← Actions or scenes?

Wins

• The question got an answer above noise, on every split. Over three splits and two seeds the S3D shuffle gap lies in 10.8 to 13.2 points and the Swin3D-T gap in 2.2 to 5.0, against a seed-pair noise floor of 2.3; R(2+1)D-18 on split 1 gave 16.5. Backdrop alone is worth about 48 % to a per-frame model. The most accurate model is the one that depends least on motion.¹
• The instrument reads in both directions. With R(2+1)D-18 cut from 16 to 4 frames, top-1 fell from 70.5 to 51 %, the permuted-frame accuracy stayed at 49 to 57 %, and the gap fell to 0.9 and 1.8 points against a 2.5-point floor; the gate’s verdict flipped to appearance-solvable on its own.¹
• The two channels fail differently. Under darkening, blur and compression the per-frame control loses up to 29 points while the convnets’ shuffle gaps hold at 13 to 18; under frame hold the control loses under 2 points and the convnets fall to their permuted scores. Appearance and motion degrade on different axes, which is what a deployment would need to know.²
• The gate refused numbers until it had proved it could fail. Shuffled-label controls sat at 1.9 % and 2.2 % against a chance rate of 1.96 %, and the 2D control was frame-order invariant to four decimals, exactly as the review had argued. Nothing downstream was read before that.³
• An edge frontier, measured. R(2+1)D-18 in TensorRT INT8 ran at 0.77 ms per clip at batch one, 3.4 times faster than fp32, with top-1 unchanged (70.3 to 70.6) and power draw roughly halved. It is the only non-dominated point above the control on the latency and accuracy frontier.⁴
• A quantisation collapse located, explained and fixed. A per-layer ablation put it in the stem’s first activation tensor; its measured statistics (median 0.46, absmax 45.9, 60 × per-channel spread over 2.5 × 10⁹ values) explain why no per-tensor calibrator can serve it. One convolution in fp32 restores INT8 to 66.1 % against 66.3 % fp32, from 2.0 % naive, in a workflow run. Detour and mechanism: rabbit holes.¹
• Every number is traceable. Each phase ran as a sandboxed unit inside a Temporal workflow with a config hash and a clean commit recorded in its metrics; a rerun after a fix is a new run with a new id. The cancel path, lease release and metrics-on-failure readback were all exercised for real, not just in tests.¹
• A real dataset fact surfaced by a routine check. Two duplicate clips cross train and test in the official splits 1 and 2; the contamination check found them, dedupe dropped the test copies, and the report records it.³