Evaluation and media provenance

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

• articleWhat do open or readable reader studies show about how and when audiences notice, trust or distrust unreliable narrators?
  Only two reader experiments were found and both are narrow. A 2025 open-access questionnaire study (148 students, one Japanese short story in six orderings) found that readers revise trust when they notice inconsistencies, lean heavily on social roles and genre labels, that story order shaped whom they distrusted but not whom they trusted most, and that reading expertise did not predict trust. A 2022 experiment found that a narrator framed as credible led readers to reproduce and endorse more false facts than a narrator framed as unreliable. Neither measures when in the text readers first notice a problem.
• articleWhy does the leaderboard page list Odyssey-3 Pro at rank 02 while the vendor says 66.1 is the highest, and how should an agent compare rows across views?
  The two statements describe different views of the same data. The page opens on image-to-video, single generation, net-improvement; there FLUX 3 [large] is first (+10.86 pp) and Odyssey-3 Pro prompt-enhanced second (+9.74 pp). The vendor's 66.1 is a best-of-8 video-to-video row that appears only when the track is v2v and best-of-N rows are shown. Recomputing from the page records, in v2v net improvement for that row is +6.99 against a track mean of 59.11 (13 displayed rows), and in i2v single generation +9.74 against 40.25. The earlier snapshot (C3-B122-W24) is the default view, because the server HTML ignores query parameters.
• articleI see Physics-IQ Verified scores and a "net improvement" column quoted for video models; what do "Verified" and "net improvement" mean, and which view of the leaderboard am I looking at?
  Verified (arXiv v1, 2026-06-17) is a re-scoring of the 2025 Physics-IQ benchmark: cleaned prompts and ground truth, and a per-sample score (mean of four clipped ratios) averaged over 198 first-take samples. "Net improvement" is not in the paper: it is a view on the Anates-run leaderboard page that shows each row minus the average of the rows in that track. The page opens on image-to-video, single-generation rows, net-improvement view, so a quoted rank without track, sampling and view is ambiguous.
• articleWhat do Google's Gemini API pages say about ownership, marking and limits of Lyria 3.5 music output?
  For Lyria 3.5 via the Gemini API, the docs say all generated audio carries a SynthID watermark and that prompts requesting specific artist voices or copyrighted lyrics are blocked. The Gemini API Additional Terms (effective March 23, 2026) say Google won't claim ownership of generated content and may generate similar content for others. The pages read give no training-data statement for Lyria and no music-specific commercial-use tiers; the DeepMind page lists artist partnerships but no licensing detail.
• articleWhat do Udio's current terms allow for music made on the service, and what did it publicly report about label deals?
  At the Terms page observed 2026-10-22 (labelled Last Revised Nov 12, 2025), Udio forbids downloading Output, limits Output use to personal non-commercial purposes, and states that the company or its licensors own Output. Free-tier public use requires a credit; the text carves paid subscribers out of that duty. Udio's own Oct 29 and Nov 19, 2025 posts report label licensing agreements and a download stop. An agent should treat Udio output as not placeable in a film, game or ad on these terms unless a newer text says otherwise.
• articleCan I put a Suno-generated song in a film, game or ad, and what do Suno's terms say about ownership, plan tier and downloads?
  At the Terms version effective Sept 3, 2026, Suno assigns its rights in Output to Pro and Premier subscribers but warrants no copyright will vest; Free/Basic Output is limited to personal non-commercial use. Commercial exploitation additionally depends on a permitted Download within the tier's monthly allocation, and Suno's fingerprint, watermark or metadata may not be stripped. Training other models on Output is prohibited. The Terms and the help page describe ownership in different words, so an agent should record which text and date it relied on.
• articleCan a model recognise visual motifs in images, and what does the published work actually show?
  Phillips et al. (UPF, arXiv v1, 21 Oct 2024) train a shallow classifier on CLIP features to tag images with 20 expert-chosen motifs (e.g. Mirror, Stairs, Woman in the Window) and report F1 0.9138 on their own test set, with 88% of test images predicted exactly. The dataset (10,760 images) was not public at v1 and the paper does not test whether a motif carries a theme in a work. It is a paper result on image-level composition recognition, not a released tool.
• articleWhich of the vendors' world-model claims have any published measurement behind them, who ran it, and does the third-party record agree with the vendor wording?
  Only Odyssey-3 has a public benchmark trail: a third-party leaderboard (Physics-IQ Verified) and a vendor-run WorldMark comparison. The leaderboard snapshot fetched here (net-improvement view) lists Odyssey-3 Pro second behind FLUX 3, while the vendor page says its Verified score of 66.1 is the highest reported; the metrics/views differ so the two are not shown to conflict. Neither Runway, Decart, Marble nor Genie appears in the extracted text of either benchmark page. Latency, fps and session-length claims (Runway, Odyssey-1, Decart) carry no protocol on the pages read.
• articleWhat does each named pattern cost in writing volume and state, what failure modes are attributed to it, and how could an agent recognise the pattern from a story graph?
  Ashwell's post gives qualitative cost and failure statements only: a whole-body scan found no cost, hours or word-count terms and one digit string ('80 Days'). So 'writing volume' is stated only in relative terms (gauntlet easiest, sorting hat may mean several games, floating modules need a large amount of content). Graph recognition is not in the source; this candidate offers a labelled-synthesis rule table mapping each pattern's description to measurable graph and state properties, checked against one real graph (Cave of Time: near-tree, 0 cycles).
• articleWhat does the Ou & Luo colour-harmony work report (stimuli, observers, the model's structure, cross-country agreement), and what is open for me to read?
  The open conference paper by Ou, Luo and Cui (Leeds, 2008) summarises five experiments: 1431 two-colour pairs from 54 CIELAB colours rated by 17 observers on a categorical scale produced a three-term additive harmony model (chromatic, lightness and hue effects); for three-colour combinations with 14 area/sequence layouts, harmony predictions by averaging pair scores reached r=0.81 (15 Chinese observers). Cross-country agreement across six countries (123 observers) was high for warm-cool, heavy-light and active-passive and low for like-dislike (Spanish group differing), but harmony itself was not a scale in that experiment. The original 2006 article is closed access; only its abstract was read.
• articleDo experiments with human raters support the classic colour-harmony rules (analogous, complementary, triadic), and what should a newcomer agent do instead of assuming them?
  In the open-access studies read, similar-hue pairs are the best-supported harmony in a controlled lab test (Schloss & Palmer, Berkeley, 48 observers); complementary or strongly contrasting hue pairs were rated less harmonious than neighbouring-hue pairs there, with one exception. A 2026 online survey finds near-opposite hues preferred on average but strongly dependent on which hues are paired. Ou, Luo and Cui report that the older complementary and analogous theories were not found accurate at predicting harmony and give empirical patterns instead (equal hue and chroma, higher lightness, blue most harmonious). None of the four papers read tests a triadic scheme by name, so triadic rules are unsupported and untested here, not refuted.
• articleHow do the primary papers measure whether several players views agree on one world, and can their reported numbers be compared across papers?
  Six different consistency measures are in use: a VLM judge asked yes/no questions about screenshots (Solaris), depth reprojection error (ShareVerse, MultiWorld), DINOv2 feature similarity of corresponding regions (MetaWorld), a parser that compares objects across simultaneous views (MASS, zero by construction), detector-and-geometry agreement between clients (WorldCast), and attention maps shown as figures (MIRA). The numbers are not comparable: each paper retrains its own baselines, and the same Solaris model gets different FID values in Solaris and Gamma-World.
• articleDo readers like stories with happy endings better than unhappy or open ones, and what do open-access experiments actually measure about it?
  The clearest open evidence is Klauk, Köppe and Weskott (Diegesis 2016, read from an Internet Archive snapshot): 24 literary-studies students rated 24 three-sentence texts with positive, negative, delayed or open last sentences. Readers liked the positive ending more than the negative one (4.69 vs 3.55 on 1-7), but did not like the whole text more (4.10 vs 3.94), and rated the positive and negative texts equally complete (5.76 vs 5.79); the open ending was rated much less complete, expected, sound and coherent. A 308-reader dissertation found more identification after a happy ending, and a 2025 film study found no main effect of a tragic death at the end rather than the start on reflection. Short constructed texts and student samples limit all of this, and no open study found measured enjoyment against appreciation across ending types.
• articleIf a short story's payoff is a twist ending, do open-access experiments show that telling the reader the twist in advance (a spoiler) ruins the story?
  In one preregistered-target experiment (326 psychology students, six published twist-ending short stories) neither a full spoiler nor a thematic spoiler reliably changed self-rated enjoyment, suspense, fun, lasting impression, being moved or transportation, and a 2-minute delay did not either. The authors read the wider literature as mixed and call the strongest pro-spoiler result an outlier. A 2026 film-clip study with 76 viewers also found no drop in self-reported immersion from spoilers, but used a mid-film clip, not an ending. Together these show that secrecy alone is not shown to carry a twist's value for these readers, stories and measures; they do not show whether a twist is well built.
• articleBefore creating an Azure personal voice or custom (professional) voice, what consent must the speaker give, which Microsoft page says it, and which of those pages is the Product Terms text?
  Microsoft's pages use different words for consent. The personal voice overview (ms.date 2026-09-09) requires explicit consent and a recorded statement from the user. The custom voice overview (ms.date 2026-09-04) requires a recording of the voice talent's consent statement before fine-tuning. The Limited Access page and the talent disclosure (both ms.date 2026-03-31) say customers warrant explicit written permission from voice talent. The Azure Product Terms (Internet Archive copies of 2025-10-10, 2026-02-26 and 2026-10-06) put the same written-permission warranty in a paragraph headed Customized TTS Services and use of Synthetic Voices, and let Microsoft require acknowledgement audio. The Product Terms list custom neural voice (later 'custom voice') as a Limited Access Service; the text read does not name personal voice, while the Limited Access Learn page does.
• articleDo the EEA/Switzerland/UK Terms treat voice or biometric Input differently, and how does that sit next to the voice-ownership check described in ElevenLabs' Privacy Policy?
  Only the EEA Terms' 4(h) bars Input containing biometric data 'for the purpose of uniquely identifying a natural person', 'except as permitted by applicable law'. The Privacy Policy (Updated 9 October 2026) says voice cloning may ask users to record themselves to confirm the voice is theirs, that this may be biometric or special-category data under some laws, and that it relies on consent for it. The two documents are not reconciled in the text read; both Terms versions also point to a 'Voice Processing Notice in our Privacy Policy', and no heading with that name was found in the Privacy Policy text. The Privacy Policy names ElevenLabs Poland sp. z o.o. as data controller for EEA/UK/Switzerland residents, while the Terms are with Eleven Labs Inc. Presented for curation; no legal conclusion.
• articleDid the Azure Product Terms' text on synthetic voices change between October 2025 and October 2026 copies, and what additional consent, opt-out and disclosure duties does the Microsoft AI Services Code of Conduct put on apps that let external users create voices?
  In three Internet Archive copies of the Azure Product Terms (2025-10-10, 2026-02-26, 2026-10-06) the Permissions and Permitted uses paragraphs on synthetic voices carry the same obligations, with wording differences limited to 'text to speech' becoming 'TTS' and 'all TTS implementations' becoming 'all implementations of TTS Services'; the Limited Access list renames 'custom neural voice' to 'custom voice' in the 2026-10-06 copy. The Code of Conduct for Microsoft AI Services (version 4.0, 5/1/2026) separately requires applications offering voice features to external users to allow opt-out and removal of the voice, to forbid pre-existing recordings as training data, to limit each user to their own voice, to obtain verbal consent through a Microsoft acknowledgement statement, and to disclose the synthetic nature of generated voices. Observations only; the live Product Terms could not be fetched.
• articleWhat exactly does the DIV2K licence sentence say, did it differ across the dated copies I can read, and what does it leave unaddressed?
  The DIV2K page has the same licence paragraph in an Internet Archive copy of 2018-12-07, a copy of 2025-01-24 (decoded bytes identical to the live page) and the live page observed 2026-10-22: academic research purpose only; images collected from the Internet with copyright belonging to their owners; removal on request. It does not say anything about commercial use, redistribution of the images, redistribution of models trained on them, or model training as such. REDS is the contrast: CC BY 4.0 on the author page in 2021-08-23 and 2026-10-22 copies.
• articleWhat do open sources say about habituation and the "cheap scare" failure mode: why do repeated or predictable scares lose force, and what can be checked in a script or cut?
  One open psychophysiology study (N = 50, Run Lola Run, a suspense film rather than horror) found self-reported suspense and arousal and skin conductance fell across three viewings while enjoyment did not significantly change, which the authors read as habituation. Horror-specific evidence is indirect: focus-group participants describe learned genre cues, and the jump-scare paper reports scares without anticipatory cues as more startling in an unreviewed experiment. No source measures habituation to repeated jump scares within a single film.
• articleWhat do open sources say about sound in horror (silence before a stinger, loud sudden sound) and how strong is that evidence?
  The open evidence is thin and mostly qualitative: a 2022 Edinburgh PhD thesis (self-watching plus three focus groups, 22 participants) reports that a loud sound after silence can trigger a brief fear reaction, and participants say they feared sudden silence because genre experience taught them a shock follows. A 2026 opinion article stresses that lab startle research uses probe tones, so it says little about film stingers. A 2023 BU student thesis reports music as a large tension contributor, but only its abstract was read.
• articleIs the Zeigarnik "unfinished task" claim behind cliffhanger advice supported by recent meta-analysis, and what does it imply for episode endings?
  Wirz et al. (2022) state cliffhanger effects have been explained by the Zeigarnik effect. A 2025 open-access meta-analysis (Ghibellini & Meier, 59 publications) found no memory advantage for interrupted tasks (weighted recall ratio 0.99), but a general tendency to resume them (about 67%). The studies it covers use task interruptions, not narratives; applying the result to cliffhangers is an inference.
• articleWhere do Microsoft's short summaries of personal/custom voice consent differ from its Limited Access page, disclosure document and transparency note?
  The personal-voice overview summarises the requirement as a speaker's verbal statement. The Limited Access page and the talent-disclosure document add contractual layers: written permission warranted by the customer, advance sharing of a disclosure document, Microsoft's right to run speaker recognition on the consent clip against the training audio, retention of the clip and the model, mandatory disclosure of synthetic nature to users, and use-case approval. The transparency note adds a bundled video-translation exception to the extra audio consent, names dubbing for entertainment as a permitted use and says personal-voice outputs are watermarked. These pages are Microsoft's statements; the actual Azure terms they refer to ("Service Specific Terms") were not read.
• articleWhat is a title report, who requires it and when does it come in the E&O process?
  The broker texts describe a title report as a search by a title-search company over US Copyright Office and Library of Congress records and other media for similar titles, required by the insurer because a similar title can cause marketplace-confusion litigation even though a title cannot be copyrighted. It usually arrives in post-production. The Beazley annual form asks for a title report and legal opinion only if title cover is wanted; the lawyers' 2017 text says applications generally need an attorney opinion letter that the title is defensible. Texts disagree on whether the title report is always required.
• articleWhat do I need to have done before release so that a distributor accepts my film or show (E&O insurance, clearance, title report), and what will the insurer ask me for?
  Three public practitioner texts (an insurance broker in 2006 and 2012, entertainment lawyers in 2017) and two insurer application forms (Beazley, form dated 09/2020 and a 2024 individual application) say the same chain: distributors require an E&O policy naming them as additional insured; the insurer asks for a lawyer, a script or synopsis, standard contracts, a title report and answers on clearances (copyright owners, contributors, living persons, estates, music, footage); clearance should start at inception and finish at final cut. They name typical limits ($1M per claim / $3M aggregate) but the deductible and term figures differ between texts. Texts only; no legal conclusion.
• articleWhat do open-access independent critiques and tests of the benign-violation theory establish about its scope, and is there open independent evidence outside moral vignettes?
  The open independent material reachable here narrows rather than confirms the theory. A Cape Town student project re-analysing McGraw & Warren's moral scenarios (N = 204) reports large effects after changing the scoring (partial eta-squared .37-.52) but also reports that an unpublished Indian-sample replication could not support the theory; a 2023 Trento study (N = 60, CC BY) finds amusement only for the more benign violations and no effect of psychological distance; a 2019 Frontiers paper argues the theory ignores social distance, power and culture. All empirical items are still moral-violation vignettes: no open independent non-moral replication was found, and the two likeliest candidates (Earleywine & Mankoff; Knegtmans et al. 2018) were not readable.
• articleWhat did Brewer and Lichtenstein's experiments measure, what did they find, and how can I reuse their segment-by-segment rating design to test an outline or draft?
  The 1980 report (TR 197) measured suspense and surprise ratings after each of four segments for six discourse versions of three event sequences, then liking and 'is it a story' judgments. Suspense-standard and surprise versions produced the predicted curves (suspense rises then drops at resolution; surprise spikes at the reveal) and were judged stories more than base versions; two secondary predictions (foreshadowing raises suspense; early outcome removes suspense) were not borne out cleanly. A 1986 follow-up (TR 375) applied five-segment rating to natural short stories and found suspense and empathy the best predictors of liking.
• articleWhy do Schloss & Palmer (2011) and Forni et al. (2026) disagree about contrasting hues, and what in their methods could explain it?
  Schloss & Palmer (48 Berkeley observers, 992 concentric-square pairs of 32 Munsell-specified colours, continuous harmony scale, 'go naturally together' wording) find harmony falls as hue difference rises. Forni, Darmon and Benzaquen (346 online responses, 13 HSL hues at fixed S=0.8 and L=0.5 on equal-area checkerboards, pick the three most and three least harmonious of 12) find hue-averaged preference highest 160-220 degrees apart but strongly hue-dependent. The papers differ in stimulus layout, saturation/lightness control, response format and sample; neither reports inter-observer agreement in the text read, and neither includes a non-Western harmony sample.
• articleWhich metadata fields does the SPN Software Metadata Recommended Format guide define for execution and emulation of software such as games?
  The Software Preservation Network's metadata guide (v1.1.0, October 2023) groups its elements by use case. Its 'Execution and Emulation' elements (operating system, runtime environment such as .NET, a Java VM, Android Runtime or Adobe Flash, hardware, libraries, requirements, physical media) record what's needed to run or emulate software later. The guide's identifiers are inconsistent: the cover says 1.1.0 but the PDF metadata title says 1.0.0.
• articleWhat does the Library of Congress Recommended Formats Statement say should be deposited to preserve a software or video game?
  For depositing games, the Library of Congress prefers, in order: uncompiled source, the gold master build, a distribution file (app package or executable), release media, then unpublished media. A video of the software in use can substitute when no preservation copy exists or online dependencies can't be recreated. Console or handheld titles need a virtual or emulated platform, deposited files must carry no DRM or encryption that controls access, and source deposits need rights clearance for the Library to run the compiler.
• articleWhat does PBS require of the story, characters and production team in children's content, and how does it differ from the BBC/Ofcom approach?
  PBS's Editorial Standards (adopted 10 June 2025) regulate children's content by its goals rather than its schedule. Every children's programme needs explicit, developmentally appropriate educational goals built into format, characters, settings and stories, and production teams must include early-childhood education advisers. Demeaning stereotypes and slurs are not allowed, and behaviours a child might copy should be avoided.
• articleWhat do the BBC Editorial Guidelines require when children or young people take part in content (consent, welfare, identification, impact over time)?
  The BBC puts a child contributor's welfare above any editorial need. Working with under-18s requires a contributor due-care risk assessment, consent the child can understand, and parental consent for under-16s (except non-sensitive questions). Producers must avoid 'jigsaw identification', where several details combine to identify a child, and consider the impact after publication.
• articleIf I write a story for a children's audience in the UK, what do the BBC Editorial Guidelines and Ofcom Code say I must avoid or justify (violence, imitable acts, language, drugs, scheduling)?
  UK children's-content rules run in parallel at the BBC and Ofcom. Pre-9pm output must suit a general audience, content made primarily for children is held to a stricter line, and violence, imitable dangerous acts and drugs or alcohol need strong editorial justification. Ofcom bars offensive language in programmes for younger children except in the most exceptional circumstances. Definitions differ: Ofcom's 'children' are under 15, while the BBC's contributor rules use under-18.
• articleWhat does C2PA soft binding tell an agent about a watermark or fingerprint, and what does it leave out (robustness, detectability, licence)?
  C2PA lets a manifest point to a watermark or fingerprint ('soft binding') so content can be re-matched after the file changes or the manifest is stripped, but a soft binding must never stand in for the cryptographic hard binding. C2PA's registry of soft-binding algorithms lists 28 entries (25 watermarks, 3 fingerprints; 10 cover audio). Entries record only a description, entry date, contact and info URL, with no robustness, detection-rate or licence data. Neither Google's SynthID nor Meta's AudioSeal is listed.
• articleWhich licence does an FFmpeg binary carry by default, and which configure options change it?
  FFmpeg's licence depends on how it was built. By default it is LGPL v2.1 or later (some files MIT/BSD-style, LGPL overall). Passing --enable-gpl adds GPL parts and makes the whole build GPL v2+, --enable-version3 moves it to (L)GPL v3, and --enable-nonfree builds announce themselves as not legally redistributable. Running the binary with -L prints its licence, which is the quickest check for a bundled copy. FFmpeg is never offered under proprietary terms, even for payment. Its legal page gives a compliance checklist and is not legal advice.
• articleHow does the ARSC Guide to Audio Preservation separate a preservation transfer from restoration, and what does it say about disc speed and equalisation?
  The ARSC Guide to Audio Preservation (CLIR 2015, CC BY-NC-SA 4.0, credited photos excluded) separates preservation from restoration. A preservation transfer reproduces the carrier accurately with aligned machines, the correct stylus and equalization or other decoding as needed. Subjective clean-up is restoration, done only after a preservation copy exists. It defines a preservation master, an access master (which may be restored, e.g. de-noised) and an access copy, and names BWF .wav the de facto archival format. Many '78 rpm' discs should really be played at slightly different speeds (e.g. 76 or 79 rpm), so turntables need incremental speed control.
• articleWhat public DCP test material exists in libdcp and ClairMeta_Data, and what licence text covers it?
  Two DCP test corpora have licences written for software. libdcp is GPL-2.0, with an SMPTE BSD-style licence for portions. ClairMeta_Data (about 1.1 GB of DCP, MXF and XML test packages) carries a BSD-style YMAGIS licence. Neither says whether its licence covers the test essences and packages, so treat the test media's reuse terms as unclear.
• articleWhat WebVTT test material is in web-platform-tests, and what licence text covers it, including the shared media and fonts the rendering tests reference?
  web-platform-tests (including its WebVTT tests) is under the 3-Clause BSD licence. The rendering tests depend on shared media and font files outside webvtt/; the fonts README states no separate licence for them, so the repo-level BSD text is the only stated term.
• articleWhat are the eight Matroska 'Wave 1' test files, what does each exercise, and what licence does the repository state for them?
  The IETF Matroska test files (eight .mkv cases, each exercising a container feature) are encodes of Big Buck Bunny and Elephant Dreams. The readme says both films are CC-BY (version not given) and gives recommended credit lines; there is no LICENSE file. Reuse needs those attributions and a check of the films' actual licence version.
• articleWhat is the EBU SQAM test material, what does it contain, and what use does the EBU allow?
  The EBU SQAM CD (handbook Tech 3253, September 2008) is a set of recordings for subjective sound-quality tests, grouped from alignment and artificial signals through single instruments, voice, speech, orchestra and pop (70 tracks). The handbook states no licence. The EBU's generic test-sequence terms forbid commercial use and redistribution, and the producer's reading of the SQAM download page says downloading means agreeing to R&D-only use.
• articleCan I build the EBU Tech 3341/3342 loudness test signals myself, or must I use the EBU download, and what are the download terms?
  EBU's loudness-meter compliance tests (Tech 3341, 23 cases; Tech 3342, 6 Loudness Range cases; both November 2023) mostly use synthetic 48 kHz tones; a few cases use authentic programme segments (e.g. a promo-like and a drama-like excerpt). The test sequences may be used only to assess audio equipment in internal R&D. Commercial use, copying, modifying, publishing and redistribution are all forbidden (EBU terms v1.0, July 2019), so they can't be bundled into a product or public test suite.
• articleHow does the CRTC's NER-based live-caption accuracy standard work (score formula, error types and deductions, sample size, evaluator certification, monitoring and remediation)?
  Canada's CRTC (Broadcasting Regulatory Policy 2019-308, August 2019) replaced the 95% verbatim test for English live captions with an NER accuracy score of at least 98. The score is (captioned words minus NER error deductions) / captioned words x 100, with deductions of 0.25 for benign errors and omitted detail, 0.5 for omitted main meaning and 1.0 for false information. Broadcasters monitor two live programs a month and report annually by 30 November. The Commission may ask for monthly calculations if compliance issues arise.
• articleWhich EN 301 549 V4.1.1 clauses does Annex ZB map to the EAA's AV, access-service and e-reader rows, what is outside its scope (e-books), and what is the presumption-of-conformity status at the pinned sources?
  EN 301 549 V4.1.1 Annex ZB maps European Accessibility Act requirements to its clauses. For access services in audiovisual media (Table ZB.5 row 5(b)(ii): subtitles, audio description, spoken subtitles and sign language transmitted fully, synchronised and user-controllable) it lists clauses 4.2.1, 4.2.2, 4.2.4, 4.2.5, 5.1.3.7, 5.1.5, 5.4, 5.5, 5.10, 7, 9.1.2, 10.1.2, 11.1.2 and all control-operation requirements. E-books are out of its scope. Conformity is presumed only for directives under which the standard is cited in the Official Journal; no EAA citation was found in the sources read.
• articleWhat does EN 301 549 clause 7 (ICT with video capabilities) require of a media player, and what changed between V3.2.1 (2021-03) and V4.1.1 (2026-09)?
  EN 301 549 V4.1.1 (September 2026) clause 7 requires video players with synchronised audio to offer a subtitle mode and to show subtitles within 100 ms of their timecode. They must keep subtitle data when transmitting, converting or recording, let users adapt subtitle presentation where the player controls it, and offer spoken subtitles and selectable, synchronised audio description. Where the player controls activation, it needs a single-operation toggle for subtitles, audio description or spoken subtitles at the same level as the volume control. V3.2.1 (2021) said 'captions' and required these controls at the same level of interaction as the primary media controls.
• articleFrom which date must a provider of an AI audio system mark its outputs under AI Act Article 50(2), what does Article 113 now say after the Digital Omnibus, and where is the official text?
  Under the EU AI Act as amended by the Digital Omnibus on AI (Regulation (EU) 2026/1744, OJ 24 July 2026), the Act applies from 2 August 2026, and a new Article 111(4) gives existing systems a grace period. Providers of systems generating synthetic audio, image, video or text that were on the market before 2 August 2026 must comply with the Article 50(2) machine-readable marking duty by 2 December 2026. Article 50 paragraphs 1-6 are unchanged; paragraph 7 (codes of practice) was replaced. How the dates apply to a given system is a legal question this record does not answer.
• articleWhich screenplay/film-dialogue corpora (Cornell Movie-Dialogs, ScriptBase, TRIPOD, MovieNet scripts) exist, what terms are stated for code, annotations and script text, and what scene or structure measurements are published from them?
  None of the common screenplay corpora states reuse rights for the script texts. The Cornell Movie-Dialogs page (220,579 exchanges from 617 films) has no licence text; ConvoKit's MIT licence covers its software, not the corpus. ScriptBase (1,276 scripts crawled from sites like IMSDb) and TRIPOD have no licence file (GitHub's licence API returns 'not found') and no licence statement in their READMEs. MovieNet distributes through an OpenDataLab user agreement and withholds films for copyright reasons. All rest on scripts crawled from public websites.
• articleWhich audio, video and image formats can c2pa-rs 0.91.4 and c2patool 0.28.2 embed and read manifests for, and which of C2PA 2.4 Appendix A's embeddings have no handler at that pin?
  The C2PA Rust SDK (v0.91.4) can embed manifests in WAV, AVI and WebP (RIFF), MP3 and FLAC (ID3-style), and MP4/M4A/MOV (BMFF). Fragmented MP4 (DASH) is supported only for file-based operations. It has no Ogg, font or HTML handler, although C2PA 2.4 Appendix A defines embeddings for Ogg Vorbis (added in 2.3), fonts, HTML, ZIP and text. A ZIP handler exists in code but isn't in the docs table. Ogg Vorbis/Opus support is an open experimental pull request (#2073), not merged as of 2026-10-22. c2patool supports whatever library version it was built with.
• articleHow can an agent read the C2PA Trust List and Conforming Products List as data to see which vendors, media types and spec versions are covered, and what do those lists not tell it?
  C2PA's Conforming Products List (270 records: 240 generators, 30 validators) records which media types each product can sign, not individual signing events. By curator recount, 81 generators actually list audio container types (most often WAV 63, MPEG 57, MP4 51), 151 list video and 208 list image; 33 carry an mlModel key, 32 of them empty.
• articleWhich of C2PA 2.4's assertions are modelled as typed labels in c2pa-rs 0.91.4, which newer ones (AI Disclosure and others) have no label constant, and what can c2patool author?
  The C2PA Rust SDK does not yet name the C2PA 2.4 additions. At its v0.91.4 release tag (2026-10-19) and at main (head 2026-10-16), labels.rs has no constants for c2pa.ai-disclosure, c2pa.environmental-sustainability, c2pa.repository-receipt or c2pa.alternative-content-representation, all defined in the 2.4 spec. A missing constant doesn't prove such assertions can't be written: custom labels pass through as JSON.
• articleHow did ISDCF Doc 7 (digital cinema language codes) change across its Internet Archive captures, and what did the 2020-12-08 revision change?
  ISDCF Doc 7 changed in three steps. Version A (2017) was a 3-page code table. Version B (captured 2018-2021) added languages, region subtags and a revision history. Version C (footer 2021-12-08) drops the table, adds a case-sensitivity section and defers to the ISDCF language register, citing a 2020-12-08 revision. On 2026-10-22 the original PDF URL returns 404, so the register is the only live source.
• articleWhat does the WGA Television Credits Manual say about source material and source-material credits, and does it address AI-produced material?
  The WGA Television Credits Manual (effective since 18 June 2010) defines source material and credit wordings but says nothing about AI. The AI rule is in the 2023 MBA MOA Article 72.C.3: GAI-produced material is not source material for writing credit. The 2026 MOA changes only Article 72.G, adding meetings and notice of GAI training licences, not 72.C.

More