PROJECT FIVE - MULTI-AGENT AND PRODUCTION HARDENING

RUNNING THE FINISHED LAB Generate all the data and run the supervisor. [START OF BASH CODE] # Create the original five fictional resumes and the job posting python make_sample_data.py # Add resumes 06 through 25 python make_more_data.py # Add the planted prompt-injection resume python make_injection_data.py # Add the planted empty resume python make_empty_data.py # Build the fifteen-item evaluation set python make_eval_set.py # Run the supervisor on a completely fresh batch python supervisor.py --fresh [END OF BASH CODE] The resumes folder now holds twenty-seven files, which are the twenty-five normal resumes plus the injection sample and the empty file. In our own run the last two lines looked like this. finished: 27, failed: none steps: 137, tokens by tier: {'cheap': 14558, 'strong': 4961}, cost: 7.87 lab units (budget 60) Now do the arithmetic that justifies routing. Had every token been charged at the strong rate of 1.0 per thousand, the total would have been about 19.5 lab units. With routing it was about 7.9, roughly sixty percent lower, and the reports came out the same. These units are made up, so the lesson is the method: measure per tier, then compare against the all-strong cost. Run python eval_runner.py as well, and you should still see Pass rate: 15 of 15, which proves the refactor did not break the injection defense or the empty-file fix.