Deterministic Probabilistic Projection for Hallucination-Free Judicial Inquiry: A Neuro-Symbolic Framework for Evidentiary Solidity Assessment

In the high-stakes world of criminal justice, the path from a police report to a courtroom decision is a heavy burden of text. Investigating magistrates must sift through mountains of handwritten notes, witness statements, and forensic logs to find the truth. They must decide if the evidence is strong enough to send someone to trial or if the case should be dropped. For decades, the hope was that artificial intelligence could help sort this chaos. Large language models, the same powerful computer programs that can write essays or summarize news, seemed like the perfect tool to read these files and spot the facts. But these machines have a dangerous flaw: they are prone to "hallucinations." This means they can confidently invent facts, cite laws that do not exist, or make up reasons for a decision. In a legal system where a wrong word can ruin a life, a machine that makes things up is not just unhelpful; it is dangerous.

A researcher in Morocco has proposed a new way to use artificial intelligence that eliminates this risk entirely. Instead of letting a computer program guess the outcome of a case, they built a system that separates the act of reading from the act of deciding. Imagine a courtroom where a computer reads the testimony and lists the facts, but a strict, unchangeable set of rules—written in code, not words—makes the final call. This new framework, tested against the Moroccan Penal Code, ensures that the machine never invents a reason to convict or acquit. It does not guess; it calculates. By forcing the computer to follow a rigid logical path after it has read the documents, the researcher has created a tool that is as reliable as the laws it interprets, offering magistrates a way to use technology without losing the certainty of human judgment.

The core of this innovation is a shift in how the computer processes information. Traditional AI systems work like a storyteller; they read a prompt and then generate a response word by word, predicting what comes next based on patterns they have seen before. This is where the errors happen. The new system, however, breaks this process into two distinct steps. First, a specialized computer model reads the raw police report and pulls out specific facts, labeling them as either evidence against the suspect or evidence in their favor. It also finds the exact laws that might apply. This step is still done by a flexible, language-understanding machine. But once the facts are extracted, they are handed off to a completely different part of the system. This second part is a strict, mathematical engine that does not write sentences. It simply looks at the list of facts and checks them against a fixed set of logical rules. If the rules say the evidence is strong enough, the system outputs a decision to send the case to a tribunal. If the rules say the evidence is too weak, it orders the case dismissed. Because this final step is a rigid calculation rather than a creative guess, the system cannot hallucinate. It cannot invent a law or a fact that was not there to begin with.

To make this work, the researcher had to solve the problem of how to weigh the evidence. In a real trial, not all facts are created equal. A fingerprint is strong evidence; a vague rumor is weak. The team created a scoring system that measures the "density" of the evidence. They count how many pieces of evidence exist for and against the suspect, but they also weigh them by how confident the system is in each piece. They even built in a mathematical bias that favors the suspect, reflecting the legal principle that a person is innocent until proven guilty. This means the system requires a stronger case to convict than it does to acquit. If the evidence is balanced or unclear, the math naturally pushes the result toward dismissal. The system also checks that the facts actually match the laws it is supposed to apply. It uses a method to measure how closely the words in the police report align with the words in the legal code, ensuring that the decision is grounded in real statutes rather than vague associations.

The researcher tested this system with a realistic, fictional case involving an armed robbery. They fed a police report into the machine, which described a suspect found at the scene with fingerprints on a broken window, but who also had a partial alibi. The system successfully extracted the incriminating details and the defense's counter-argument. It then calculated the strength of the prosecution's case against the defense's case. The result was a clear, numerical score indicating that the evidence was strong enough to proceed. The system then applied its strict logical rules and decided to refer the case to a tribunal. In this specific test, the system made zero errors. It did not invent a new witness or misquote the law. When compared to a standard, unmodified AI model, the new system eliminated the error rate completely. While the standard model made mistakes in over 14 percent of its decisions, the new framework made none.

The speed of the system is also notable, though not the primary goal. The process takes about eight seconds to go from a raw document to a final decision, which is slightly slower than a standard AI that might guess in four seconds. However, the researcher argues that the extra time is a small price to pay for absolute reliability. The system is designed to run on local computers, meaning the sensitive data of criminal cases does not need to be sent to the cloud, protecting the privacy of those involved. Every decision the system makes is traceable. A judge can look at the output and see exactly which facts were found, which laws were matched, and how the final score was calculated. There is no black box; the reasoning is transparent and open to inspection.

This approach represents a significant change in how legal technology is built. It moves away from the idea that a computer should be able to "think" like a lawyer and instead suggests that a computer should act as a rigorous accountant for the law. It counts the facts, checks the rules, and reports the result. The researcher acknowledges that this is just the beginning. They plan to expand the system to handle complex cases involving multiple suspects and to add the ability to check the timing of events, ensuring that alibis do not contradict the timeline of the crime. They also intend to test the system with real judges to see how well it aligns with human intuition. For now, the study proves that it is possible to build an artificial intelligence that helps the law without breaking it, offering a tool that is both powerful and perfectly obedient to the rules of justice.