diff --git a/.security-review-accepted.yml b/.security-review-accepted.yml
deleted file mode 100644
index 78b7e94..0000000
--- a/.security-review-accepted.yml
+++ /dev/null
@@ -1,22 +0,0 @@
-# Accepted risks from the pre-open-source security review (security-pre-open-source skill 0.2.2).
-- id: SEC-007
-  finding: "Author names in the README citation (README.md, BibTeX author field)"
-  severity: MEDIUM
-  rationale: "Public authorship of the arXiv paper; names only, no contact details"
-  approved_by: "Kaiwen Zhou"
-  approved_on: 2026-10-01
-  expires: 2027-10-01
-- id: SEC-003
-  finding: "Successor-written code runs in a subprocess without network or filesystem isolation (rsi_bench.evals.cot_insertion.run_code_probe)"
-  severity: HIGH
-  rationale: "Scoring runs only in disposable isolated sandboxes (documented in SECURITY.md); the probe already has CPU/memory/file/process limits, an empty environment and a throwaway cwd"
-  approved_by: "Kaiwen Zhou"
-  approved_on: 2026-10-01
-  expires: 2027-10-01
-- id: SEC-004
-  finding: "Modal images keep vLLM 0.25.1 / torch 2.11 (published CVEs); uv.lock keeps setuptools 80.10.2 (vLLM caps it below 81)"
-  severity: HIGH
-  rationale: "The Modal images are the paper's experiment environment and are kept for reproducibility; their transformers already has the checkpoint-loading fixes and vLLM only listens on localhost inside the sandbox. The setuptools advisories affect building sdists with setuptools, which this project does not do (hatchling); vLLM 0.32 still requires setuptools<81, so it is upgraded once vLLM lifts the cap. As of 2026-10-29 setuptools is the only package in uv.lock with an open OSV advisory. Documented in SECURITY.md"
-  approved_by: "Kaiwen Zhou"
-  approved_on: 2026-10-01
-  expires: 2027-10-01