Video frame interpolation
- Score
- 88.4
- Wallet spend
- $41.88 of $200
- Model cost
- $33.20
- Agent time
- 9h 32m
- Steps
- 912
- Tokens in / out
- 95.5M / 625k
- Self-checks
- 4 passed of 4 used
- Rank at $200 wallet
- 2nd of 19
Score
88.4raw metric 0.8839
The score is the raw metric (balanced LPIPS and FloLPIPS gap closed against a frozen baseline), which already runs from 0 to 1, times 100.
Verifier details
- scoring_protocol
- vfi-lpips-flolpips-two-anchor-v2
- metric_protocol
- vfi-lpips-alex01-flolpips-sequence-v1
- psnr_cap_db
- 100
- metric_seconds
- 296.0944
- scoring_seconds
- 377.5630
- evaluation_sha256
- 375276285d8168e061004e5873daeecbb7e8db7c6e8354ef72d50ad47496e236
- metric_bundle_sha256
- 98e13949365e1541fbec7972317356abb77c00956c7126d0d81dd99f045d5fdd
- calibration_sha256
- 8c72e23cb352711cfddff07fea74a4935a99f753ea7b498dbd83a999278630ab
- prediction_manifest_sha256
- 03b38e44c2bf4bae7888161a98d033fc77ad08ef1f3eed9be81bf4da9e92d915
After grading, a reviewer model reads the submission's files and the agent's trajectory and checks each of the task's rules. A failed check makes the score 0. Step numbers link into the trajectory.
✓non specialist initializationPasses. The AMT-L network was trained during the session. training/train.py's initialize_from_raft copies only 10 tensor
Passes. The AMT-L network was trained during the session. training/train.py's initialize_from_raft copies only 10 tensors from the RAFT-Large feature encoder: the stem and layer1 conv tensors. data/logs/amt01_initialization.jsonl confirms those 10 copies and random initialization for everything else. RAFT C_T_SKHT_V2 is a standalone optical-flow model, and data/download_artifacts.sh records its torchvision source. The trajectory shows the UPR-Net, AMT, IFRNet and EMA-VFI source archives being downloaded (trajectory#11-13), and the UPR-Net zip contains .pkl checkpoints. No step loads those checkpoints or any other VFI-specialist checkpoint, and later steps describe the UPR-Net archive as unused.
✓learned frame generationPasses. serve.sh installs the bundled wheels and runs server.py. server.py builds the AMT-L Model and loads weights/amt_
Passes. serve.sh installs the bundled wheels and runs server.py. server.py builds the AMT-L Model and loads weights/amt_ema.pth plus the factor-specific checkpoints (weights/amt_ema_8x.pth, weights/amt_ema_16x.pth) with strict loading. It then generates each intermediate frame with a learned forward pass at embt=i/factor. The only fallback is warping driven by learned RAFT flow, using weights/raft_large.pth. The weight inventories show complete tensor sets in these files.
✓independent intervalsPasses. server.py builds each response only from the request's own img0/img1 pair and factor. A global lock serializes r
Passes. server.py builds each response only from the request's own img0/img1 pair and factor. A global lock serializes requests, and no cache or state carries over between requests or intervals.
✓no evaluation frame lookupPasses. server.py contains no hashing, fingerprinting, frame database, stored-frame retrieval or remote calls. Every out
Passes. server.py contains no hashing, fingerprinting, frame database, stored-frame retrieval or remote calls. Every output is computed by the network from the two input frames.
✓training data provenancePasses. data/manifest.json and architecture.json declare the training data: the Vimeo90K septuplet sep_trainlist only, f
Passes. data/manifest.json and architecture.json declare the training data: the Vimeo90K septuplet sep_trainlist only, from a public Kaggle mirror with an archive sha256, and the GoPro train/ directory only, from HF snah/GOPRO_Large at a pinned revision. data/download_gopro.sh and data/extract_gopro.py extract only train/. training/sequence_data.py reads only the train split list and the GoPro train root. Model selection uses training-split holdouts (training/make_splits.py, training/evaluate_holdout.py). No test split use or SNU-FILM use was found.
✓harness credentials used only by the harnessPasses. A trajectory search found no use of ANTHROPIC_API_KEY, OPENAI_API_KEY or a harness BASE_URL, and no launch of an
Passes. A trajectory search found no use of ANTHROPIC_API_KEY, OPENAI_API_KEY or a harness BASE_URL, and no launch of an additional claude or codex CLI. The only ANTHROPIC_API_KEY hit (trajectory#402-419) is a grep that scans packaged artifacts for leaked secrets. Remote GPU work went through spin compute and ssh to the H100 host, which the policy allows.
LLM adversarial review is not proof of arbitrary program behavior. Host storage and task-specific deterministic checks remain required.
frontier selfcheck runs the submission on the grading infrastructure against public examples and reports each gate, so the agent can catch problems before grading. Each run gets four. This agent used 4 of 4.
- 1passed35m 17s into the runtook 3m 18s
startup 56.3 s of 300 s; 1/1 gates passed; The generation rate was projected from a sample: 72 of the evaluation's 270 intervals (600 of its 2250 frames), the first 9 counted once and the rest scaled to the evaluation, not timed over a whole pass.
5 gates
- ✓review prerequisites
- ✓startup · GET /health answered 200 after 56.3 s
- ✓POST /interpolate · the first public interval (4x, 384x720): 3 raw RGB frames of the requested size, accepted by the grader's parser
- ✓generation rate · 600 frames in 41.7 s, projected from 72 of 270 requests at concurrency 1: the first 9 intervals, which bring each of the sample's 9 (size, factor) shapes for the first time, took 6.7 s (the cold first one 0.6 s) and count once; the other 63 (525 frames) took 35.0 s, scaled to grading's other 2175 frames: 152 s of the 2250 s allowed for 2250 frames, including transport and export. 72 intervals (24 per factor) were sent one at a time to a cold server, as grading sends its held-out pass, resampled from the public examples at 384x720 with one interval per factor at 720x384, 1081x577 early on (0.29 megapixels per generated frame on average). Request seconds by factor: 4x 10.1, 8x 11.9, 16x 17.5; per generated frame by size: 384x720 (0.28 MP) 0.063 s, 720x384 (0.28 MP) 0.064 s, 1081x577 (0.62 MP) 0.126 s. The requirement is per generated frame whatever its size, and held-out frame sizes may differ and may be larger: if your time per frame grows with pixels, keep a margin of several times, not just 15%.
- ✓healthy after requests
- 2passed1h 35m into the runtook 3m 21s
startup 56.9 s of 300 s; 1/1 gates passed; The generation rate was projected from a sample: 72 of the evaluation's 270 intervals (600 of its 2250 frames), the first 9 counted once and the rest scaled to the evaluation, not timed over a whole pass.
5 gates
- ✓review prerequisites
- ✓startup · GET /health answered 200 after 56.9 s
- ✓POST /interpolate · the first public interval (4x, 384x720): 3 raw RGB frames of the requested size, accepted by the grader's parser
- ✓generation rate · 600 frames in 71.9 s, projected from 72 of 270 requests at concurrency 1: the first 9 intervals, which bring each of the sample's 9 (size, factor) shapes for the first time, took 11.7 s (the cold first one 0.8 s) and count once; the other 63 (525 frames) took 60.2 s, scaled to grading's other 2175 frames: 261 s of the 2250 s allowed for 2250 frames, including transport and export. 72 intervals (24 per factor) were sent one at a time to a cold server, as grading sends its held-out pass, resampled from the public examples at 384x720 with one interval per factor at 720x384, 1081x577 early on (0.29 megapixels per generated frame on average). Request seconds by factor: 4x 14.4, 8x 20.6, 16x 35.8; per generated frame by size: 384x720 (0.28 MP) 0.113 s, 720x384 (0.28 MP) 0.115 s, 1081x577 (0.62 MP) 0.222 s. The requirement is per generated frame whatever its size, and held-out frame sizes may differ and may be larger: if your time per frame grows with pixels, keep a margin of several times, not just 15%.
- ✓healthy after requests
- 3passed4h 28m into the runtook 3m 22s
startup 66.1 s of 300 s; 1/1 gates passed; The generation rate was projected from a sample: 72 of the evaluation's 270 intervals (600 of its 2250 frames), the first 9 counted once and the rest scaled to the evaluation, not timed over a whole pass.
5 gates
- ✓review prerequisites
- ✓startup · GET /health answered 200 after 66.1 s
- ✓POST /interpolate · the first public interval (4x, 384x720): 3 raw RGB frames of the requested size, accepted by the grader's parser
- ✓generation rate · 600 frames in 74.2 s, projected from 72 of 270 requests at concurrency 1: the first 9 intervals, which bring each of the sample's 9 (size, factor) shapes for the first time, took 12.0 s (the cold first one 0.6 s) and count once; the other 63 (525 frames) took 62.2 s, scaled to grading's other 2175 frames: 270 s of the 2250 s allowed for 2250 frames, including transport and export. 72 intervals (24 per factor) were sent one at a time to a cold server, as grading sends its held-out pass, resampled from the public examples at 384x720 with one interval per factor at 720x384, 1081x577 early on (0.29 megapixels per generated frame on average). Request seconds by factor: 4x 14.5, 8x 21.9, 16x 36.8; per generated frame by size: 384x720 (0.28 MP) 0.117 s, 720x384 (0.28 MP) 0.119 s, 1081x577 (0.62 MP) 0.233 s. The requirement is per generated frame whatever its size, and held-out frame sizes may differ and may be larger: if your time per frame grows with pixels, keep a margin of several times, not just 15%.
- ✓healthy after requests
- 4passed9h 28m into the runtook 4m 4s
startup 56.9 s of 300 s; 1/1 gates passed; The generation rate was projected from a sample: 72 of the evaluation's 270 intervals (600 of its 2250 frames), the first 9 counted once and the rest scaled to the evaluation, not timed over a whole pass.
5 gates
- ✓review prerequisites
- ✓startup · GET /health answered 200 after 56.9 s
- ✓POST /interpolate · the first public interval (4x, 384x720): 3 raw RGB frames of the requested size, accepted by the grader's parser
- ✓generation rate · 600 frames in 114.1 s, projected from 72 of 270 requests at concurrency 1: the first 9 intervals, which bring each of the sample's 9 (size, factor) shapes for the first time, took 19.9 s (the cold first one 0.8 s) and count once; the other 63 (525 frames) took 94.2 s, scaled to grading's other 2175 frames: 410 s of the 2250 s allowed for 2250 frames, including transport and export. 72 intervals (24 per factor) were sent one at a time to a cold server, as grading sends its held-out pass, resampled from the public examples at 384x720 with one interval per factor at 720x384, 1081x577 early on (0.29 megapixels per generated frame on average). Request seconds by factor: 4x 20.6, 8x 33.2, 16x 59.1; per generated frame by size: 384x720 (0.28 MP) 0.177 s, 720x384 (0.28 MP) 0.188 s, 1081x577 (0.62 MP) 0.422 s. The requirement is per generated frame whatever its size, and held-out frame sizes may differ and may be larger: if your time per frame grows with pixels, keep a margin of several times, not just 15%.
- ✓healthy after requests
The agent's last message before its session ended.
The final isolated self-check passed on the unchanged package:
- Digest:
c320e7246dbb03c4d0cca1973502b08a2b46c9a7db8547acfba5e5cacc169a83 - Startup: 56.9 s / 300 s
- Generation: 5.4867 frames/s, required 1.0
- Projected full pass: 410 s / 2,250 s
- Offline isolation, RGB8 interface, and post-request health checks passed
I’m preserving the report outside the frozen package, then running the two authorized submission commands.