Datasets:

---

Rory726

/

paperbenchx

like 0

Adds 12-robotics/octo-simplerenv, the Octo (arXiv:2405.12213, MIT) SimplerEnv
WidowX-Bridge zero-shot replication task, as a byte-identical copy of its canonical Harbor
task plus the three diagnostic overlay files.

Why this task now. It is the first task in this batch to complete a matched pair of runs
under the clean low-privilege replay contract, so its published evaluator is the one that
actually produced scores rather than an untested revision.

Contents

• 52 source files frozen in source_integrity.sha256, including
  tests/evaluable_replay.py and tests/run_evaluator.sh (the replay contract) and
  tests/refs/widowx_bridge/reported_metrics.json (thresholds and their provenance).
• tests/rubric_mapping.{json,yaml} and tests/evidence_policy.json, generated by the same
  overlay generator used for the nine tasks in #6 (verified byte-identical when re-run
  against that PR's rt1x-simplerenv overlays). 11 rubric leaves, each mapped exactly once,
  effective weights unchanged.
• evidence_policy.json is version: 2 (clean-replay): generated roots cleared per attempt,
  candidate source hashed before and after, replay as uid/gid 65532 with cleared capabilities
  and loopback-only networking, evidence frozen to /logs/verifier/candidate_outputs, replay
  dependencies baked into the image.

Repository-level changes

• package_manifest.json added, and tools/{validate_package.py,finalize_outputs.py} replaced
  with the generalized multi-task versions from #6. The versions currently on main are
  hardcoded to TASK = ROOT / "pi05-libero" and to a tasks_raw/... source path that is not
  part of this dataset, so they cannot validate a second task — or run at all as published.
• source_integrity.sha256 moves from the 5-files-per-task pilot format to the full per-file
  manifest that the generalized validator requires. The 46 pi05-libero lines are
  hash-identical to the ones in #6; none of its 49 files are touched by this PR.
• README.md and UNIFIED_SCORING_FRAMEWORK.md updated for two tasks. They state plainly
  that the two are not at the same hardening level: octo-simplerenv carries the replay
  contract, pi05-libero predates it.

Scoring is unchanged. No reward formula, target, gate, cap, rubric hierarchy, or leaf ID
is modified. python3 tools/validate_package.py passes: validated 2 tasks and 98 frozen
public source files — which also clears its host-path, credential, and private-answer-key
scans.

No private answer keys are withheld for either task: neither grades against a goal table
or an oracle path, so PRIVATE_VERIFIER_ASSETS.md (needed by #6 for mapgpt/navid/navila) has
no counterpart here.

One thing worth flagging rather than quietly fixing: the task's own README.md and task.toml
comments carry stale in-repo notes and links to docs/*.md paths that are not part of this
dataset. They are published as-is, because the package's core promise is that each task
directory is byte-identical to the canonical Harbor source — #6 publishes the same pattern.