# PBS production corpus qualification Owner: compiler/pbs. This maintained runner preserves the frozen 28-case input and writes each run to a new directory. It never rebuilds or writes the runtime provider and does not modify a game workspace or memcard. Build the candidate with the required regression gate, then run the corpus: ```bash ./gradlew :prometeu-compiler:frontends:prometeu-frontend-pbs:test :prometeu-compiler:prometeu-build-pipeline:test :prometeu-lsp:prometeu-lsp-v1:test :prometeu-app:test :prometeu-app:installedKnowledgeTest --console=plain ``` ```bash python3 tools/pbs-production-corpus/run.py --corpus prometeu-compiler/frontends/prometeu-frontend-pbs/src/test/resources/production-regressions/cases.json --distribution prometeu-app/build/install/prometeu-app --runtime-headless /absolute/existing/prometeu-headless --output build/pbs-production-corpus/new-run ``` Each case is compiled in a fresh JVM/project outside the checkout. Checkout mode puts current classes/resources first; installed mode includes only candidate JARs and runs outside the checkout. The runner verifies candidate product class freshness and records Java, SDK/compiler origins, JAR/source/PBX/manifest/runtime hashes, Git/worktree state, diagnostic phases and protocol transcripts. The required corpus produces 168 observations: 28 cases × checkout/installed × optimized production stages, investigation-only unoptimized stages, and the public optimized facade. Unoptimized means omitting OptimizeIRVMPipelineStage and emitting original IRVM; it is not a public compiler flag. The expected runtime matrix is 108 executions: 27 compiled cases × checkout/installed × two optimization configurations, with 104 normal three-frame successes and four deliberately faulting controls. The file-private case must remain a structured source rejection in all six compile configurations. Use --select id1,id2 or --origins checkout/installed for focused investigation. Without those restrictions, exit zero requires the complete expected matrix and classifications. Output directories cannot be reused or target the original AGD-0113 baseline. Runtime processes are terminated and waited for on every path. Logs are local evidence, while summary.json and per-case JSON preserve structured results. The focused-cases.json resource is an additional matrix for contextual overloads, noncommutative receiver/argument order, original global storage, service singleton identity and privacy. It does not replace or alter the frozen corpus. Successful compiler proof does not imply qualification of an installed runtime distribution or a live Studio session. The Studio tests independently preserve MCP/LSP diagnostic contracts. The completed DEC-0088/PLN-0166 result, original baseline and separate post-fix reports are consolidated in [LSN-0106](../../discussion/lessons/DSC-0111-pbs-production-reproductions-and-fixes/LSN-0106-pbs-production-binding-and-lowering-corrections.md).