Chain Consistency Checks

Deterministic fidelity verification: ensuring the final synthesis honors all intermediate conclusions of the reasoning chain.

Verified Fidelity

Language models often generate fluent text that contradicts their own prior reasoning—concluding that a party prevails after establishing facts that support the opposition. LegalChain addresses this through five deterministic checks that require no additional LLM inference.

ID	TARGET	SOURCE	HALLUCINATION CAUGHT
CC1	Rule Section	S1 Citation	Missing or incorrect target case reference.
CC2	Rule Section	S3 Status	Citing overturned law as good law (or vice-versa).
CC3	Conclusion	S4 Disposition	Affirmed/Reversed mismatch relative to S4 extraction.
CC4	Conclusion	S4 Winner	Incorrect party assertion based on extracted data.
CC5	Application	S5 Relationship	Fabricating relationship logic when S5 was skipped.

Context-Adaptive Scoring

Checks like CC5 are adaptive. The system adjusts its expectations based on what the model actually had available. If S5 (Relationship Analysis) failed due to coverage gaps, the model is not penalized for omitting it—but it is penalized if it "hallucinates" a relationship that it did not actually analyze.

Linguistic Logic

Assertion vs. Quotation

The system uses structural pattern matching to distinguish between the model's own conclusions and text it is merely quoting from a party's argument.

Assertion (Checked)

"The court correctly reversed the lower ruling..."

Quotation (Excluded)

"Petitioner argued that the and court should have affirmed..."

Regex-based boundary detection prevents false positive contradictions.

Scaling Quality Assurance

By verifying internal consistency algorithmically, LegalChain achieves reproducible quality assurance that scales. A synthesis that claims "the Court affirmed" when the model’s own S4 extraction found "reversed" fails CC3, regardless of how persuasively the text is written.