Parity results: parity.md

parity.md

@@ -0,0 +1,80 @@
+# Parity: explorer (original) vs Harbor
+
+**Original side:** the MiMo RL Environment Explorer's runner (`app/runner/domains.py`), which follows Xiaomi's
+harness.
+
+**Harbor side:** the tasks this adapter generates, run with `agents/mimo_opencode.py` and `jobs/*.yaml`.
+
+Both sides use the same settings:
+- Agent: OpenCode 1.18.32.
+- Model: `zai-org/GLM-5.3` via deepinfra, thinking low, replies up to 65,536 tokens.
+- Step limit: the reference limit for each dataset.
+- Sandbox: HF Sandbox, same flavors.
+- Judges: Inkling (text) and Llama-4-Maverick (vision).
+- Graders: the same vendored code.
+
+## Sanity check: 1 task per dataset, 1 run per side (2026-09-26)
+
+The first parity task of each dataset, chosen by `--split parity`.
+
+| Dataset | Task | Explorer | Harbor | Notes |
+|---|---|---:|---:|---|
+| code | format-code-task-000192 | 0.0 | 0.0 | hidden tests fail on both sides |
+| cyber | arvo_13428 | 1.0 | 1.0 | Harbor's earlier valid run scored 0: the model patched the bug instead of submitting a PoC |
+| general | s3k_0057_accounting_audit_tax_en_t2_rl_008 | 0.625 | 0.625 | the same rubric score |
+| terminal | candidate-0109-science-robotics | 0.0 | 1.0 | the explorer run passed 6 of 7 tests (one edge case failed); the task scores all or nothing |
+| webdev | dasyn_260630_00561 | 0.76 | 0.747 | the same judge on two different sites |
+| music | music-gK-0229 | 0.0 | 0.0 | both rejected by the validity gate for bars of the wrong length |
+
+These runs cost about $2.40 on the explorer side and about the same on Harbor.
+
+This is one run per side. It shows that both sides set up, run and grade each dataset the same way. It doesn't yet
+measure parity, which Harbor defines as overlapping score ranges over at least 2 runs per side.
+
+### What the sanity runs found and fixed
+
+- **Webdev's grader couldn't import its rubric.** The vendored files use package-relative imports, so they now ship
+  as the `webdev` package.
+- **General's verifier printed the first characters of the judge key.** Harbor scrubs whole secrets only, so the
+  grader now redacts key-shaped text in its output and in every verifier file.
+- **Harbor blanked every `65536` in the results.** It treats env vars named `*TOKEN*` as secrets, and
+  `OPENCODE_EXPERIMENTAL_OUTPUT_TOKEN_MAX` qualifies. The agent class now sets that variable itself.
+- **A trial with no model access still records reward 0.** Harbor keeps the trial and scores it 0, with no verifier
+  error. When comparing, look at `agent/opencode.txt`: 0 steps with an auth error means an infrastructure failure,
+  not a model result.
+- **The explorer's Music runner had been broken since the security audit** (`httpx.stream` with `extensions=`).
+  It is fixed in `app/runner/domains.py`.
+
+## Full parity: 36 tasks, 3 runs per side (2026-09-27)
+
+The parity subset (`--split parity`): 6 tasks per dataset, fixed by the sha256 of the task id. Run 1 was done on both
+sides before runs 2 and 3, as Harbor's parity guide asks. Each run's score is the mean reward over the tasks that side
+scored. Scores are mean ± sample SEM over the three runs.
+
+| Dataset | Tasks | Original (mean ± SEM) | Harbor (mean ± SEM) | Original runs | Harbor runs | Ranges overlap |
+|---|---:|---:|---:|---|---|:---:|
+| Code | 6 | 0.556 ± 0.056 | 0.556 ± 0.056 | 0.667, 0.500, 0.500 | 0.667, 0.500, 0.500 | yes |
+| Cyber | 6 | 0.822 ± 0.011 | 0.833 ± 0.096 | 0.800, 0.833, 0.833 | 1.000, 0.667, 0.833 | yes |
+| General | 6 | 0.844 ± 0.008 | 0.844 ± 0.024 | 0.842, 0.832, 0.859 | 0.890, 0.811, 0.832 | yes |
+| Music | 6 | 0.111 ± 0.063 | 0.317 ± 0.093 | 0.000, 0.114, 0.219 | 0.138, 0.449, 0.363 | yes |
+| Terminal | 6 | 0.278 ± 0.056 | 0.111 ± 0.056 | 0.333, 0.333, 0.167 | 0.167, 0.167, 0.000 | yes |
+| Webdev | 6 | 0.749 ± 0.049 | 0.738 ± 0.054 | 0.650, 0.803, 0.793 | 0.799, 0.631, 0.785 | yes |
+
+**All six datasets meet Harbor's criterion: the run-score ranges overlap.** Code, Cyber, General and Webdev agree
+closely.
+
+- **Terminal (0.278 vs 0.111):** each task is all or nothing, and outcomes flip in both directions on both sides
+  (for example `candidate-0109` scored 0/1/1 in the explorer and 0/1/0 in Harbor). With six tasks, that is noise.
+- **Music (0.111 vs 0.317):** the validity gate makes each piece close to pass or fail, and it flips run to run on
+  both sides. Harbor passes the gate more often (7 of 18 pieces vs 4 of 18). That fits Music's one intended
+  difference: in Harbor the agent writes its reply to a file with up to 4 steps, where the original is a single
+  completion.
+
+**Infrastructure failures were rerun, not counted.**
+- Some first attempts timed out talking to the sandbox API (about 26 sandboxes ran at once): 9 original-side and 6
+  Harbor-side tasks in run 1, fewer in later runs. Those were rerun at lower concurrency.
+- `arvo_49407` run 1 failed on both sides the same way: the agent started 16 parallel fuzzers on a 2-vCPU
+  `cpu-basic` sandbox and starved the sandbox server. It is excluded from run 1 on both sides.
+
+The explorer side of these runs cost about $77; the Harbor side about the same. `parity_experiment.json` has the raw
+numbers in Harbor's format.