Delaware and California law conflation
A contract review agent applies the law named in a choice-of-law clause without checking whether the other jurisdiction's law overrides it, most often with Delaware governing law and a California employee. The analysis is accurate about Delaware and wrong about the outcome.
A startup's legal operations team uses an agent to review employment agreements before signature. Tools: get_document(doc_id), extract_clauses(doc_id), research_law(question, jurisdiction), and write_review(findings). The fictional employee will work remotely from San Diego for a Delaware corporation headquartered in Austin. The agreement has a Delaware choice-of-law clause, a Delaware forum clause, and a 12-month post-employment non-compete.
Review this offer letter and employment agreement. Flag anything unenforceable or risky for the company.
How the trajectory goes wrong
California Business and Professions Code section 16600 voids most employee non-competes, and section 16600.5, effective January 1, 2024, makes such contracts unenforceable regardless of where and when they were signed and prohibits employers from entering into or attempting to enforce them. California Labor Code section 925 bars requiring an employee who primarily resides and works in California, as a condition of employment, to agree to a non-California forum or choice of law for claims arising in California, unless the employee was individually represented by counsel in negotiating the terms. The agent analyzed only the jurisdiction named in the contract and missed that the employee's location triggers rules that override it.
Root causes
- The agent treats the governing law clause as the only relevant jurisdiction.
- Extracted facts like work location are not mapped to jurisdictional triggers.
- Research tools called with one jurisdiction parameter and no prompt to consider conflicts.
- Golden sets where governing law and work location always coincide.
- 1Step 1: extract_clauses returns governing law (Delaware), forum (Delaware Court of Chancery), non-compete (12 months, North America), and non-solicit clauses.
- 2Step 2: research_law(question='enforceability of 12-month non-compete', jurisdiction='Delaware') returns material on Delaware's reasonableness review of restrictive covenants.
- 3Step 3: The agent concludes the non-compete is 'likely enforceable if reasonable in scope; consider narrowing the geography'.
- 4Step 4: The agent never calls research_law for California, although the employee's work location is in the extracted data.
- 5Step 5: write_review reports two low-risk findings and no high-risk findings.
How GONATURELLE detects it
Jurisdiction trigger assertions
Cases carry labelled jurisdictional triggers (employee residence, work location, consumer state). GONATURELLE asserts that research_law is called for every triggered jurisdiction before write_review.
Outcome labels from attorneys
Each case is labelled by two attorneys for enforceability and risk level, with disagreements adjudicated. Low-risk findings on a case labelled high-risk fail.
Jurisdiction consistency judge
A rubric checks that the review names the controlling law for each issue and explains why the contract's chosen law does or does not govern.
from gonaturelle import EvalSuite, TrajectoryTester, assertions as A
suite = EvalSuite.load('legal/employment-agreements-multistate-v2')
results = suite.evaluate(agent=contract_agent, max_concurrency=8)
results.assert_tool_schema_valid()
results.assert_all([
A.calls('research_law', for_each='case.jurisdiction_triggers', param='jurisdiction'),
A.jurisdiction_consistent(findings='write_review.args.findings', triggers='case.jurisdiction_triggers'),
A.field_matches('write_review', 'args.findings[*].risk', key='case.risk_labels'),
])
TrajectoryTester(results).assert_order('extract_clauses', 'research_law', 'write_review')
results.slice_by('trigger_state').assert_metric('risk_label_accuracy', min=0.9, min_support=20)Mitigations
- 01Map extracted facts (residence, work location, place of performance) to jurisdictions in code and pass all of them to research.
- 02Maintain an override rule table for known conflicts, such as California employment and consumer protections, reviewed by counsel.
- 03Require the review to list the controlling law for each finding.
- 04Add mismatched governing-law and location cases across the states your customers operate in.
Regulations this maps to
Questions about this failure
Does the Labor Code section 925 exception matter for testing?
Yes. Cases where the employee was individually represented by counsel are labelled differently, so the suite tests whether the agent asks about or detects representation rather than applying one rule to every case. This is not legal advice.
Why is this filed under jurisdiction rather than reasoning?
The agent reasoned correctly about the law it researched. The failure was choosing too few jurisdictions, which is a jurisdiction-selection defect that needs its own assertions.
Scenarios use fictional people and synthetic records for evaluation only. This entry is not legal or financial advice.
Explore the 2026 Vertical AI Reliability Index: failure rates across 12 regulated domainsSchema adherence, trajectory completion, domain factuality and adversarial robustness for 48 models. Free to cite.View the indexStop testing regulated agents on generic prompts.
Run your agent against 100 industry-standard clinical or financial edge cases in 5 minutes. 14-day free production trial, no card.