Building an Ace Attorney trial with no script

Why I built it

Courtroom games are puzzles: the testimony is written in advance, and you win by pressing the right line and presenting the right exhibit at the right moment. Say anything else and you lose a life. I wanted to find out if I could build the opposite: ask whatever you think will help your client, the way a lawyer would, and have the court answer what you actually said.

Getting a model to write one courtroom line is easy. The real problem is keeping a trial fair and consistent for a hundred turns, with a judge whose ruling follows the record. So I scoped it to one case, The Gatewater Blackout: an art dealer is shot in Room 704 of the Gatewater Hotel during a blackout, your client Larry Butz left a fingerprint on the balcony rail, and a photographer says she saw him climb down.

Designing the player's side

I kept the series' shape, a private interview with your client and then court, but replaced its fixed menus with five open actions. I also cut the penalty meter and the timer: a weak question costs nothing but a weaker case.

Speak

Type a question, a challenge, an objection or an argument. The court works out who you're talking to and who should answer.

Present evidence

Pick an exhibit and say what you want to ask or argue with it.

Cite

Quote an exact line from the record. Quoting a private conversation is your claim as counsel, not sworn testimony.

Let it continue

Hand the floor to the court: the prosecutor questions the witness, the judge rules or invites closing. You can step in whenever a question is pending.

Close

Argue your case, or let the judge decide on the record as it stands.

Writing a witness with secrets

I wrote the case as data, not as a script. Lotta's file holds four things she won't volunteer: the police showed her Larry's photo before she named him, she left her post for four minutes to move her van off a fire lane, a magazine pays more for a story with a named suspect, and her camera has a 26-minute gap right across the shooting. I never scripted when, or whether, any of it comes out. In the trial below, two came out because of how I questioned her. The other two were never found.

The hard part: keeping a live AI consistent

I split the work so the model only ever does what it's good at, and code holds it to the case:

Fixed

The facts of the crime, the exhibits, and what each character actually knows. My code checks that every turn only refers to evidence and facts that exist.

Written live

Every line of dialogue, what a witness chooses to disclose, the prosecution's arguments, the judge's rulings and the verdict.

Checked

A separate model reviews each proposed turn for continuity before it's committed: an admission can't be forgotten two turns later, and a private conversation can't leak into court unless you bring it up there. A rejected turn is retried without losing your place.

Remembered

The court keeps a ledger of the points before it, each open, disputed, conceded or resolved, with references to the record, so a concession still counts when the next person speaks.

A trial I played

Lines from one of my test trials, in order, numbered from the record. My lines are exactly as I typed them, typos included; everything else was written live. It shows the pieces working together: a client who lies until he's pressed, an exhibit that breaks a witness's story, a citation, a secret coming out, and a ruling that follows the record.

In private

1. 1Phoenixso what happend?
2. 2LarryNick! You gotta believe me, I didn't shoot anybody! I wouldn't even know how to fire a gun without dropping it on my foot!
3. 4Phoenixthen how do you explain your finger prints in the balcony?
4. 5LarryW-Wait, they found my prints out there?! How did they even… I mean, balcony?! What balcony?!
5. 7Phoenixyou're lying to me aren't you? your finger prints are on the balcony, explain so i can help you
6. 9LarryI did step out onto the balcony of Room 704 and touch the railing... but it's not what it looks like!
7. 14LarryLook, I even kept the signed receipt from when he took the package! Here, take it!
8. 15RecordLarry hands over the Delivery Receipt.

In court

1. 24LottaRight after, I saw someone clambering down from Room 704's balcony. It was that fella Larry Butz, plain as day!
2. 29LottaThat fella pulled out his cell phone, and the screen lit him right up in the dark!
3. 39RecordPhoenix presents the Power Co. Report: the whole block, street lights included, was dark from 9:40 to 10:00 PM.
4. 40Phoenixstreet lights were out, how could you be sure it was him?
5. 42FranziskaYou foolishly repeat what we already know, Phoenix Wright; the blackout did not extinguish the battery of a handheld phone.
6. 46LottaI couldn't rightly make out a clear face or his exact build from all the way down there.
7. 47RecordPhoenix cites record #46.
8. 82Phoenixis there anything else that would make you identiy the fact the shooter was larry?
9. 83LottaWell... the police did come show me a photograph of that fella Larry the next mornin'.
10. 84LottaAfter seein' his picture, I just figured he had to be the one, even if I couldn't truly tell from the silhouette alone.
11. 102JudgeA lack of an alibi does not substitute for affirmative proof. The prosecution's remaining link is the undated railing print alongside the 8:30 PM delivery.
12. 112JudgeThe State has failed to prove the defendant's guilt beyond a reasonable doubt, and I hereby find the accused, Larry Butz, not guilty.

From the 113-entry record. I played the first 99 entries; for the last few, my test harness picked up the saved trial, asked the judge for a ruling and let the court finish.

What I learned

Facts in code, words from the model

The model never decides what happened, only what people say about it. Game code rejects any turn that mentions a fact or exhibit the case doesn't have, which is what keeps a live trial fair to the player.

A second model is a cheap editor

Checking each turn for continuity before it's committed stops forgotten admissions and leaked private talk before the player sees them. A rejected turn just runs again.

Watch the thinking budget

Early live runs came back cut off because the model's reasoning used up the output budget. Lowering the reasoning effort and leaving room for the answer fixed it.

Test against the real model too

Mocked tests prove the plumbing. Replaying a saved trial through the live model is what showed whether the judge rules sensibly.

What I'd do next

An original case and cast, more than one witness, places to investigate before the trial, and a turn well under the current seven seconds, using the same write-ahead tricks I used to give Façade's characters a live voice.