Release native trainer state before cached refinement, rebuild each detector from the persisted EMA checkpoint, save gradient controls incrementally, and reuse hash-verified completed controls after interruption.
Constraint: Preserve baseline checkpoints and fixed optimization budgets
Rejected: Retry candidates in retained trainer process | keeps opaque CUDA state alive
Confidence: high
Scope-risk: moderate
Not-tested: Full three-seed remote development matrix
Keep objective construction on CPU and evaluate cached detection features in source-sized chunks. Adam gradients now accumulate per chunk, avoiding a monolithic 2,500-image CUDA graph without changing the full-objective mean.
Constraint: Preserve 2,500-image objective and exact 41 Adam evaluations
Rejected: Reduce objective sample count | changes the sealed protocol
Confidence: high
Scope-risk: moderate
Not-tested: Full remote 2,500-image CUDA optimization
Allow the sealed ResNet and YOLO convergence protocol to run unchanged on the remote NVIDIA accelerator while retaining CPU and MPS support.
Constraint: Preserve fixed seeds, exact workloads, and one-device-per-run semantics
Rejected: Reuse mixed MPS and CUDA results | invalid cross-device production matrix
Confidence: high
Scope-risk: narrow
Not-tested: Full production matrix on RTX 3080 Ti
Migrate the package and examples to the tensor-native PyTorch implementation, add benchmark evidence, and add the guarded post-training convergence protocol with TensorBoard progress monitoring and hash-verified recovery.
Constraint: Preserve one-shot official-test sealing and auditable research artifacts
Rejected: Commit local .omc runs and downloaded datasets | multi-gigabyte runtime state is machine-local
Confidence: high
Scope-risk: broad
Not-tested: Production CUDA run on pieroot-server
jung-geun
jung-geun