v1.1.0 — alignment classifier v7

Cubed Core v1.1.0 (proposed)

Cubed Core reconstructs a 3×3 puzzle-cube solve from 120 fps phone video using
per-frame camera evidence and exact search over legal cube states.

Release source:
649390350a859fef30332f0c3fb7421a50a13c34.

What changed since v1.0.0

Alignment classifier promoted to v7. The candidate passed its promotion
gate against the current v1.0.0 alignment classifier:

• Recall at threshold 0.50 — aligned: 0.976 (val) / 0.942 (test). unaligned:
  0.849 (val) / 0.895 (test).
• Decode-neutrality smoke check on decode tags gt17 and gt21: byte-identical
  decode output checksums against the prior alignment classifier. This
  promotion changes read-time alignment gating; it does not change decoder
  behavior on those two smoke tags.
• These are within-corpus results against the maintainer's existing capture
  set and evaluation split, not a generalization claim. See
  Evidence and the model card in
  camera-tracker-v1-runtime.tar.gz for the full boundary.

Face-pose candidate evaluated, not promoted. A separate face-pose
candidate was evaluated the same night and failed its per-cell promotion gate:
pooled decode-frame read accuracy was 0.748 for the current v1.0.0
face-pose model versus 0.729 for the candidate, worse on all five
evidence-rich decode tags used for the gate — despite the candidate reporting
a better held-out validation corner error (31px candidate vs 187px current).
The validation-corner-error metric and the decode-frame read-accuracy gate
disagreed in direction. The promotion gate is decode-frame read accuracy, so
the candidate was not promoted; face-pose.onnx in this release is
byte-identical to the file shipped in v1.0.0. Recorded as a lesson: a large
validation-geometry improvement did not transfer to decode-frame read
accuracy for this candidate.

Dataset r3 and schema v3. raw-cube-session-v3.schema.json was added for
relative-timestamp BLE sessions (dataset release r3-2026-08-02 republishes
cube_session.json with relative t_ms; wall clock stays stripped). The
validator accepts the new shape for the ble_session role. See
Dataset.

The checked-in dataset download pin (config/public-dataset-v1.json) still
points at the previously reviewed revision
(7fae604962c590ac9c658ba6ee0350e86de9c4f5, 35 captures) as of this staged
release's source commit. If the r3-2026-08-02 HF revision is meant to be
the pinned download target for make download-dataset in this release,
that pin needs a separate reviewed update before or alongside publishing —
it was not part of the requested asset set and is flagged here rather than
changed.

Workflow

1. Open Demo for a preserved, no-compute run.
2. Download the published dataset and release assets with
   make download-dataset download-assets, or add your own video with its
   exact starting scramble and a six-color calibration.
3. Run Decode on a local or remote NVIDIA CUDA host.
4. Inspect the saved result in Runs. Runs never starts, retries, or resumes
   compute.

Warning

Native Windows is experimental and may still be broken. Use WSL2 or Linux
for a dependable path. Live Decode is blocked on native Windows.

Report only what the named result supports: "The job completed on this
input." If its sequence was independently replayed to solved, say: "The
sequence replayed to solved." For an abstention, say: "The run abstained."
These statements are not accuracy or generalization claims.

Release assets

• camera-tracker-v1-runtime.tar.gz — alignment (v7) and face-pose
  (unchanged) ONNX models with their manifest, model card, audit, checksums,
  and license.
• trust_v1_numpy.npz, calibration_gan12.json, and
  decode-support.LICENSE.txt — the read-trust model, published shared
  calibration, and notice (unchanged from v1.0.0).
• gtD1s_decoder-demo-clip_f3919-f9722.mp4 and its license notice — the clip
  used by Demo (unchanged from v1.0.0).
• SHA256SUMS — hashes for the runtime and media assets above.

This asset set is the same 7 files, same names, as v1.0.0; only the runtime
tarball's bytes changed. See MANIFEST.md in this staging bundle for the
full byte/hash table and the changed/unchanged marking of every file.

The released models were developed against the maintainer's capture setup.
See the pinned dataset revision,
camera tracker, and
read-trust model for data,
model identities, and limitations.

Read the tagged
README for
setup and
evidence guidance
before reporting a result.

Cubed Core v1.0.0

Cubed Core v1.0.0

Cubed Core reconstructs a 3×3 puzzle-cube solve from 120 fps phone video using
per-frame camera evidence and exact search over legal cube states.

Release source:
6d2d7314bb7b2bb3360c7b2b46a427e88fb36242.

Workflow

1. Open Demo for a preserved, no-compute run.
2. Download the published dataset and release assets with
   make download-dataset download-assets, or add your own video with its
   exact starting scramble and a six-color calibration.
3. Run Decode on a local or remote NVIDIA CUDA host.
4. Inspect the saved result in Runs. Runs never starts, retries, or resumes
   compute.

Warning

Native Windows is experimental and may still be broken. The current
/api/captures and media-ticket fix has not run on a real Windows machine or
in hosted Windows CI. Use WSL2 or Linux for a dependable path. Reproducible
Windows failures are high priority and will be fixed as quickly as possible.
Live Decode is blocked on native Windows.

Runs can show synchronized video, camera overlays, sampled reads, trellis
spans, reconstruction checkpoints, and an optional post-decode BLE comparison
when those fields are present. Between decoder-supported checkpoints, Runs may
evenly pace ordered moves as a display-only estimate. Those intermediate states
are labeled estimated; only the saved checkpoint state is
decoder-supported.

Report only what the named result supports: “The job completed on this input.”
If its sequence was independently replayed to solved, say: “The sequence
replayed to solved.” For an abstention, say: “The run abstained.” These
statements are not accuracy or generalization claims.

Release assets

• camera-tracker-v1-runtime.tar.gz — alignment and face-pose ONNX models with
  their manifest, model card, audit, checksums, and license.
• trust_v1_numpy.npz, calibration_gan12.json, and
  decode-support.LICENSE.txt — the read-trust model, published shared
  calibration, and notice.
• gtD1s_decoder-demo-clip_f3919-f9722.mp4 and its license notice — the clip
  used by Demo.
• SHA256SUMS — hashes for the runtime and media assets above.

The released models were developed against the maintainer’s capture setup. See
the pinned dataset revision,
camera tracker, and
read-trust model for data,
model identities, and limitations.

Read the tagged
README for
setup and
evidence guidance
before reporting a result.