Evidence Grading for OAuth Credential Exposure and Containment: A Retrospective Study of GENESIS Incident 38

Closing an OAuth credential exposure requires several distinct judgments: whether credentials entered source control, whether a refresh path still works, whether application authentication remains available, and whether repository containment has removed exposed values from a specified tree. This retrospective study examines GENESIS incident #38 using local Git objects, a contemporaneous governance record, verifier source, and current provider documentation. The record reports application deletion attested by the founder, rejection of the old refresh path with ACCESS_DENIED, and closure as PROVIDER-NEUTRALIZED and CONTAINMENT-VERIFIED. Repository inspection corroborates removal of three exporter literals and removal of a tracked token file at the containment revision. Direct rejection of the historical access token remains NOT-COVERED. The study also identifies limitations in the historical verifier: its decision logic can accept simulated transport failures as rejection, and its repository scan passes search values in subprocess arguments. These findings constrain the assurance attributable to its terminal verdict without establishing that the recorded provider response was false. The contribution is a claim-specific evidence taxonomy and a bounded verification method that keeps source inspection, reported observation, operator attestation, inference, and non-coverage separate. The case supports a disciplined account of incident closure, not a proof of universal credential invalidation or historical erasure. This is a preprint and has not undergone peer review. The supplement contains sanitized evidence summaries and editable manuscript files; it does not contain raw provider records or the private repository. Author: Christopher Musyoki, GENESIS. This work was self-funded. The author reports no known competing financial or personal interests and discloses affiliation with GENESIS, whose incident is examined. AI assistance is described in the manuscript.