Evaluation and media provenance

Benchmarks, task datasets, attribution, metadata, interoperability and quality review

• articleWhat does the enrolled Tennessee Public Chapter 588 (ELVIS Act) say about voice, simulated voice, tools and the fair-use list?
  The enrolled ELVIS Act (effective 2024-07-01) contains the narrowed tool clause ('a particular, identifiable individual's ... voice, or likeness, with knowledge'), the voice-includes-simulation definition, the First Amendment fair-use list and the 'reasonably should have known' standard; it confirms that the House amendment wording read in CA-B194 is the enacted text.
• articleDoes the Ofcom Broadcasting Code apply to the video or programme I am making, and what does Rule 2.2 mean in plain terms?
  The Broadcasting Code applies to radio and TV content on Ofcom-licensed services (plus BBC licence-fee services and S4C) and binds broadcasters through their licences; notified on-demand providers fall under separate rules; Ofcom regulates accuracy as such only in news (Section Five), and 'materially misleading' depends on context and potential harm or offence. Whether it reaches an independent creator's online video is not stated in the Code read.
• articleWhere does the '32 characters, 3 lines' closed-caption rule in a DCP come from, and what do SMPTE ST 429-5:2023 and RDD 52 (the Bv2.1 profile) say about closed-caption packaging?
  The DCP closed-caption limit of 32 characters per line and 3 lines per instance comes from SMPTE RDD 52:2020 (the DCP Bv2.1 Application Profile, a Registered Disclosure Document expressly not a Standard), which notes some languages need fewer characters (e.g. 14 or 16) and also says caption XML should not exceed 256 kB, forbids MainCaption and ClosedSubtitle and requires 48 kHz audio. ST 429-5:2023, the packaging layer, defines closed captions as text for a non-screen device and sets no character or line limits.
• articleWhat changed in the WGA's AI article (Article 72) between the 2023 MOA and the 2026 MOA, according to the two MOA texts?
  The 2026 MOA (April 4, 2026; term May 2, 2026 to May 1, 2030) amends only Paragraph G of Article 72: it keeps semi-annual meetings and adds written notice to the Guild when a company grants a compensated licence to a non-affiliated third party to train a public-facing commercial GAI system, with a meeting on request. Paragraphs A-F (AI not a writer, disclosure, consent) are not amended in Item 15, and the WGA's AI page still describes the 2023 provisions.
• articleWhat subtitling, audio description and signing percentages does Ofcom's Code on Television Access Services set for UK-regulated TV channels, and how are channels and affordability decided?
  Ofcom's 2024 access code sets statutory ten-year targets of 80% subtitling (90% Channel 3 and 4; 100% BBC), 10% audio description and 5% signing as percentages of non-advertising programme hours per delivery platform, with interim targets by anniversary (e.g. year 1 10/2/1%, year 5 60/10/3%); Ofcom's July 2025 notice lists 83 channels required in 2026 (82 at Level One). The Code does not cover on-demand services (producer reading of the 2024 statement).
• articleWhat numbers does Ofcom's Code on Television Access Services (updated 20 Nov 2019) set for audio description and subtitling, and which of them are binding targets versus best-practice guidance?
  In the 2019 edition of Ofcom's access code, binding annual percentage targets rise to 10% audio description, 80% subtitling and 5% signing, while numeric production guidance (subtitles normally at most 160-180 wpm for pre-recorded programmes, live delay at most 3 s, two lines unless three do not obscure the picture) sits in Annex 4 as guidance to 'have regard to'. This edition is superseded by the 2024 Code (C2-B61-K5-06).
• articleDoes the brain treat continuity cuts as a common event, and do cut types differ?
  An open-access EEG study (21 viewers) reports neural responses common to continuity cuts (theta synchronisation, delta desynchronisation) interpreted as encoding and segmentation; differences by cut type appeared under one taxonomy but not another, where results matched random groupings. Only continuity cuts were tested, so it says nothing about jump cuts or axis breaks.
• articleDoes crossing the 180-degree line confuse viewers or reduce how much they like a video?
  In three experiments with sketch-style videos, viewers noticed 180-degree violations (flagging .42 of critical cuts vs .08 without violation), were slightly more confused (about 8% vs 6% of critical cuts), and liked violated videos just as much (preference .519, Bayes factor favouring no effect). This supports 'detectable and mildly confusing' but not 'harms enjoyment'.
• articleWhich field in a Smithsonian Open Access (EDAN) record tells me an item is CC0?
  In sampled Smithsonian open-access records, the CC0 designation sits at content.descriptiveNonRepeating.metadata_usage.access for edanmdm records but at content.metadata_usage.access for ead_component records, and media carry their own online_media media[].usage.access; every sampled value was CC0 (curator re-parse reproduces the producer counts exactly). Agents must check the path by record type.
• articleWhat exactly does the Openverse images API return for licence, attribution, provider and source, and what do its docs and terms say about relying on those fields?
  At schema version 639e9266 an Openverse image record carries license, license_version, license_url, creator, provider, source, foreign_landing_url and a generated attribution; provider and source can differ (a live Smithsonian record shows provider 'smithsonian' and source 'smithsonian_national_museum_of_natural_history' with no creator), license_url and attribution are generated, and Openverse does not verify licences.
• articleI have an asset search result from Openverse, Smithsonian, Europeana or Wikimedia Commons: which field tells me its licence, which tells me what credit to give, and which fields must I not treat as final?
  Openverse exposes flat licence fields (license, license_version, generated license_url and attribution) but says it does not verify licensing; Commons extmetadata exposes LicenseShortName, UsageTerms, AttributionRequired, Attribution and Restrictions, with values unreliable for multi-licensed files and License only a best guess; the Smithsonian marks CC0 per item and says CC0 items need no attribution. Each provider tells reusers to verify.
• articleWhat does a major stock licence say about editorial content, model and property releases, and use of content as a trademark?
  The Getty Images EULA says editorial content is not model or property released, makes the licensee responsible for third-party approvals before commercial use, warrants publicity/privacy and trademark only where Getty states a release was obtained (and only for content as delivered), bars use as the distinctive feature of a trademark or logo, and carries a separate no-machine-learning/AI/biometric restriction.
• articleFor reference images of logos, branded props, museum interiors or artworks, which non-copyright restrictions does Wikimedia Commons say remain with the reuser?
  Commons treats trademark, personality, moral-rights and similar restrictions as separate from copyright and leaves compliance to reusers; its warning templates are a courtesy whose absence does not mean no obligation, and venue house rules (e.g. museum photo bans) normally do not change a photographed work's copyright status. The CC FAQ likewise says CC licences do not license trademarks.
• articleWhat does Wikimedia Commons say about using photos of identifiable people (consent, public versus private places, free licences, model releases)?
  The Commons guideline separates the photographer's copyright from subjects' personality rights (publicity and privacy): private-place and commercial uses usually require consent, consent to take a photo is not consent to publish or reuse, rules vary by country, a model release is more permissive than Commons requires, and a free licence covers only the photographer's rights. It is community guidance giving general advice.
• articleIf a reference photo or logo image carries a Creative Commons licence, does that licence also cover the people, trademarks or property shown in it?
  CC BY 4.0 licenses copyright and similar rights only: moral, publicity, privacy and personality rights and patent and trademark rights are not licensed, and the licensor waives only rights it holds, as far as needed for the licensed use. Rights held by others (a person pictured, a brand shown) are unaffected, CC does not guarantee model releases, and licensors are advised to mark third-party rights.
• articleHow do vrecord and QCTools document capturing and checking videotape transfers, and what do they record as fixity and metadata?
  vrecord captures via MediaArea's ffmpeg-ma with Blackmagic cards (or AVFoundation for DV on macOS) to FFV1 version 3 with GOP 1 and slice CRCs, writes a framemd5 and logs by default (QCTools XML highly recommended; frame MD5s catch some dropped frames), and checks output against a MediaConch policy (Matroska >= 4, FFV1 >= 3.4, GOP 1). QCTools stores per-frame signal metrics (signalstats, cropdetect, psnr, ssim, ebur128) in a report for review. Neither is a standards conformance tool.
• articleWhich WER variants and text normalisations change ASR subtitle scores, and is COMET usable for subtitle translation evaluation?
  WER on connected speech is asymmetric and can exceed 1 (bound max(N1,N2)/N1), which motivated MER and WIL; Whisper's authors report text normalisation cutting WER by up to 50 percent on some sets and warn of overfitting to Whisper's style; SubER's authors found case-sensitive WER with punctuation did not consistently improve correlation with post-editing. IWSLT 2024 ranked offline speech translation by COMET but subtitling by SubER, with BLEU/chrF/BLEURT for translation and CPS/CPL/LPB compliance rates.
• articleIs IMF an archive format for finished films, what do archives and the Library of Congress say about IMF, DCP and MXF as wrappers, and what remains unclear?
  LoC's RFS lists IMF (MXF essence, CPL, packaging XML) first for file-based final video; the Academy forum describes IMF as a wrapper in significant industry use, notes an LoC MXF archiving subset (SMPTE RDD 48) and IMF App 5 ACES (ST 2067-50); FIAF 2019 found only CNC and LoC requesting a mezzanine on deposit, with CST recommending IMF (ST 2067-40) as mezzanine. Evidence of IMF as an actual long-term archive master remains thin.
• articleFor a born-digital feature, what do film archives and the Library of Congress say should be deposited: DSM, DCDM or an unencrypted DCP?
  FIAF's Technical Commission (2010) accepts only a DCDM or an unencrypted DCP for long-term preservation: the DCDM is ideal but may never exist as a complete set, while the unencrypted DCP is lossy but shows the work exactly as audiences saw it. FIAF's 2019 survey counted unencrypted DCP deposit at 9 archives and DCDM at 7, and LoC's 2025-2026 RFS prefers an unencrypted interop DCP for motion pictures, asking depositors to contact it about DCDM/DSM.
• articleWhat do archives and tools document about using BagIt with OAIS/PREMIS for audiovisual collections, and what went wrong in practice?
  Archivematica documents AIPs as BagIt bags containing METS, but its 1.13 page gives inconsistent manifest algorithms across its own examples. Indiana University's MDPI relied on Matroska per-element CRCs written by FFmpeg (about 100 bytes unprotected), packaged film as BagIt tarballs (DPX, WAV, metadata) with a custom multithreaded verifier, and reported that a vendor bagit.txt reading 'BagIt-version' made them re-handle over 2 PB. Per the L07 correction on C2-B10-G2-07, ABNF quoted strings are case-insensitive, so this shows IU's tooling or acceptance criterion, not a proven RFC nonconformance.
• articleHow do I score machine-generated subtitle files (not just text) against a reference, what does SubER measure, and how far has it been validated?
  SubER scores a subtitle file against a reference including text, line breaks and timing (SRT input; Apache-2.0 code); its authors give a rule of thumb (<20% very good, >40-50% bad). In IWSLT 2024 human judgments for translation-focused DA agreed more with BLEURT than SubER (producer reading), so its validation depends on the evaluation target.
• articleWhat do reviews say about the limits of marker gap filling methods, including learned methods, for motion capture used in animation?
  A CC BY 2025 Sensors review reports learned mocap recovery overfits benchmarks (CMU, HDM05) and degrades on long gaps and new subjects, clothing or marker sets, while classical interpolation is simple and interpretable but relies on manual input; it cites tree-based regression for long gaps. Method-level claims are secondary.
• articleDo biomechanics sources agree on how to choose a Butterworth cutoff, and does residual analysis give one answer?
  On eight data sets, four automatic cutoff-selection methods (Challis, Yu, Winter, Jackson) gave widely different optimal cutoffs for the same trial (walking toe x: 12.5, 3.0, 10.0, 4.0 Hz); the effect depends on the parameter (minimum toe clearance varied under 1 mm), and cutoffs above 20 Hz were judged inappropriate for force-plate data. Applying these biomechanics results to animation capture is a proposal.
• articleWhat does a mocap vendor document for filling marker gaps and choosing a smoothing cutoff, including limits on how much gap to fill?
  OptiTrack Motive recommends trimming about 3 frames at gap ends, filling only gaps below a user-set Max. Gap Size (large gaps not recommended), smoothing with a 4th-degree bidirectional low-pass Butterworth at a commonly used 7-12 Hz (higher for fast, sharp motion), and retaking when a significant marker is occluded; edits change raw trajectories and overuse is discouraged.
• articleWhat do public summaries of the 2023 SAG-AFTRA TV/Theatrical AI provisions say about digital replicas of voice and likeness, synthetic performers and consent, and what primary text is unavailable?
  A law-firm summary (2023-11-13, before ratification) describes the 2023 SAG-AFTRA TV/Theatrical AI terms: consent requirements for employment-based and independently created digital replicas and digital alteration (arbitration, monetary remedies), and notice/bargaining for synthetic performers. Union primary texts were not readable.
• articleWhat does the Tennessee ELVIS Act text say about voice, including simulated voice and voice-cloning tools, and how does the adopted amendment differ from the bill as introduced?
  The ELVIS Act amendment defines protected voice to include a simulation, creates liability for knowingly making an unauthorised voice/likeness available and for distributing tools whose primary purpose is producing a particular, identifiable individual's voice or likeness (narrowed from the introduced text's 'an individual's ... without authorization'), lists First Amendment fair uses, and lets exclusive recording-artist contract holders sue.
• articleWhat does Article 50 of the EU AI Act say about marking synthetic audio and disclosing audio deep fakes, who is addressed, and what exceptions does the text state?
  As reproduced on a secondary site, AI Act Article 50 requires providers of systems generating synthetic audio/image/video/text to mark outputs machine-readably (except assistive standard editing), requires deployers to disclose deep fakes, and limits that disclosure for evidently artistic or fictional works to a form that does not hamper enjoyment.
• articleHow are licences embedded in image and media files (XMP xmpRights, Creative Commons XMP, IPTC rights fields), and what does documentation say survives common handling?
  An image can carry licence information in parallel XMP places (dc:rights, xmpRights:UsageTerms/WebStatement/Marked, IPTC Extension and PLUS licensing fields, cc:license/morePermissions); no text read says which wins on disagreement; IPTC states copyright management metadata must never be removed, yet its 2019 platform tests (reported by IPTC) show several sites strip metadata from saved or downloaded files.
• articleWhat licence do the Valentini/Voicebank-DEMAND item and the OpenSLR pages for MUSAN, LibriTTS, LibriTTS-R and Hi-Fi TTS state, and what do they say about training?
  OpenSLR pages for MUSAN, LibriTTS, Hi-Fi TTS and LibriTTS-R each state CC BY 4.0 at page level (byte-identical to the producer pins).
• articleWhich RightsStatements.org statement fits a cultural-heritage item, and which of the 12 assert a determination versus an unknown status?
  RightsStatements.org offers 12 statements (In Copyright, No Copyright, Other); its page says all but 'the two statements for objects with an unclear copyright status' should be applied only after copyright status is established, while the Other group lists three statements (CNE, UND, NKC), so NKC's status under that sentence is unclear.
• articleIf a record or old dataset carries a www.europeana.eu/rights/... rights URI (rr-f, rr-p, unknown...), does Europeana still accept it, and what should an agent do?
  Europeana retired its own www.europeana.eu/rights statements on 2016-11-11 in favour of rightsstatements.org equivalents (Paid Access retired outright); legacy URIs may still appear in old data, and Europeana code still recognises them (producer, code not re-read), but no source read says they are mapped automatically.
• articleWhere is the Comfortaa font that the YarnSpinner-Unity notices mention, and why is it listed under 'Samples:'?
  The Comfortaa OFL notice in YarnSpinner-Unity has no Comfortaa-named file in the core tree at 90fb4c46 or in the YarnSpinner-Unity-Samples tree at 13f258df; the samples tree ships other fonts (four .ttf/.otf files) with their own licence files (producer).
• articleWhich binaries or source folders in YarnSpinner-Unity at 90fb4c46 correspond to each entry in its Third-Party Notices file?
  Each headed library notice (ANTLR4, Protocol Buffers, CsvHelper) has one Yarn.-prefixed DLL in Runtime/DLLs; three shipped analyser binaries in Editor/Analysis/DLLs (Microsoft.CodeAnalysis, Microsoft.CodeAnalysis.CSharp, System.Collections.Immutable, System.Reflection.Metadata) have no notice entry by name, and the notices file lists Microsoft/System names with no DLL in the tree.
• articleWhat third-party notices and licence statements ship with YarnSpinner-Unity at the pin used?
  YarnSpinner-Unity at 90fb4c46 ships a Third-Party Notices file separate from its MIT LICENSE.md, naming ANTLR4 and Protocol Buffers (BSD 3-Clause), CsvHelper (Apache 2.0), a Microsoft/System package list (MIT, .NET Foundation), Unity-SerializableDictionary (MIT) and, under Samples, the Comfortaa font (OFL 1.1, RFN Comfortaa).
• articleWhat do the OS/2 fsType bits mean for embedding, and what do fontTools and fontbakery report or enforce about them?
  fontbakery v1.1.0 requires fsType 0 in its googlefonts profile and 8 in its microsoft profile (profile-specific, not a contradiction in the spec). No file in the fontTools 4.67.0 subset package mentions fsType or embedding, so the subsetter does not consult bit 8 (no subsetting) by code reading.
• articleWhere does Google Fonts record which licence a font family uses, and what are the possible values?
  In google/fonts at 45e3d526 the top-level directory indicates the licence of the files within it, and the METADATA spec (gf-docs 8ad6ca1e) lists APACHE2, OFL and UFL as license values.
• articleWhat does the W3C imsc-tests repository contain (TTML source files and reference PNG renderings) and under what licence text can its files be used?
  w3c/imsc-tests (14dfb678) licenses its documents under both the W3C Test Suite License (which grants no right to modify and says tests may not be altered) and the W3C 3-clause BSD licence, at the licensee's choice, unless a document carries another compatible licence; the imsc1 README attributes 58 TTML files to IRT tests under an Apache-2.0 licence and 196 to the W3C TTML and IMSC implementation reports.
• articleWhat rights information does the Europeana Search API expose, and how does its reusability filter relate to rights URIs?
  Europeana's Search API documentation (archived 13 Apr 2024) offers a reusability filter whose 'open' bucket covers PDM, CC0, CC BY and CC BY-SA and whose 'restricted' bucket covers NC and other restricted statements, with a 'permission' bucket for the rest; the API requires a registered key.
• articlePROPOSAL on 1916 practice: which failure modes does Sargent name for motive and plausibility, which are claims and which examples, and what check could an agent run on an outline using only those?
  From Sargent's ch. X, XII and the passages they point to (ch. XI, XXXIII, XLVI) I list ten failure modes: unexplained action, a plot-needed event with no cause, motive held to the denouement, an unreasonable motive, a possible but improbable incident, an extraordinary premise offered as fact, convenient arrivals, overworked conveniences, an unshown enabling trait, and a plant shown too openly. Each is a Sargent claim or a Sargent example, labelled as such. The outline check below is the worker's proposal that applies those 1916 rules; Sargent gives no test and nothing here was run. The check is not a validated detector and the 'average spectator' standard is a 1916 US editor's.
• articleWhat is the simplest test of whether a character agent refuses facts that the character could not know, and what numbers have been reported with it?
  Two papers use a multiple-choice test with an explicit refusal option. KBF-QA (arXiv 2606.25632v1): for a named character, each question has four answers plus 'I cannot answer this from my own knowledge'; items whose fact the character could witness have a normal gold answer, items it could not witness have the refusal as gold; 4,386 items over eight novels. CHARM (arXiv 2609.01352v1): separate 'is this outside the character's knowledge' yes/no questions and abstention-enabled questions, 40 characters, 2,748 questions. Reported: retrieve-everything RAG recalled 63.8 to 66.4 and refused only 8.7 to 11.3 on KBF-QA; the visibility-bounded ReverieMem refused 81.2; on CHARM GPT-4o recognised the boundary 91.3 percent of the time but still answered in 72.1 percent of cases.
• articleWhich licence terms are recorded for the Unity Recorder package components (source code, other package content, third-party GIF encoder) and what does the Unity Companion License v1.4 text say about use?
  The package's LICENSE.md (copyright 2022 Unity Technologies) says source code is under the Unity Companion License (UCL) "otherwise" under the Unity Package Distribution License (UPDL); Third Party Notices lists gif.ski as a commercial licence (Kornel Lesinski). The UCL v1.4 text served by unity.com on 2026-10-22 grants copyright rights only in connection with authoring/distributing content under a valid Unity Engine License, bars competitive analysis and competing products, and assigns derivative works of the Work to Unity. UPDL text and gif.ski terms were not read. No legal clearance is concluded.
• articleDoes any openly readable on-set VFX data template record who captured each item (operator/person) and whether it was checked or unverified, and if not what is the nearest documented substitute?
  Negative result with partial analogues. In the VES Camera Report spec the only person field is slate-level 'Wrangler' (person entering data); the VES On-Set VFX Data Guide (v1.1.0 file) assigns a Creator role per data item (role names, not persons). Across the pinned VES texts read there is no verified/unverified or checked field; the nearest items are the 'Under Protest' VFX Work flag, the guide's Clips Report 'data confirmation' and the timestamped virtual-production scene snapshot. A synthesis proposal adds 'Tk Captured By' and 'Tk Verification Status' columns, which the spec's own extension rule and sample reader would carry (untested).
• articleIf I run MEt3R, which reconstruction backbone do the paper and the repository say it uses, and which one do I get by default?
  The paper text (arXiv v1 and v2) names DUSt3R as MEt3R's reconstruction model and says MASt3R was not used. The released code at commit ee0e175 defaults to a 'mast3r' backbone (MASt3R weights) and offers 'dust3r' and 'raft' as options, and the README overview still says DUSt3R. So a default run is not the configuration the paper text describes; to match the paper wording pass backbone="dust3r". Whether the paper's tables were produced with DUSt3R weights is not stated by any source read (see candidate 02).
• articleWhat exactly do Wayve's GAIA-2, GAIA-3 and GAIA-4 pages claim about fidelity, controllability and correlation with real driving, and which of those claims come with a value?
  Across the three company pages the numeric values are few and are about scale or comparison, not about fidelity or correlation: GAIA-3 gives 15 billion parameters, a tokenizer twice GAIA-2's size, about five times the compute and ten times the data over 9 countries on 3 continents; GAIA-4 gives one number, a 2.5x improvement in how faithfully the recorded world is preserved, with no metric or baseline named in the page text. The claims that matter for evaluation (GAIA-3 predicts relative policy performance; GAIA-4 outcome, closed-loop and component fidelity) are stated without coefficients, sample sizes, policy counts or tables. The GAIA-2 page has no metric values; its numbers live in the arXiv paper (FDD, FID, FVMD, agent IoU as training curves).
• articleWhere do the Vimeo-90K, Flickr2K and OST images come from according to their own authors, and which third-party terms might therefore sit underneath them?
  Vimeo-90K: 89,800 clips cut from videos downloaded from vimeo.com (project page); the README says 5,846 selected videos, a subset of the AoT dataset. Flickr2K: collected by the EDSR authors with the Flickr API. OST: images from search-engine queries by category with ImageNet images as backgrounds, per the SFTGAN paper. In each case the authors' pages state origin, not the terms of the underlying source; those terms (Vimeo, Flickr, search-engine results, ImageNet) were not examined in this batch.
• articleWhy does hiding the threat (darkness, obscurity, withheld information) make horror scarier, and is that measured or craft lore?
  Three openly readable craft texts (Burke 1757, Radcliffe 1826, Lovecraft) argue that fear rests on the unseen or the unknown, and Radcliffe adds a useful limit: obscurity must leave the imagination something to work with, while confusion does not. The only measured support found is indirect: a non-film shock-learning experiment where stress responses peaked at maximal uncertainty. No open-access study was found that tests darkness or monster-reveal timing in a film, so those claims remain craft lore.
• articleWhat do the RVM, MatAnyone and MatAnyone 2 papers report on video-matting benchmarks (MAD, MSE, Grad, Conn, dtSSD), how do synthetic and real-video numbers differ, what limits do they state, and what are the code and weights licences?
  RVM (arXiv v1, GPL-3.0 code) is auxiliary-free human matting and states it fails when people are in the background. MatAnyone (v2) and MatAnyone 2 (v2) are mask-guided: they take a first-frame mask, MatAnyone states it has the best MAD and dtSSD on every synthetic benchmark block (not every MSE or Grad), and it says the older VideoMatte240K ground truth lacks fine detail such as hair. Numbers do not transfer between benchmarks: MatAnyone reports core-region MAD 0.14 on a real-video benchmark with segmentation-mask ground truth, while MatAnyone 2 reports MatAnyone at MAD 5.76 on 19 real videos with manually annotated alpha every 10 frames. Even the newest model has 15.2% of boundary-band pixels flagged unreliable by the paper's own learned evaluator, whose checkpoint is not released at the pinned README. MatAnyone and MatAnyone 2 code carries the non-commercial S-Lab licence.
• articleWhat J&F accuracy does the SAM 2 paper report for video object segmentation, where does it say SAM 2 fails, which checkpoint does the public README actually describe, and what licence covers each component?
  Paper v2 reports SAM 2 (Hiera-B+) J&F of 76.6 on MOSE val, 90.2 on DAVIS 2017 val, 78.0 on LVOS val and about 77 on SA-V val/test, ahead of Cutie-base+ (71.7, 88.1, -, 61.3/62.8) when prompted with the first-frame ground-truth mask. The authors list failures: shot changes, crowded scenes, long occlusions, extended videos, and thin or fine details when fast-moving; they recommend refinement clicks on later frames. The public README's SAM 2.1 per-checkpoint numbers (large: SA-V test 79.5, MOSE val 74.6) match the paper's appendix rows marked as trained without the VOS datasets, not the headline row (MOSE 77.9), which is a documented difference an agent must not collapse. Licences: checkpoints, demo and training code Apache-2.0 (README), SA-V dataset CC BY 4.0 (paper), a third-party connected-components file carries its own licence.
• articleHow do I compare two characters' idiolects in a script using frequent-word rates, the way Burrows did for Jane Austen?
  Buckland (2019/2020 file) retraces Burrows's 1987 method on a screenplay: extract each main character's dialogue, count rates per 1000 words of frequent words (Burrows used the 30 most common word types), then compare characters pairwise by correlation or by individual words ('the', 'of', 'I', 'not'; 'we', 'us', 'our'). In Burrows's Austen data character pairs correlate above 0.7 on those 30 words; weaker correlations (e.g. 0.930 Darcy-Elizabeth vs a weak Collins-Lydia pair) mark more different idiolects. Buckland's screenplay run is exploratory, uses a web tool and one film, and does not report a full pairwise matrix for the brothers.
• articleWhich craft elements did professional writers say failed in Dramatron output, what did they say about authorship and ethics, and what are the paper's own limits?
  Participants reported logical gaps and weak long-range consistency, no subtext (dialogue 'telling'), characters without motivation or complete arcs, stereotyped and biased content, repetition loops, and scene-to-scene dialogue inconsistency caused by parallel generation. Over half of survey respondents felt they did not own the result; the authors frame it as inspiration/provocation, raise plagiarism, bias and labour-impact concerns, and propose (untested) future work on character arcs and hierarchical dialogue.

More