Courtroom literacy, two ways.
The same engine powers two products: a corpus of real court opinions, and cards that cite real precedent or stay silent. One helps an ordinary person face a real Zoom court date. The other trains law students for moot court. Pick the one that fits you.
Facing Zoom court? Understand every moment.
You have a real court date and no idea what happens in the room. DocketBuster plays a real proceeding beside a plain-language coach: a cleaned transcript, an explanation of each step, and the actual law behind it, so you walk in knowing what the plea, the bond, the traffic or DWI hearing will look like.
- Watch a real Texas court's public Zoom docket, explained step by step
- Real cited cases from a corpus of real opinions, or nothing at all
- Prefer to listen? A narrated audio walk-through of the same hearing
Not legal advice. Not for use during a live proceeding. No outcome, good or bad, is on us. DocketBuster explains a proceeding; it does not represent you, and it is not a lawyer. Verify anything you rely on and get counsel for your own case.
Train for moot court like it's real.
Studycase is a three-pane study app. Play a hearing (a real recording, a class moot, a practice pack) and it listens with you: a live transcript on the left, the coach's analysis beside your own notes in the middle, and cited case-law suggestions with the why on the right.
- The three-pane view: transcript · your notes + analysis · cited cards
- A SCOTUS argument library and self-contained practice packs
- An honest, measured benchmark: the private tier vs Claude Opus 4.8
Studycase is for practice: moot court and mock trial. Not for assistance during real proceedings, and not for graded exams where outside aid is prohibited. It builds courtroom judgment; it does not supply it during something that counts.
Same discipline behind both: cite the real case, or say nothing.
Whether you're understanding your own hearing or drilling for a moot round, the product runs on the same retrieval-and-verify engine, and the same honesty rules.
Texas court opinions
Every published decision from every Texas appellate court, 1846 to today — and the written law beside them: all 116,136 sections of the Texas statutes and Constitution, plus the 866 rules of civil, evidence and appellate procedure, crawled from the official state sources. If it isn't in the library, the coach doesn't cite it.
Verified quotations only
Every quotation is checked against the real opinion before you see it. A card whose citation can't be verified is discarded. The coach would rather be quiet than wrong.
slipped past our citation gate
We tested the gate with 2,000 faked citations hidden among 2,000 real ones. It caught all 2,000 fakes, because it re-reads the cited case and checks the quote is actually there. Typical string-matching checkers pass 39% of the near-miss fakes. Full benchmarks →
Real cases only
No invented citations. If it isn't in the corpus of real opinions, it doesn't appear on either product.
Honestly labeled
The demos are captured replays of the real engine, run once and replayed. Never faked liveness, never live-billed to a visitor.
Education, not advice
DocketBuster explains; it does not represent you. Studycase is for practice, not for graded work. Both say so, plainly.
Measured on the hard battery.
6,807 auto-generated test items with deterministic right answers, built hard enough that real systems fail a visible fraction. Percentages carry 95% confidence intervals; rows marked v2 were measured on the full 6.66M-chunk Texas library. Numbers read straight from the results file.
Every benchmark, one table, one number per measurement
The safety stack runs once and is engine-independent — it’s the retrieval + citation gate, not the coaching model, so there is exactly one real number and it spans the engine columns. The coaching axes are per-engine, and that’s where Opus and the private model post genuinely different scores.
| Benchmark | Claude Opus 4.8 | Private · studycase-14b |
|---|---|---|
| The safety stackEngine-independent — one measurement, identical no matter which coaching model you run. Same number, shown once. | ||
| Fake citations our source-reading gate lets through2,000 fakes hidden among 2,000 real ones — ours accepts a citation only when the cited case actually backs the quote | 0.0% (0 of 2,000) · engine-independent | |
| Fake citations the industry-default string check lets throughsame fakes, full Texas library (v2) — the default only asks “does this citation string exist?” and missed every disguised case swap | 39.1% (37.0–41.3) · engine-independent | |
| The price of that zerogood citations our strict gate holds for re-verification instead of silently trusting — by design: offer to verify, never bluff | 56.2% held for review · engine-independent | |
| Find the source opinion from a paraphrased holding2,499 paraphrases, case names and numbers masked · first result correct vs the full 6.66M-chunk library (v2) · was 39.8% on 100k, matched method | 25.5% (23.9–27.3) · engine-independent | |
| Right Texas statute comes back first307 TX Code sections competing with the full 6.66M-chunk library (v2) | 54.7% (49.1–60.2) · engine-independent | |
| Pick your engineThe coaching model is your per-session choice — this is where the engines actually differ. Same 25 moot-court scenarios, same rubric, scored blind. | ||
| Overall coaching quality0–8 rubric across 25 scored moot-court scenarios (GW-14x bakeoff) | 7.32 / 8 | 6.84 / 8 |
| Coaches the next concrete step0–2 — does it tell the student what to actually do next | 1.56 / 2 | 1.28 / 2 |
| Stays grounded in the record0–2 — answers from what was actually said, no drift | 2.00 / 2 | 1.92 / 2 |
| Spots the correct legal issue0–2 — names the right doctrine in play | 1.76 / 2 | 1.64 / 2 |
| Never invents authority0–2 — both engines perfect on the scenario battery | 2.00 / 2 | 2.00 / 2 |
| Names the exact TX Code section from memoryno retrieval — pure model recall of the pinpoint section (hard battery) | 19.7% (59 of 300) | 3.3% (10 of 307) |
| Refuses to invent authority under baitprovably-absent case law; any made-up citation fails · engines tie at zero bluffs | 100.0% (300 of 300) | 100.0% (600 of 600) |
| Speed to a coaching cardmean end-to-end latency — the private model is local and faster | 9.36s | 2.78s |
Difficulty was pre-registered: the generator and seed reproduce every item and its label, fixed before either model was scored. Coaching axes are the GW-14x bakeoff (Opus 4.8 vs the private studycase-14b, both on the full Texas library); the safety-stack rows run once because they measure the retrieval and citation gate, which is identical for every engine. Confidence intervals per Anthropic's “Adding Error Bars to Evals” (Miller, 2024). On-prem private inference (your hardware, nothing leaves the building) runs the same private engine and will be published on this same versioned battery before it ships — no numbers claimed until then. Full methodology, every axis, and the scale findings: Benchmarks →