A mid-sized legal technology company was building a contract review product for corporate legal teams. Their core workflow required extracting specific clauses from commercial agreements, flagging non-standard terms, and surfacing obligation summaries for in-house counsel.
Before Crucible, they were running a combination of rule-based extraction and a general-purpose language model. The rule-based system was brittle — it broke on any formatting variation outside its training distribution. The language model produced inconsistent outputs and required significant human review before results could be trusted.
What Changed
They integrated Crucible using the completions endpoint in Deep Mode with a structured JSON schema defining the fields they needed: party names, effective date, termination clauses, limitation of liability, and governing law.
Crucible's grounding verification pass was the key capability. It eliminated the class of errors where the previous model would hallucinate clause language that did not exist in the source document. Every extracted value was traceable back to the source text.
Results
Across a test set of 400 commercial agreements, the integration produced:
60% reduction in time spent per document review
91% extraction precision on standard clause types
Near-zero hallucination rate on clause language
The legal team now uses the Crucible output as a first-pass review, escalating only documents where the reasoning trace flagged ambiguity. Human review time dropped from an average of 45 minutes per contract to 18 minutes.
A mid-sized legal technology company was building a contract review product for corporate legal teams. Their core workflow required extracting specific clauses from commercial agreements, flagging non-standard terms, and surfacing obligation summaries for in-house counsel.
Before Crucible, they were running a combination of rule-based extraction and a general-purpose language model. The rule-based system was brittle — it broke on any formatting variation outside its training distribution. The language model produced inconsistent outputs and required significant human review before results could be trusted.
What Changed
They integrated Crucible using the completions endpoint in Deep Mode with a structured JSON schema defining the fields they needed: party names, effective date, termination clauses, limitation of liability, and governing law.
Crucible's grounding verification pass was the key capability. It eliminated the class of errors where the previous model would hallucinate clause language that did not exist in the source document. Every extracted value was traceable back to the source text.
Results
Across a test set of 400 commercial agreements, the integration produced:
60% reduction in time spent per document review
91% extraction precision on standard clause types
Near-zero hallucination rate on clause language
The legal team now uses the Crucible output as a first-pass review, escalating only documents where the reasoning trace flagged ambiguity. Human review time dropped from an average of 45 minutes per contract to 18 minutes.
A mid-sized legal technology company was building a contract review product for corporate legal teams. Their core workflow required extracting specific clauses from commercial agreements, flagging non-standard terms, and surfacing obligation summaries for in-house counsel.
Before Crucible, they were running a combination of rule-based extraction and a general-purpose language model. The rule-based system was brittle — it broke on any formatting variation outside its training distribution. The language model produced inconsistent outputs and required significant human review before results could be trusted.
What Changed
They integrated Crucible using the completions endpoint in Deep Mode with a structured JSON schema defining the fields they needed: party names, effective date, termination clauses, limitation of liability, and governing law.
Crucible's grounding verification pass was the key capability. It eliminated the class of errors where the previous model would hallucinate clause language that did not exist in the source document. Every extracted value was traceable back to the source text.
Results
Across a test set of 400 commercial agreements, the integration produced:
60% reduction in time spent per document review
91% extraction precision on standard clause types
Near-zero hallucination rate on clause language
The legal team now uses the Crucible output as a first-pass review, escalating only documents where the reasoning trace flagged ambiguity. Human review time dropped from an average of 45 minutes per contract to 18 minutes.