One brief, every path, worked end to end
The demo ships four sample briefs. We walk two: the Mixed Motion brief, which carries a real cite, a fabricated cite, and an out-of-corpus cite at once, and the Clean Motion brief, where every citation is real. Every image below is a screenshot of the running app. The matter caption is illustrative, not a real case.
The pipeline runs live against real case law
On the left is an AI-drafted motion to suppress. On the right, the system console lights up stage by stage and streams each real CourtListener request with its URL, status, and latency, alongside the support-check reasoning. This is the first citation, Miranda v. Arizona, 384 U.S. 436, resolving against the ground-truth corpus while the support-check reads the opinion text.
A fabricated case, caught with certainty
The brief cites Halstead v. Ferngate Holdings, 823 U.S. 1199. The existence check sends a live GET to CourtListener and gets 200 with 0 results, because U.S. Reports has no volume 823. The verdict is FABRICATED, proven absent rather than guessed. In the same run Brown v. Board of Education, 347 U.S. 483 is a real case that is not in the cached subset, so it is honestly marked OUTSIDE_COVERAGE instead of a false verified. The gate reads NOT_FILING_READY: 1 fabricated, 2 unresolved, 1 auto-cleared.
A real case cleared green, on the opinion's own holding
This is the Clean Motion brief, where every citation is a real, on-point case. When the opinion text establishes the proposition, the support-check confirms it. Miranda v. Arizona, 384 U.S. 436 clears Verified because the majority itself holds that a suspect must be advised of the right to counsel and to remain silent, with a link to the real case on CourtListener. Which specific on-point cites clear green versus route to review can vary run to run, because the support-check is the one non-deterministic step, but an on-point cite like Miranda reliably clears.
It routes to a human rather than bluff a verdict
On the Mixed Motion review, the scoreboard reads 1 Verified, 1 Fabricated, and 2 Need Review. Bell Atlantic Corp. v. Twombly, 550 U.S. 544 is a real case, but the excerpt does not squarely establish the proposition it is offered for, so the support-check marks it Needs Review with its reasoning rather than confidently grading it. That conservative abstention is the design: a confident red Unsupported verdict is rare on purpose, because the honest move on an ambiguous opinion is to hand it to a human.
Even a clean brief is held until every cite resolves
This is the certificate for that same Clean Motion brief. Miranda cleared Verified, but Terry v. Ohio and Strickland v. Washington were routed to review, so the gate holds at Not Filing-Ready and the seal reads Hold for Review, even though not a single citation is fabricated. The certificate lists each cite with its verdict and support-check result and certifies that every one was checked for existence against CourtListener, that fabricated cites are identified to a certainty, and that citations requiring contextual judgment are routed to a human. It exports as a signed JSON record you attach to the filing. A clean-looking brief is not the same as a filing-ready one.
Grounded in an independent authority
The corpus is 20 real decisions cached from CourtListener, the public U.S. case-law database, including Miranda, Twombly, Iqbal, Terry, Strickland, International Shoe, Erie, Baker v. Carr, Chevron, and Daubert. Existence and fabrication checks also query CourtListener live at request time, which is why fabrication-catching is authoritative rather than limited to the subset. The 20-case corpus is a cached subset, so support adjudication is demonstrated at small N on real cases, and that scope is stated plainly rather than dressed up as universal accuracy.
