# Actual source rechecks, 2026-10-14

- S2: live OpenAI July audit, Methodology and OpenLibrary-77c16d5. Confirmed the 286 flagged-task denominator, 200/249 follow-up findings and one-versus-two leading spaces in prompt/test. The existing local source includes the same example; the old report omitted it. Added a concrete explanation, not a new defect-rate estimate.
- S6: live official SWE-Lancer repository, Running SWELancer. Confirmed internet-disabled runs are the valid protocol and Linux implements the iptables restriction; Mac with internet enabled is not equivalent. Existing stored source remains usable. Added an environment prerequisite to the procurement plan.
- S9 and S10: live Scale V2 page and repository README, update header and What changed from v1. Confirmed 69 contradictory-instruction tasks versus 529 rewritten statements, 214 test patches and 38 reference patches. These describe different changes; no disjointness or one-to-one mapping to the July 249 findings is inferred. The stored source observations already contained these figures; the old report did not explain the distinction.
- S10: The locked protocol and The v1 pipeline. Confirmed agent-only model endpoint access, normal setup/verifier access, clean-image regrading and different V1/V2 tooling. Retain V2 as a candidate-screening source with this actual protocol, not as production productivity evidence.
- New S11: release gate. Confirmed three infrastructure failures per control arm followed by successful reruns; separate final correctness controls from first-pass reliability and cost. No benchmark was executed here.

The S10 page was opened twice because the first combined tool result was truncated in the conversation. That second open repeated source retrieval and provided no new source. Later checks used local retained text and focused locators instead. Six unique original sources S1/S3/S4/S5/S7/S8 did not need live refetching for these changes; their bytes and identities were retained.