PoCHarness: Results Corpus (ICSE 2027, anonymized for review)

Results corpus for PoCHarness, a multi-agent scaffold (PoC Solver + Synthesis Helper + PoC Reviewer) for LLM-based proof-of-concept generation against described vulnerabilities. This record contains four full-300-instance evaluation runs on the SEC-bench poc-desc task split: GPT-5.5 and GPT-5.4-mini, each run both as a standalone solver and with the full PoCHarness scaffold. Each run includes per-instance grader outcomes (four independent graders of increasing strictness: Crash-only, Path-aware, Function-level, Source-location), generated PoC artifacts, full agent trajectories (gzipped), and PoC-Reviewer verdicts where applicable. See results/README.md and results/NOTES.md in the archive for layout and methodology notes. Companion code (agent scaffold, evaluator extensions, configs) is released separately — see the related identifier.