Skip animation, jump to the honest-latency summary
STUDYCASE

Studycase · the whole chain, animated

From a courtroom to a card. Every link, in flow.

Sound goes in one end; a verified, speakable case card comes out the other. Watch the data travel each stage: audio, a transcript that locks and never walks backward, one detected moment, retrieval across a real corpus, the citation gate, the card, and finally your own notes. This is the moving-picture companion to how a session works.

An explainer of the pipeline, not a sign-up
Stage 1 / 7 · Court audio
studycase · the chain LIVE

Stage 1 · in

It starts with sound

A live hearing: a class moot, a booked recording, a court stream. The audio enters, and transcription begins on-device: on a booked private Mac mini the sound never leaves the room.

Live audioCaptureOn-device ASR
0.33real-time factor on the dev M4, faster than real timeASR-18m
0audio leaves a booked private sessionMACMINI-SCHEDULING

Stage 2 · the stable prefix

The transcript locks, and never walks backward

Words arrive fast and ghosted in gray. The instant a prefix is stable it locks to full ink, and a locked word is never rewritten under your eyes. The stream may only grow, never retract.

The Court Counsel, he was pulling his arm away at the eighteen-second mark … wasn't that resistance
Locked: final, immutable Partial: still ghosted, may change
0retractions across 1,364 words / 678 churned partialsASR-6

Stage 3 · the trigger

Most of the hearing is quiet. One moment isn't.

Plain detectors watch the stream scroll past, and stay silent. When the crux lands (a judge's pointed question, an objection, a hostile hypothetical), the trigger gate fires once.

…and the officers had detained him for roughly a minute…
…the bodycam picks up at the second cruiser's arrival…
"…wasn't that resistance that justified the takedown?"
Gate fired · judge question

The economics, honestly: we don't run the AI on every second of audio. That's how the incumbents keep a flat price. One analysis fires per detected moment, not per second.

Stage 4 · retrieval

The moment fans out across real law

The query goes two ways at once (meaning and words), then the two rankings fuse. Candidate opinions surface from a corpus of real, published cases. No language model has spoken yet.

query · "resistance justifying a takedown: reasonableness"
Vector search
arctic-embed-s · 384-dim · nearest meaning
Keyword search
BM25 / FTS5 · the exact terms
↓ reciprocal-rank fusion ↓
Graham v. Connor Tennessee v. Garner Kingsley v. Hendrickson Plumhoff v. Rickard County of L.A. v. Mendez
100,000 opinions 1.31M passages 2.03M citation links
Real corpus counts, measured at scalePERF-6 · caselaw-100k

Stage 5 · verify or vanish

The citation gate: cite a real case, or don't exist

Every candidate must resolve against the verified index and every quote must match word-for-word. What passes glides through. What can't be verified is discarded. Never shown, never softened into a maybe.

Graham v. Connor, 490 U.S. 386: resolves in the indexCitation and pinpoint verified against the retrieved corpus. Allowed to speak.
Tennessee v. Garner, 471 U.S. 1: resolves in the indexVerified. Held as supporting authority.
"See Halloway v. State, 612 U.S. 44…"No such opinion in the verified index → discarded. A suggestion may be authority-free; it may never be authority-false.
0unverified citations may escape the gateGENIUS-LINES §2.3

Stage 6 · the deliverable

The card lands: real case, real pinpoint, why now

A skeleton card paints almost instantly because it carries no model output, just a verified case, its citation, and the why-now. The speakable argument line then streams into that same card without moving it.

Binding authority · verified
Graham v. Connor
490 U.S. 386 (1989) · pin cite at 396

Why now: reasonableness is judged from the on-scene perspective. The judge's "resistance" framing invites exactly this holding. Lead with the fact-bound test.

54msskeleton card, measured end-to-endGW-32b

Stage 7 · out

And you keep thinking

The chain ends where the learning lives: your notebook. Verbatim capture hurts learning, so your words stay primary in full ink; the coach's work sits behind them as a muted substrate you fold in or ignore. Two inks, never blurred.

Lead with Graham's fact-bound test. Hammer: prone + cuffed + one ambiguous reflex ≠ resistance. Have the 40-second reply ready.

Graham v. Connor, 490 U.S. 386: objective reasonableness from the on-scene perspective; three factors (severity, threat, resistance). Risk: opponent may distinguish on duration.

Your words: full ink, yours alone AI substrate: muted, behind your work

How fast, honestly

Nobody in this category hits sub-second, and we don't pretend to. Here is the honest clock for the chain above. The first useful thing lands fast, the fully-verified argument lands in seconds.

~54 ms
First paint: the skeleton card (verified case, cite, why-now) with no language model in the loop.
measured · GW-32b
2–5 s
The gate-passed argument line streams into that same card as the analyst finishes. The honest window for a complete, verified completion.
STRATEGY-REALTIME · RESEARCH
Never sub-second
Incumbents advertise 116 ms and measure 4–5 s. We show a thinking state and tell you what's settled. No placeholder lies.
RESEARCH · teardowns