Protocol and access

Documented, hash-bound protocol document.

All binding scientific choices were confirmed, including mandatory final training on all 100 auxiliary actions. Independent research governance is recorded in docs/protocol/independent-research.md. The result-blind NTU input-contract amendment was adopted on 2026-09-24. Current protocol SHA-256: c8b08b6867bc14dc0a947ebfa94e40b16ea3f0dae38dfa6d86187afecb4fd45f.

Input inventory and checksum report.

Verified 2026-09-24: dataset access remains active, the complete usable HRNet aggregate, official missing-skeleton list, one-shot definition, and MotionBERT checkpoint are readable and checksummed on the GPU host. The aggregate contains 113,945 usable annotations and all 20 official exemplars; the official 535-item missing-skeleton list accounts for the remaining nominal captures. The result-blind aggregate-aware input contract is adopted in docs/protocol/input-contract-amendment.v1.md, and w01-gate-02 is met. The novel test remains sealed.

Compute profile with a justified physical batch-size decision.

Completed 2026-09-16: a real frozen MotionBERT encoder profile ran on a scheduler-assigned BU SCC A40 using dense protocol-shaped float32 inputs. Physical batch 32 completed with 2.572 GiB peak allocated, 2.855 GiB peak reserved, and an 842.164 ms median forward; batch 64 completed with 4.977 GiB allocated, 5.514 GiB reserved, and a 1,685.400 ms median. Required device, CUDA, scheduler, quota, and job-limit evidence was captured. Physical batch 32 remains selected because protocol v1 permits 64 only after all core methods fit. This is compute evidence only, not preprocessing or encoder-parity evidence.

Decision evidence or a visible blocking record.

The computational scope is recorded in docs/compliance/computational-research-scope.md. On 2026-09-24, the researcher confirmed that BU has issued the applicable human-subjects/data-governance determination and requested completion without attaching proof. The confirmation is recorded and w01-gate-04 is met. This is researcher confirmation, not independent document verification; no specific institutional classification or original decision date is asserted.

A complete Week 1 progress record.

Verified inputs, remote setup, the adopted result-blind input contract, and the researcher-confirmed BU determination are recorded. Final closeout remains blocked until researcher-supplied actual minutes are recorded. No hours or week closure are inferred.