How Kingsfield earns your trust
You sign the brief, so when a citation is wrong the court sanctions you, not the AI vendor that wrote it. A checker is only worth trusting if you know how often it is wrong, and if it admits when it cannot be sure. Here is how we test that.
We run Kingsfield against citation benchmarks built by outside researchers to break tools like ours, and we report two numbers, never one: how accurate the check is, and how much of the benchmark our corpus covers. When testing finds a flaw, we fix it, prove the fix, and wire the proof into our own test run so it cannot silently come back. Kingsfield rules on citations, not on the merits of your argument.
We test against benchmarks we did not build
Anyone can pass their own demo. So we run Kingsfield against citation benchmarks built by outside researchers to break tools like ours. One of them is LegalCiteBench, a public academic dataset. Its Category 3 hands the judge a legal paragraph whose citation is either genuine or altered in one controlled way: a wrong page, a wrong volume, a wrong reporter. The judge has to catch the altered ones and pass the clean ones.
The 95% and 4.3% are measured on the ~80% of the benchmark our corpus covers. The rest cite Westlaw or LEXIS vendor numbers, not official reporters, that no case index resolves by citation, so we say nothing about them rather than pretend we checked them.
The 4.3% is the number that matters most. For a tool whose whole job is accountability, flagging a real citation as fake is the worst thing it can do, because it sends you back to re-checking by hand. 4.3% is low enough to actually rely on the flag.
The benchmark caught a real flaw in our own judge. We fixed it.
The run did not just produce a score. It found a genuine defect in our code.
Court reporters are written two ways. The official record writes 110 So. 3d 399 with a space; the tools lawyers use every day write 110 So.3d 399 without one. Our judge matched them literally, so a real, correctly cited case, Tiara Condominium Ass'n v. Marsh & McLennan, 110 So.3d 399 (Fla. 2013), came back as "not found, possible hallucination." A genuine citation, condemned as fake, by the exact kind of tool that is supposed to prevent that.
It had stayed hidden because the reporters written without a space worked fine, so the failures looked like gaps in our library rather than a bug.
We fixed it the same day. We proved the fix: Tiara now resolves, and the reporters that had failed recovered to near 100%. And we locked it, with a regression test wired into our own test run, so if this defect ever comes back it shows up in testing, not in front of you.
When it cannot be sure, it says so
The fix did not swap a wrong rejection for a wrong acceptance. Tiara now returns inconclusive: the case is real, but the exact page cited cannot be confirmed against what the corpus currently holds. Kingsfield says "I cannot confirm this" rather than guess in either direction. That honesty is what makes the signed record worth handing to a court.
What this adds up to
- Independent testing, not self-grading. Outside benchmarks, scored honestly.
- Two numbers, never one. How accurate the check is, and how much it covers, reported side by side. A single headline figure would hide the second.
- A process, not a one-time claim. When testing finds a flaw, we fix it, prove the fix, and wire the proof in so it cannot silently return. A risk officer can inspect that record.
- It checks citations, not merits. Kingsfield tells you whether a citation holds up. Whether your argument wins is your judgment.