Skip to content

Versioning

TOWER versions software, benchmark data, and persisted wire formats separately.

Surface Scheme Release contract
nero-kinematics, tower-eval, tower-sim Semantic Versioning (MAJOR.MINOR.PATCH) All three reach 1.0.0 at the public release. Breaking public API changes increment major; compatible features increment minor; fixes increment patch.
Benchmark Calendar Versioning (YYYY.MM.REVISION) 2026.11.1 is the public benchmark revision, using tower-scene-v2 and a published all-600 physical equivalence map.
Wire schemas tower-*-vN Increment N for an incompatible format; do not infer it from a package or benchmark version.

Current package versions are development versions, not a published 1.0 release. The package release workflow uses coordinated stable tags vMAJOR.MINOR.PATCH; all three package metadata versions must match the tag. It does not change versions or dependency constraints itself. Package maintainers must align their inter-package constraints and the workspace lock before the public release.

Persisted jenga_* schema strings remain readable through historical data readers. Renaming the project does not authorize rewriting old recordings, changing their schema identifiers, or dropping old readers. New formats use the tower-*-vN family. For example, tower-infer-v2 is a protocol version, not package version 2.0 or a benchmark revision.

The public repository starts with fresh history. The internal predecessor 2026.11.0 has identical resolved physics and remains internal archival evidence; its tag and scene files are not part of the public repository. Historical receipts never substitute for current posture and safety acceptance.

Release machinery

.github/workflows/release.yml runs only on package tag pushes. Branch pushes, PRs, benchmark tags and asset tags do not start it. It requires matching package versions and final docs/release-notes/<version>.md notes (the Status: Draft marker blocks release), verifies the public HF asset tag against its locked commit and manifest SHA-256, then builds three wheels and three sdists. The GitHub Release includes these six distributions, the asset lock and SHA256SUMS. The current lock is packages/tower-sim/src/tower_sim/data/benchmark/assets.lock.json; any future relocation must update the workflow's ASSET_LOCK in the same revision.

Builds use isolated build environments and do not install Isaac or run GPU jobs. There is no PyPI publication step; publication requires maintainer approval.

See the changelog and 1.0.0 notes draft.

Scene revision 2026.11.1

The scene schema contains fixed evaluation state, neutral scene identifiers and content-addressed resources. Scene payloads are external data; the committed manifest and equivalence table bind them. Only evaluation inputs and content hashes are public. Internal lineage is not part of the repository or public dataset.

The request remains tower-public-benchmark-v2: its scene-selection values identify fixed cases, and its policy/action semantics are unchanged. Envelopes remain tower-public-envelope-v1 and record the selected benchmark, data version and manifest hash. Case numbers are independent of historical source identifiers.

The complete v1 reader and lossless compatibility view require archival evidence. Maintainer rechecks accept an explicit --v1-root; no archive is discovered from checkout history. Without that explicit source, compatibility checks skip clearly and cannot produce a successful equivalence receipt. A supplied invalid archive is an error. Full-output golden values are unchanged; scene identifiers use neutral names.

The new scene manifest binds only new metadata and external evaluation payloads. All new files have zero host/IP/path leak exemptions. The data lock pins the public dataset at immutable commit d78e4ac0b2bfc1e10a15d31ac7e7d723be2764f1. Local verification uses the same hashes as the public release. Package versions are unchanged. Staging performs no upload.

The public wire stores each fixed state under episode. The closed schema retains measured_asset_id, which binds the collision/visual asset, and removes archival provenance. After hash checks the decoder supplies the runtime asset-closure flag. The full resolved runtime input is compared for all 600 scenes after the explicit archive-only dropped-field list and mechanical identifier changes.

Data identity binds payload and metadata hashes plus the upload MANIFEST hash. The HF repository/revision/tag/directory is a separate locator. Filling those fields leaves data_manifest_sha256 and existing GPU receipts valid; changing manifest_sha256 is rejected. The locator is not part of the upload.

Shared IK specification identity

The packaged shared IK specification has provenance-redacted bytes with SHA-256 3247cd2eed229c367341f26e2d7bf95dca845b94f74b6ade575c345cf32a7911. Its canonical identity remains efa69beb12a5bb972a2f570a9b2310f9dcfa852a91bd095295b205a66b8a0d03, read from the nero-kinematics command profile. The runner verifies the packaged byte hash before binding that canonical identity. Bound values, IK profiles and command limits are unchanged; request ik_bounds must still match the package file bytes.

Station distribution and calibration identity

The public station digest is 54fe7ea5f90b8919f6fece5b2d3e926f9f9f14a0bf7c69659b5e3aa38d474ad9. Only this pinned digest maps to the canonical calibration identity 59d3443b7020a365beb648ebbf428cae750559ca320824ca99d77824d4c26d27. The metadata-only re-authoring replaces internal path strings with redacted, including the layer documentation. Every USD field outside the enumerated string metadata, all numeric arrays, references and texture bytes were compared exactly. The runtime configuration comparison is exact again; changed configuration values fail with their JSON pointer. TCP and fit-equivalence calibration pins retain their canonical identity.

The public asset manifest is independently pinned by its lock. Maintainer GPU comparison maps only the sealed public manifest digest to the historical asset identity, after verifying every actual public file. The private proof covers the required asset subset, unchanged texture/physics data and redacted golden members. This mapping changes no comparison thresholds or scoring behavior.

Policy transport revision

The unreleased tower-public-benchmark-v2 request now requires XPolicyLab WebSockets. policy_url and --policy-url keep their names. This is a runtime transport revision, with no benchmark, data or scene identity change. See the Isaac runtime contract for lifecycle, response reconstruction, metrics and MME availability.

Benchmark run schemas

The unreviewed request protocol is tower-public-benchmark-protocol-unreviewed-v1; the reviewed protocol remains tower-public-benchmark-protocol-v1. Only the explicit unreviewed schema skips maintainer review. Missing or other schemas still require full review. Aggregate scores use tower-benchmark-score-v1; informational posture reports use tower-posture-flags-v1 and tower-ordinary-tcp-reference-v1. The benchmark revision 2026.11.1, ruleset bench-2026.11, package versions and data hashes are unchanged.

Run selection uses tower-benchmark-run-v1; retained request bindings use tower-benchmark-attempt-v1. Keep these records with recordings for offline coverage, attempt lineage and relocatable aggregation. They establish local content consistency, not signed authenticity or protection against scene selection.