it has earned trust.
Grounded critics and feature-specific checks, the Panel of Judges challenges high-impact outputs through independent agreement, source evidence and criteria tailored to the legal task. This makes clear what has been verified, what requires human review and why.
Independent by design
Every prediction must pass an independent check
The model that produces an answer is not treated as the authority on whether that answer is trustworthy. Instead, a different model or group of judges scrutinizes it against evidence or defined criteria. A clear decision then determines whether the answer is verified or requires review.
The model proposes. Luminance decides what can be trusted.
Predict: The selected model produces an answer or set of candidate predictions.
Verify: A separate critic or consensus path scrutinizes the result.
Decide: A fixed gate returns a clear outcome: verified, or requires review.
Independent by design
Every prediction must pass an independent check
The model that produces an answer is not treated as the authority on whether that answer is trustworthy. Instead, a different model or group of judges scrutinizes it against evidence or defined criteria. A clear decision then determines whether the answer is verified or requires review.
The model proposes. Luminance decides what can be trusted.
Predict: The selected model produces an answer or set of candidate predictions.
Verify: A separate critic or consensus path scrutinizes the result.
Decide: A fixed gate returns a clear outcome: verified, or requires review.
Independent by design
Every prediction must pass an independent check
The model that produces an answer is not treated as the authority on whether that answer is trustworthy. Instead, a different model or group of judges scrutinizes it against evidence or defined criteria. A clear decision then determines whether the answer is verified or requires review.
The model proposes. Luminance decides what can be trusted.
Predict: The selected model produces an answer or set of candidate predictions.
Verify: A separate critic or consensus path scrutinizes the result.
Decide: A fixed gate returns a clear outcome: verified, or requires review.
Validation built around the work
Apply the right test to every answer
Multi-model consensus is used where independent agreement matters. Grounded critics test whether an answer is supported by the source. Feature-specific checks assess requirements such as completeness, consistency, and groundness.
Independent Agreement: Do separate models reach the same conclusion?
Supporting Evidence: Is the answer supported by the underlying contract?
Task requirements: Is the answer complete, consistent and appropriate for that specific use case?
The result is validation designed around the work being performed; not a single confidence setting applied to every answer.
Validation built around the work
Apply the right test to every answer
Multi-model consensus is used where independent agreement matters. Grounded critics test whether an answer is supported by the source. Feature-specific checks assess requirements such as completeness, consistency, and groundness.
Independent Agreement: Do separate models reach the same conclusion?
Supporting Evidence: Is the answer supported by the underlying contract?
Task requirements: Is the answer complete, consistent and appropriate for that specific use case?
The result is validation designed around the work being performed; not a single confidence setting applied to every answer.
Validation built around the work
Apply the right test to every answer
Multi-model consensus is used where independent agreement matters. Grounded critics test whether an answer is supported by the source. Feature-specific checks assess requirements such as completeness, consistency, and groundness.
Independent Agreement: Do separate models reach the same conclusion?
Supporting Evidence: Is the answer supported by the underlying contract?
Task requirements: Is the answer complete, consistent and appropriate for that specific use case?
The result is validation designed around the work being performed; not a single confidence setting applied to every answer.
When validation finds a problem
Every failed check leads to an action
Verification is only useful if it catches real mistakes and acts on what it finds. Luminance tests its validation system to uncover and address any source of uncertainty.
Test: Evaluate the validation system on difficult cases where the original prediction was already wrong.
Challenge: Check the answer against independent models, supporting evidence or task-specific criteria.
Recover: Use the critic’s feedback to retry the answer with the identified gap in hand.
Escalate: When trust still cannot be earned, clearly flag the answer for review.
When validation finds a problem
Every failed check leads to an action
Verification is only useful if it catches real mistakes and acts on what it finds. Luminance tests its validation system to uncover and address any source of uncertainty.
Test: Evaluate the validation system on difficult cases where the original prediction was already wrong.
Challenge: Check the answer against independent models, supporting evidence or task-specific criteria.
Recover: Use the critic’s feedback to retry the answer with the identified gap in hand.
Escalate: When trust still cannot be earned, clearly flag the answer for review.
When validation finds a problem
Every failed check leads to an action
Verification is only useful if it catches real mistakes and acts on what it finds. Luminance tests its validation system to uncover and address any source of uncertainty.
Test: Evaluate the validation system on difficult cases where the original prediction was already wrong.
Challenge: Check the answer against independent models, supporting evidence or task-specific criteria.
Recover: Use the critic’s feedback to retry the answer with the identified gap in hand.
Escalate: When trust still cannot be earned, clearly flag the answer for review.
Explore the architecture
From the right context to the right answer.
Document Search Engine
Pre-computed legal knowledge and Context Graphs connect contracts, amendments and related documents, enabling users and AI agents to retrieve the right evidence faster and analyse transactions in context.
Model-Opinionated
A multi-model architecture routes each task to the intelligence best suited to its accuracy, speed, reliability and cost requirements, without dependence on a single model or provider.
Luna Crescent
Luminance’s proprietary LLM is purpose-built for specialist legal tasks, improving the speed and economics of high-volume contract extraction, classification and analysis.
Explore the architecture
From the right context to the right answer.
Document Search Engine / Context Graphs
Pre-computed legal knowledge and Context Graphs connect contracts, amendments and related documents, enabling users and AI agents to retrieve the right evidence faster and analyse transactions in context.
Model-Opinionated
A multi-model architecture routes each task to the intelligence best suited to its accuracy, speed, reliability and cost requirements, without dependence on a single model or provider.
Luna Crescent
Luminance’s proprietary LLM is purpose-built for specialist legal tasks, improving the speed and economics of high-volume contract extraction, classification and analysis.
Explore the architecture
From the right context to the right answer.
Document Search Engine / Context Graphs
Pre-computed legal knowledge and Context Graphs connect contracts, amendments and related documents, enabling users and AI agents to retrieve the right evidence faster and analyse transactions in context.
Model-Opinionated
A multi-model architecture routes each task to the intelligence best suited to its accuracy, speed, reliability and cost requirements, without dependence on a single model or provider.
Luna Crescent
Luminance’s proprietary LLM is purpose-built for specialist legal tasks, improving the speed and economics of high-volume contract extraction, classification and analysis.
GET STARTED
Get more from your contracts
See how Luminance can help your teams negotiate smarter, surface what matters, and keep contracts moving across the enterprise.
GET STARTED
Get more from your contracts
See how Luminance can help your teams negotiate smarter, surface what matters, and keep contracts moving across the enterprise.
GET STARTED
Get more from your contracts
See how Luminance can help your teams negotiate smarter, surface what matters, and keep contracts moving across the enterprise.
© 2026 Luminance. All rights reserved.