I have been playing around with a thought experiment where agents play human roles and come to a conclusion. I created agents for each jury, and the judge is another agent that can make an informed decision based on what the juries come up with. I fed the movie script into the setup, and each jury uses the Recall tools to refine their belief, and how they came to a conclusion.
I built this using Polign Recall - https://github.com/Polign/recall , langGraph and agent-sandbox.
1. Each jury agent processes the conversation in its own terms.
2. The memory gets folded, and contradiction and change of mind get processed as part of their own belief system.
3. The judge decides how each juror finally concluded their belief and whether they came to be with an informed decision or were coerced into the decision.
--------------------------------------------------------------------------------
Here is the conclusion by the judge (This is AI generated for a reason) -
As the judge, I'd rule: not guilty, and the verdict stands. The prosecution's case was meant to prove guilt beyond reasonable doubt, and each piece of it fails once tested in the jury room:
The knife isn't unique. Juror 8 bought an identical switchblade in the boy's neighborhood, so the "one of a kind" weapon ties no one to the killing.
The old man couldn't have seen the boy run. He had had a stroke and drags one leg. Walked out in the jury room, he couldn't reach his door in the 15 seconds he claimed.
The woman across the street wasn't wearing her glasses. She had marks on her nose from glasses, and she saw the stabbing from bed, at night, through the windows of a passing train. She wouldn't have been wearing them in bed. "I'm going to kill you" couldn't be heard. The train was passing at that moment, and it's loud enough to drown out speech.
The stab wound doesn't fit. The wound angled downward, but someone experienced with a switchblade stabs underhand. A shorter boy stabbing downward at a taller man is unlikely. The boy forgetting the movie proves little. Juror 3, sitting calmly in the jury room, can't remember a film he saw days earlier.
Did the boy kill his father? Neither the film nor the record answers that, and a judge doesn't rule on it. A judge rules on whether the state proved its case. On this record it didn't: every eyewitness account and every piece of physical evidence fails. Not guilty is the definite answer the evidence supports.
--------------------------------------------------------------------------------
The premise behind the memory updates and storage-
A change of mind is a correction: When a later scene changes what a juror believes, Recall keeps the newer belief as current and keeps the older one as what it replaced, with the scene it came from. A juror who moves from guilty to not guilty carries both beliefs, the scene where each began, and the scene where the first one ended. Recall's diff returns exactly what started and stopped being believed in the final act.
Opinions from memory alone: At the end, each juror is asked what they make of the outcome. The model gets that juror's memory from Recall, and not the script. It answers in the juror's voice and cites the scene behind every belief. If a juror was misled and never found out, they still believe it.
A conclusion from separate minds: The last step reads the final pages and every juror's memory, side by side, and names the question the film leaves open. It answers that question with evidence that names the juror and the remembered scene, for example, a late change of mind, or something one juror knows that another does not. Because each memory was kept apart, the conclusion can see where the jurors disagree.