[ydnyshhh]

Summary

This PR adds Synthetic Workspace Gym (SWG) as a native Verifiers V1 environment under environments/synthetic_workspace_gym/. D1-D3 tasks support teacher-trace generation for mid-training/SFT; D4-D5 tasks and held-out compositions support RL and evaluation. The taskset uses Prime's standard Verifiers harness/runtime and has no custom model loop or legacy adapter.

Generation, immutable curriculum construction, validation, and artifact publication are submitted separately in #25.

Native Verifiers V1 design

• Native vf.Taskset and vf.Task implementations.
• Typed task data, task state, and configuration.
• Strict binary @vf.reward (1.0 only on trusted evaluator success, otherwise 0.0); the former partial score is a diagnostic @vf.metric and cannot affect training reward.
• NEEDS_CONTAINER = True, /workspace, and the standard Bash harness.
• A real validate(runtime) gold control that applies the reference solution, runs trusted finalization, and requires success with reward exactly 1.0.
• verifiers>=0.4.0,<0.5; the task uses the 0.4 setup(trace, runtime) hook and applies network restrictions on both published and newer runtime APIs.

Security and reward robustness

Finalization separates three trust domains:

1. The agent runtime contains only the selected split's visible workspace.
2. A fresh answer-free execution runtime receives a bounded workspace snapshot and runs candidate-controlled code.
3. A separate trusted grader receives the executed snapshot, hidden expected values, evaluator code, and host-owned execution metadata, but never executes candidate code.

Review-driven hardening includes:

• bounded whole-workspace snapshots preserve candidate-created files and deletions;
• score-bearing execution outputs live outside the agent-writable workspace;
• candidate commands have a 120-second process-group timeout;
• output paths are cleared before determinism reruns;
• nondeterministic command output receives zero reward on every difficulty;
• output with structure but no semantic overlap is capped at 0.2;
• observation payloads are framed in a child process and written by the trusted wrapper only after candidate shutdown;
• network access and transferred paths/file sizes are bounded;
• untouched baselines are measured through the evaluator that actually ships.

Immutable V5 pins

• Public dataset: ydnysh/synthetic-workspace-gym
• Dataset revision: 3ba83f647b96556be8c5084584a2d86e93225684
• Exporting public generator commit: 3bd13bb949846e7b7393afe81727f1127e984475
• Export integrity SHA-256: 128c3c984d91e7d46918e0f2f77cab4543943a25e11083bf18d0ad00dcf9f7f8
• Public Prime namespace/tag: prime/yadnyesh, v5-128c3c98
• Default curriculum: sft-easy-v5

artifacts.json records the seven completed public Prime VM artifact IDs and CAS paths. Frozen V2-V4 artifacts remain unchanged.

Validation performed

uv run ruff check .
uv run ruff format --check .
uv run pytest -n auto tests -v
uv build environments/synthetic_workspace_gym

• Ruff and formatting passed.
• All 7 public-repository package tests passed, including the SWG package build. A focused four-case regression check confirmed that partial diagnostic scores, even 1.0 without success, return zero reward.
• Taskset sdist and wheel 0.2.4 built successfully. V5 dataset and image pins are unchanged because only the package-level reward mapping changed.
• V5 passed 1,328/1,328 reference solutions, all 1,328 untouched ceilings, and the focused reward-hacking regressions.
• The official validate(runtime) gold path completed on Prime with success and reward 1.0.
• On Verifiers 0.4.0, two Linux/Prime smoke rollouts completed task setup, candidate execution, trusted grading, and reward recording without trace errors. Runner and grader network policies both enforced block=["*"].

Note

Add synthetic-workspace-gym taskset with evaluators, schemas, and trusted grader

• Adds a new taskset package for the Synthetic Workspace Gym, including task data/state models, taskset loading from a pinned dataset revision, and the SyntheticWorkspaceTask setup, finalization, reward, and validation hooks in taskset.py
• Adds five family evaluators (tabular, script-repair, pipeline, retrieval-workspace, composite-workspace) with capability scoring, weighted match metrics, required-artifact checks, and a D5CalibratedEvaluator wrapper that cubes scores for failed level-5 tasks, in evaluators/
• Adds shared schemas in common.py, JSON/IO/path helpers in utils/, and a trusted grader entrypoint in trusted_grader.py that runs the evaluator and writes result JSON
• Finalization runs the candidate execution plan in an isolated runtime, grades in a no-network grader runtime, and stores evaluator results and file counts in task state
• Risk: D5CalibratedEvaluator.evaluate alters score semantics for unsuccessful difficulty-5 results (score cubed); evaluators reject unsupported protocol modes (e.g. non-observations-v1 in ScriptRepairEvaluator.evaluate) with ValueError

^{Macroscope summarized f5a7987.}