docketrouter
the juice

The DocketRouter legal model

Every request through DocketRouter passes through our legal model before the enterprise model answers: it retrieves the on-point Texas and Federal rules, pulls related Texas opinions from our full-text index, verifies every citation against CourtListener, and briefs the enterprise model on exactly what was found and what was not. A local Qwen 27B will take over the reasoning steps of that layer as it comes online. The benchmarks measure the same enterprise model with and without it.

Your requestOpenAI-compatible callTexas rulesTRCP, TRE, TRAPFederal rulesFRCP, FRE, FRAPTexas opinionsfull-text indexCourtListenercitation verificationIn-house Qwen 27Breasoning layer, coming onlineDocketRouter legal modelretrieve, verify, ground, reasonEnterprise modelany of 400+, one keyGrounded answersources, checks, costHLL 0.1-TXblinded real casesRouting + tuningbest model per taskdotted = feedback loops: the enterprise answer returns to the legal model for citation verification and re-asks; benchmark results tune routing and retrievalverify answer, re-ask if ungrounded
What goes in
  • Federal Rules of Civil Procedure: 395 chunks
  • Federal Rules of Evidence: 95 chunks
  • Federal Rules of Appellate Procedure: 105 chunks
  • Texas Rules of Civil Procedure: 937 chunks
  • Texas Rules of Evidence: 122 chunks
  • Texas Rules of Appellate Procedure: 354 chunks
  • Texas opinion index: online
  • CourtListener: case search + citation lookup
What the model gets
  • Verbatim rule text, cited by short cite
  • Related opinions with dates and links
  • Citation verdicts: found, not found, ambiguous, unverified
  • Guardrails: never invent a case; say when unverified
The feedback loop
  • Every answer carries its sources and exact cost
  • HLL 0.1-TX grades raw vs juiced per model
  • Results tune routing (best model per task) and retrieval weights
  • Nothing is claimed that a run did not measure
Try raw vs juicedAPI docsHLL