Restore MiMo reward-hacking audit and completed residue follow-up

mimo_v1/README.md

@@ -34,6 +34,26 @@ Intended benchmark transfer: **SWE-bench Verified, Terminal-Bench 2.0 and OpenTh
 
 Includes the full validation run and reruns of repaired tasks. Task `002797` also showed a Jest worker serialization crash, and `000133` showed connection-refused errors in other tests; each has independent assertions reproducing the requested bug. Their acceptance establishes a valid no-op zero, not complete verifier health.
 
+## Reported reward hacks
+
+[Vals AI](https://www.vals.ai/blogs/mimo-reward-hacking) reported readable solution commits in 1,795 of the 2,698 original Code environments, plus file modification times that could reveal which files the reference patch changed. Our comparisons of old and re-uploaded images found that upstream had removed unreachable Git objects in the inspected images. This is evidence about those images, not a claim that every original task is now leak-free.
+
+All 2,509 task Dockerfiles in this release include a build-time repository repair: remove non-ancestor refs, expire reflogs, prune unreachable objects, and normalize tracked-file and directory modification times to the HEAD commit time. The release contains these Dockerfiles and pinned upstream image references, not rebuilt Docker images. In the saved 39-task repair pilot, 22 environments had timestamp outliers before repair; the history and timestamp checks passed for all 39 after repair. This pilot is separate from full datasource validation.
+
+We also revisited [Charlie Ruan's audit](https://github.com/CharlieFRuan/rl-env-cleaner/tree/836c6061a2d8605155c47041a1ee02e5e5554e6e/mimo_swe) against the re-uploaded images. The recorded disposition of the **27 reported cases** was:
+
+- **21 resolved by the upstream re-upload.** The previously reported leakage was recorded as removed; this group includes the trajectory-only case `002607`.
+- **3 archived.** Image `000484` was removed upstream. Task `000092` still exposed feature-specific tests in an installed package, and `001397` still exposed the implementation fix.
+- **3 retained after inspection:** `000407`, `002070`, and `002517`. The inspected differences were older code or required package-release content, without the reported fix being identified.
+
+The **56 additional residue flags** were also rechecked on 2026-10-18; they are not 56 confirmed exploits or an outstanding backlog:
+
+- **27:** the reported leftovers were absent or matched the repository.
+- **22:** installed project files still differed. Copying those files into the repository and running the verifier returned reward 0 for every task. This specific transplant did not solve the tasks; it does not establish that every leftover is harmless.
+- **7:** cache or bytecode leftovers required content inspection instead. Task `000253` contained the requested feature's API in an installed package and was archived. The other six (`000046`, `000283`, `000556`, `000699`, `000114`, `001372`) were retained after the inspected leftovers did not contain the searched feature identifiers; no successful exploit was demonstrated by that inspection.
+
+These are the residue-review decisions. Other validation findings can independently exclude a task. The three content-residue exclusions in the release are `000092`, `001397`, and `000253`; the removed image `000484` is counted separately among missing upstream images.
+
 ## Runtime and validation time
 
 - **Versions:** Apptainer **1.5.4** on the main Helma run, Python **3.12.14**, Harbor **0.8.1**.