# Astra: frozen script policies (Agent-D20-E20) One frozen, non-neural script policy per ManiLadder task, written by a GPT-6 Astra coding agent under the `Agent-D20-E20-Compiled` protocol: for each task the agent received 20 demonstrations (D20) and 20 exploration episodes (E20), then froze a controller that is evaluated without any further changes. These are Python controllers plus their reference data. They are not PI0.5 weights and load no neural network. All paths in this file are relative to `astra/`. ## Layout ```text astra/ README.md manifest.json every file of every task with its SHA-256 and size results_snapshot_20261105.csv 50-episode evaluation results available on 2026-11-05 shared/ protocol, environment lock and runner files common to all tasks / the frozen policy of one task (112 folders) controller.py entry point: reset_episode() and act(obs) policy_runner.py, worker_api.py freeze_manifest.json SHA-256 of every file in this folder, bundle hash, selection record model_tool_manifest.yaml runtime dependencies and declared assets environment.lock handoff.md, findings.md, skills.md, demo_manifest.json demonstrations_readonly/demo_00.h5 ... demo_19.h5 *.npz / *.npy / helper *.py task-specific reference data and helpers ``` Each `/` folder is the unmodified content of the agent's original `tasks//frozen/` directory. File names beyond the common ones differ from task to task. ## Coverage All 112 tasks have a complete frozen bundle: every file listed in the task's `freeze_manifest.json` is present and matches its recorded SHA-256. `manifest.json` gives the origin of each bundle in `policy_source`: | `policy_source` | Tasks | Meaning | |---|---|---| | `original_frozen_20261102` | 100 | first frozen snapshot (2026-11-02) | | `reexplore8_20261104` | 8 | policies frozen after a second exploration run (2026-11-04): DualArmKeyboardType-v1, DualArmOrganizeCabinet-v1, DualArmPingpongBalance-v1, DualArmPlugCharger-v1, DualArmSandShovel-v1, FitnessBall-v1, KitAssembly-v1, TweezersPickPlace-v1 | | `added3_20261105` | 2 | BagDropIn-v1 and BagHanging-v1, frozen on 2026-11-05 against the task code of that date (before the later wall-scene releases of these two tasks) | | `continue2_20261108` | 2 | DualArmAssembleBracket-v1 and PressFitSoftPeg-v1, frozen on 2026-11-08 | In seven of the re-explored tasks some or all of the 20 demonstration files are byte-identical copies of each other; this is how those bundles were frozen and they are published unchanged. ## Download one task ```python from huggingface_hub import snapshot_download snapshot_download( repo_id="Linslab/ManiLadder_Models", allow_patterns=["astra/StackBlock-v1/*", "astra/shared/*", "astra/manifest.json"], local_dir="./models", ) ``` A task folder is between roughly 0.5 GB and 11 GB, almost all of it demonstrations. The controller reads only the assets named under `runtime_assets` in its `model_tool_manifest.yaml`; the demonstrations are part of the frozen bundle and are needed for the integrity check, not at run time. ## Evaluate a frozen policy The evaluation harness (simulation workers, isolation sandbox, accounting) is the `agent_e20` package of the ManiLadder code repository: . To use a policy from here with that harness: 1. Copy `astra//` to `agent_e20/tasks//frozen/` in a checkout of the code repository, keeping all files and names. 2. Check the bundle with `agent_e20.isolation.verify_frozen`, or compare the files against `freeze_manifest.json` / `manifest.json` yourself. 3. Run the evaluation entry point described in the code repository. The protocol is 50 evaluation episodes per task with the task's native horizon; `policy_runner.py` shows the control loop (`reset_episode()`, then `act(obs)` until termination, truncation, the horizon, or `None`). `shared/` holds the files that define the protocol and runtime for all tasks: `protocol.yaml`, `environment.lock`, `model_tool_manifest.yaml`, `task_manifest.json` (task list with level, suite and inclusion flags), `observation_contract.json` (the observation keys a policy may read per task), and the runner-side `policy_runner.py`, `worker_api.py`, `timed_policy.py` and `isolation.py`. They are copies taken from the code repository on 2026-11-16; the code repository is authoritative. `task_manifest.json` and `protocol.yaml` describe the original campaign of 108 included tasks; the four tasks listed there as excluded were frozen later (see the coverage table). ## Results snapshot `results_snapshot_20261105.csv` lists the 103 tasks whose 50-episode evaluation was final on 2026-11-05 (task, level, suite, policy source, episodes, successes, success rate). Mean success over these 103 tasks is 0.397 (level 1: 0.674 over 28 tasks, level 2: 0.461 over 28, level 3: 0.317 over 20, level 4: 0.102 over 27). Nine tasks are not in the snapshot: BagDropIn-v1, BagHanging-v1, DualArmAssembleBracket-v1, DualArmFlipTable-v1, DualArmMakeLine-v1, DualArmOrganizeCabinet-v1, DualArmPlugCharger-v1, DualArmShapeDoll-v1 and PressFitSoftPeg-v1. For final numbers see the ManiLadder paper and project page. ## Integrity `manifest.json` lists, for each task, the bundle hash, the SHA-256 of `freeze_manifest.json`, and the path, size and SHA-256 of every file. The 100 bundles of the first snapshot were moved here from `agent_d20_e20/frozen_20261102/tasks//frozen/` (previous repository name `Linslab/ManiLadder_PI05_Models`) without changing any file; the other 12 were checked file by file against their `freeze_manifest.json` before upload.