Designing for the Else: Generating AI Specifications That Handle What Is Not There

Abstract

I can generate a working AI specification in about five minutes, and it runs across five different AI platforms without being rewritten for each one. This paper describes the method that got me there and how I checked that it was real. The method is a loop. Build a specification for a job, test it, then pull out the parts that were not about that job, strip the job-specific language, and put them where the next build will load them. Over fourteen builds the extraction got smaller each pass until the specification-writing work was nearly gone. What survived extraction was almost entirely else clauses: what to do when an answer is absent, unverifiable, or open to more than one reading. Handed an unspecified case, a model does not stop. It fills the gap fluently, and the filled version is indistinguishable from the real one. The failure is missing instructions rather than wrong ones. The method's weakest step is the extraction, and skipping it feels exactly like the loop converging. Version discipline kept separate from the loop is what let me find out that I had stopped. I checked my own reading of the corpus by giving the build records to a separate model session with the hypothesis withheld. The limits are stated plainly: one author, one domain, and no measured results.