Foundational Concepts
This method helps a computer combine answers from several judges. A judge can be an AI model, a rule, or a person. Each judge looks at the same case and gives an answer. The method does not force one final score too early. The method keeps three things clear. One part shows support for an answer. One part shows support against an answer. One part shows the part that is still unknown.
The rest of this brief gives the history, the method, the 2026 gap, and a first project plan.
A home health agency sends one claim for 4,800 dollars.
The claim lists twelve nurse visits in seven days.
A rules engine checks dates, codes, and visit counts.
The rules engine assigns 0.55 to Fraud, 0.05 to Not Fraud, and 0.40 to ignorance.
The notes describe nine visits and omit three billed visits.
The language model assigns 0.65 to Fraud, 0.10 to Not Fraud, and 0.25 to ignorance.
Fusion of the first two judges gives 0.83 for Fraud and 0.06 for Not Fraud.
Ignorance falls to 0.11 and conflict K stands at 0.09.
A history model then scores the mixed five-year record of the agency.
The history model assigns 0.35 to Fraud, 0.30 to Not Fraud, and 0.35 to ignorance.
The second fusion step raises K to 0.27.
The fused result gives 0.85 for Fraud, 0.10 for Not Fraud, and 0.05 for ignorance.
Belief for Fraud stands at 0.85 and plausibility stands at 0.90.
The team holds the payment and opens an investigation.
The fused record shows each judge vote and each conflict value to the investigator.
A support agent drafts a refund action for one customer.
The draft sends 480 dollars to the customer.
The order total in the account stands at 240 dollars.
A policy checker reads the draft against the refund rules.
The policy checker assigns 0.70 to Safe, 0.10 to Unsafe, and 0.20 to ignorance.
A fact checker compares every figure in the draft with the account.
The fact checker finds the doubled refund figure.
The fact checker assigns 0.25 to Safe, 0.55 to Unsafe, and 0.20 to ignorance.
Fusion of the two checkers gives conflict K at 0.41.
The gate rule stops any action with K at 0.40 or higher.
The system blocks the refund and routes the draft to a human.
The human finds a duplicate line in the draft and corrects the refund to 240 dollars.
A clinic group scores the security file of one software vendor.
A scanner checks 212 control answers in the file.
The scanner assigns 0.60 to Pass, 0.10 to Fail, and 0.30 to ignorance.
A document reader checks the policy documents behind the answers.
The reader assigns 0.50 to Pass, 0.05 to Fail, and 0.45 to ignorance.
Fusion of the first two tools gives K at 0.08.
One past audit shows an expired certificate.
The history check assigns 0.40 to Pass, 0.20 to Fail, and 0.40 to ignorance.
The fused result gives 0.84 for Pass, 0.09 for Fail, and 0.07 for ignorance.
Belief for Pass stands at 0.84 and plausibility stands at 0.91.
The team approves the vendor for a standard contract.
The team logs the expired certificate as a watch item for the next audit.
A delivery robot carries parts between two warehouse zones.
A camera, a lidar unit, and an ultrasonic sensor each assign mass values.
The frame of the task holds two answers: Path Clear and Path Blocked.
Band one covers belief for Path Clear at 0.90 or higher with K under 0.30.
The robot proceeds at full speed in band one.
Band two covers K at 0.30 or higher at any fusion step.
The robot slows, stops, and scans again in band two.
Band three covers belief for Path Blocked at 0.60 or higher.
The robot stops and calls a human operator in band three.
Sunlight through a dock door strikes the camera and the camera reads a dark shape across the lower frame.
The camera assigns 0.15 to Path Clear, 0.65 to Path Blocked, and 0.20 to ignorance.
The lidar assigns 0.75 to Path Clear, 0.05 to Path Blocked, and 0.20 to ignorance.
Fusion of camera and lidar gives conflict K at 0.50.
The pair result gives 0.58 for Path Clear and 0.34 for Path Blocked.
The ultrasonic sensor then reads the doorway space.
The ultrasonic sensor assigns 0.55 to Path Clear, 0.10 to Path Blocked, and 0.35 to ignorance.
The full fusion gives 0.75 for Path Clear, 0.21 for Path Blocked, and 0.04 for ignorance.
Belief for Path Clear stands at 0.75 after the full fusion.
The K value of 0.50 from the first pair still controls the behaviour band.
Band two applies. The robot slows, stops one metre before the doorway, and scans again.
The robot stands half a metre closer after the stop.
The glare angle changes and the camera separates shadow from frame.
The camera assigns 0.65 to Path Clear, 0.10 to Path Blocked, and 0.25 to ignorance.
The lidar assigns 0.70 to Path Clear, 0.05 to Path Blocked, and 0.25 to ignorance.
Fusion of the pair gives 0.88 for Path Clear and K at 0.10.
The ultrasonic sensor assigns 0.60 to Path Clear, 0.05 to Path Blocked, and 0.35 to ignorance.
The full fusion gives 0.95 for Path Clear, 0.02 for Path Blocked, and 0.03 for ignorance.
Band one applies. The robot proceeds through the doorway at full speed.
A pallet jack stands at the side of the doorway, clear of the path.
The robot logs both scans, all mass values, and both K values in the trip record.
Conflict changes the behaviour of the robot in direct steps.
Goal. The desired end state is one fusion service in daily use for typed judge panels. Any caller sends one case and receives one fused record. The record holds a belief interval, an ignorance value, a conflict value, and one decision word.
Proof target. The project proves one claim on labeled fraud dossiers. Dempster-Shafer fusion gives better calibrated decisions than the best single judge. The same fusion also beats plain score averaging on act decision errors at the same act rate.
Product at the end. The product is the DST Fusion Harness, version 1. The harness contains judge adapters, the fusion engine, the decision policy, and a full JSON audit trail. First users are judge panels in fraud screening and AI review work.