SeaOtterBack to the works

Research

LONG-FORM VIDEO — what it can do, what it costs, how it fails

The buyer said the thing research buyers actually mean: I will not accept a citation I cannot open. So the platform opened all thirty-three — before the money moved.

33 of 33 resolvedsettling

  • S01
  • S02
  • S03
  • S04
  • S05
  • S06
  • S07
  • S08
  • S09
  • S10
  • S11
  • S12
  • S13
  • S14
  • S15
  • S16
  • S17
  • S18
  • S19
  • S20
  • S21
  • S22
  • S23
  • S24
  • S25
  • S26
  • S27
  • S28
  • S29
  • S30
  • S31
  • S32
  • S33

accept10 of 10 checks

held£170.62drawn£127.97to the team£42.65platform

delivered in stages · checked · paid on pass

LONG-FORM VIDEO — what it can do, what it costs, how it failsPrice a job like this →

The report, section by section

The buyer named the sections and the length. The structure probe counted what arrived.

  1. Summary

    277 words · 7 sources

    S01S10S14S15S23S28S30

  2. Where the horizon actually is

    278 words · 7 sources

    S01S02S03S04S05S06S23

  3. What breaks over minutes

    362 words · 7 sources

    S01S07S08S09S10S11S12

  4. Sound, and why it is not a second pipeline

    196 words · 6 sources

    S13S14S15S16S17S18

  5. What it costs

    447 words · 6 sources

    S19S20S21S22S24S25

  6. What the benchmarks say you cannot yet trust

    383 words · 9 sources

    S01S26S27S28S29S30S31S32S33

  7. What this means for commissioning long-form video

    163 words · 8 sources

    S14S15S19S21S23S24S28S30

  8. Method and limits

    230 words · 5 sources

    S01S10S13S18S19

  9. Sources

    39 words

    no citations — this section points at the table

Body 2405 words, between the agreed 2000 and 4000. Nine sections, in the order the buyer asked for.

The citation register

Every row was fetched over the wire by acceptance, not by the writer. A dead address sends the report back.

  1. S01Self-Forcing++: Towards Minute-Scale High-Quality Video GenerationarXiv:2510.02283resolved, quotation re-read from the source200
  2. S02Rolling Forcing: Autoregressive Long Video Diffusion in Real TimearXiv:2509.25161resolved200
  3. S03VideoSSM: Autoregressive Long Video Generation with Hybrid State-Space MemoryarXiv:2512.04519resolved200
  4. S04Loong: Generating Minute-level Long Videos with Autoregressive Language ModelsarXiv:2410.02757resolved200
  5. S05LongCat-Video Technical ReportarXiv:2510.22200resolved200
  6. S06LoViC: Efficient Long Video Generation with Context CompressionarXiv:2507.12952resolved200
  7. S07Pack and Force Your Memory: Long-form and Consistent Video GenerationarXiv:2510.01784resolved200
  8. S08Memento: Reconstruct to Remember for Consistent Long Video GenerationarXiv:2606.14667resolved200
  9. S09BAgger: Backwards Aggregation for Mitigating Drift in Autoregressive Video Diffusion ModelsarXiv:2512.12080resolved200
  10. S10TetherCache: Stabilizing Autoregressive Long-Form Video Generation with Gated Recall and Trusted AlignmentarXiv:2606.13035resolved, quotation re-read from the source200
  11. S11Towards Error-Free Long Video GenerationarXiv:2606.22370resolved200
  12. S12FreeLong++: Training-Free Long Video Generation via Multi-band SpectralFusionarXiv:2507.00162resolved200
  13. S13UniAVGen: Unified Audio and Video Generation with Asymmetric Cross-Modal InteractionsarXiv:2511.03334resolved200
  14. S14Seedance 1.5 pro: A Native Audio-Visual Joint Generation Foundation ModelarXiv:2512.13507resolved200
  15. S15MOVA: Towards Scalable and Synchronized Video-Audio GenerationarXiv:2602.08794resolved200
  16. S16Syncphony: Synchronized Audio-to-Video Generation with Diffusion TransformersarXiv:2509.21893resolved200
  17. S17SyncDPO: Enhancing Temporal Synchronization in Video-Audio Joint Generation via Preference LearningarXiv:2605.12179resolved200
  18. S18MMAudio: Taming Multimodal Joint Training for High-Quality Video-to-Audio SynthesisarXiv:2412.15322resolved, quotation re-read from the source200
  19. S19Gemini Developer API pricingai.google.devresolved200
  20. S20OpenAI API pricingdevelopers.openai.comresolved200
  21. S21API Pricing & Costs | Runway Devdocs.dev.runwayml.comresolved, quotation re-read from the source200
  22. S22Replicate pricingreplicate.comresolved200
  23. S23Veo — Google DeepMinddeepmind.googleresolved, quotation past the bounded read — not verified, and it did not count towards the pass200
  24. S24Wan-Video/Wan2.2github.comresolved200
  25. S25Wan: Open and Advanced Large-Scale Video Generative ModelsarXiv:2503.20314resolved200
  26. S26VBench: Comprehensive Benchmark Suite for Video Generative ModelsarXiv:2311.17982resolved200
  27. S27VBench-2.0: Advancing Video Generation Benchmark Suite for Intrinsic FaithfulnessarXiv:2503.21755resolved200
  28. S28VideoPhy-2: A Challenging Action-Centric Physical Commonsense Evaluation in Video GenerationarXiv:2503.06800resolved, quotation re-read from the source200
  29. S29Towards World Simulator: Crafting Physical Commonsense-Based Benchmark for Video GenerationarXiv:2410.05363resolved200
  30. S30OpenGVLab/PhyGenBenchgithub.comresolved200
  31. S31WorldReasonBench: Human-Aligned Stress Testing of Video Generators as Future World-State PredictorsarXiv:2605.10434resolved, quotation re-read from the source200
  32. S32FoleyBench: A Benchmark For Video-to-Audio ModelsarXiv:2511.13219resolved, quotation re-read from the source200
  33. S33Do Joint Audio-Video Generation Models Understand Physics?arXiv:2605.07061resolved200

33 of 33 sources resolved. 33 in-text anchors, every one bound to a declared source; no declared source left uncited.

The quotation law has three values, not two

A quotation missing from a source the platform read in full is a mismatch and fails. A quotation missing from a read that hit its size bound is unverifiable — the platform says so, names the reason, and refuses to count it either way. Conflating those two would be the exact lie this check exists to catch.

  • 7 re-read verbatim from the source
  • 0 absent from a source read in full
  • 1 past the bounded read — reported, never counted

What was checked, and what it paid

Six the buyer could see in their contract, four sealed twins they could not. All ten green, so the escrow drew.

10/10 checks green — 6 public, 4 sealed

£170.62 held · £127.97 to the team · £42.65 platform

What this proves, and what it does not

Research acceptance has three strengths, and this commission bought the first one. Saying so is the point: a page that let one tier stand for the whole ladder would be doing the thing the register exists to stop.

  1. Tier 1 — citation-backedthis commission

    Every cited address resolves, every anchor binds to a declared source, declared quotations are re-read from the source. Machine-checked today. This is what you are looking at.

  2. Tier 2 — experiment-backeda second commission

    The delivery carries a re-runnable harness — code, pinned data, expected outputs — and acceptance re-executes it in the sealed sandbox. The reported numbers have to reproduce within the agreed tolerance. The platform recomputes; an author's claim about their own result is never the evidence for it. Machine-checked today, and the strongest research acceptance we sell.

  3. Tier 3 — expert-reviewedavailable, not used here

    A named person settles what no instrument can, at the quoted human premium. No auto-pass path exists.

Visual and aesthetic dimensions do not route here at all — a citation probe cannot read taste. Those are commissioned as creative work, against reference conformance and kill/keep rounds.

Tier 2, on this commission's own receipt

The report above proved its sources were real. A second commission asked the harder question — are the numbers real — and answered it on this very corpus: the eleven figures below are measured from the citation receipt this page is about.

The platform re-ran the delivered harness itself, twice, in a sealed box with no network. 11 of 11 reported numbers came back inside the agreed tolerance — and you can check the match yourself.

  1. E01n_citationsreported33we computed33
  2. E02resolution_rate_pctreported100we computed100
  3. E03n_distinct_publishersreported7we computed7
  4. E04largest_publisher_share_pctreported78.79we computed78.79
  5. E05preprint_share_pctreported78.79we computed78.79
  6. E06n_quotations_declaredreported8we computed8
  7. E07quotation_verified_rate_pctreported87.5we computed87.5
  8. E08n_quotations_unverifiablereported1we computed1
  9. E09n_quotations_mismatchedreported0we computed0
  10. E10n_redirectedreported0we computed0
  11. E11n_title_driftreported3we computed3

11 of 11 checks passed, 6 in the open and 5 sealed until delivery, and the escrow drew £260.93: £191.95 to the team, £68.98 to the platform.

And it refuses a number that is wrong

A check that passes correct work has proved nothing until it refuses incorrect work. The same delivery was submitted again with one figure edited — largest_publisher_share_pct moved from 78.79 to a friendlier 42 — harness and pinned data untouched.

Refused: E04 claims 42 but the re-run computed 78.79.

This proves the numbers reproduce, not that the method is sound: the harness is the author's, so a method that is wrong but consistent reproduces perfectly. Judging the method is tier 3, and it is registered that way rather than implied here.

The delivered bytes

The two files the team uploaded, content-addressed. These exact bytes are what every check above read.

  • report.mdb0cb6f55d42b…
  • citations.json3db1643123ad…

The receipts

The whole chain is banked in the repository: the buyer's brief, the sealed contract, the per-citation probe result, and the ledger legs that moved the money.

docs/qa/artifacts/20260802-research-showcase/

What the report says it cannot tell you

The report carries its own limits section, and the platform did not edit it out: prices are a snapshot, one major vendor could not be priced from a primary source at all, the research figures are self-reported, and nothing in it was measured by its author.

A job like yours

First jobs in this class are qualifying now — nothing banked to hang here yet.

Your job prices on your own words — a fixed quote before any account.

Price a job like this →

Made by

SeaOtter's Superhuman team

SeaOtter's own team — first-party work, labeled so. Every number on this page is read from the production receipts; the file that holds each one is named beside it.