Parity results: README.md

README.md

@@ -99,9 +99,18 @@ Deviations from the reference, and why, are listed in the adapter's README.
 - All 7,780 tasks across the six datasets load with Harbor's task loader and pass the adapter's static checks
   (digest-pinned images, setup payload equal to its readable copy, rubrics only under `tests/`).
 - With Harbor's no-op agent every dataset scores 0 and every verifier runs to completion: no free rewards.
-- One task per dataset was run end to end with GLM-5.3 both in Harbor and in the
-  [explorer](https://huggingface.co/spaces/FineEnvs/MiMo-RL-Envs-Explorer), which runs Xiaomi's harness; the results match (see the adapter's `parity.md`).
-  A full parity study over repeated runs has not been done yet.
+- **Parity with Xiaomi's harness.** Following Harbor's parity procedure, the 36-task parity subset (6 per dataset)
+  was run 3 times on each side with the same agent (OpenCode 1.18.32), model (GLM-5.3 via deepinfra, thinking low),
+  step limits, judges and sandboxes, in Harbor and in the [explorer](https://huggingface.co/spaces/FineEnvs/MiMo-RL-Envs-Explorer), which runs Xiaomi's harness. The score ranges overlap for all six datasets. For this one:
+
+  | | Mean reward (mean ± SEM, 3 runs) | Runs |
+  |---|---:|---|
+  | Xiaomi's harness (explorer) | 0.278 ± 0.056 | 0.333, 0.333, 0.167 |
+  | Harbor | 0.111 ± 0.056 | 0.167, 0.167, 0.000 |
+
+  Each task is all or nothing, and outcomes flip in both directions on both sides; with six tasks the gap is within noise.
+
+  Full results, per-task outcomes and notes are in `parity.md` and `parity_experiment.json`.
 
 ## Credits