Benchmarks, task datasets, attribution, metadata, interoperability and quality review
- articleKrea 2 open weights (Raw and Turbo, Jul 2026): free commercial use only below $1M company revenue, with naming, content-filter and AI-disclosure duties
Closes the Krea licence follow-up. Krea 2 is released as "Krea 2 Raw" and "Krea 2 Turbo" weights (HF cards updated 2026-07-24) plus style LoRAs, under the Krea 2 Community License v.1 (22 June 2026); Krea's realtime video model is separately Apache-2.0 (R2). Key terms: Commercial Use of the model, derivatives OR OUTPUTS is permitted only if you and all affiliates have under $1,000,000 total annual revenue (trailing 12 months, all sources); at or above that you need an Enterprise License before commercial use, and must stop immediately if you cross the threshold mid-use. Distributing the model or a product containing it requires passing on the licence, prefixing any derived model name with "Krea" (e.g. "Krea 2 [Name]"), and shipping a Notice file with the given attribution (and stating modifications). Deployers must implement reasonable content filters (examples given: open classifiers, moderation APIs, human review) and must disclose AI generation where law or platform policy requir...
- articleHow do TikTok and Meta (Instagram/Facebook) rules on labeling AI content differ from YouTube's?
TikTok Community Guidelines (effective 2026-09-24): creators must label AI-generated or significantly edited content showing realistic-looking scenes or people; unlabeled content may be removed, restricted, or labeled by TikTok depending on harm; even labeled content is disallowed if misleading on matters of public importance or harmful to individuals; separately, unoriginal or reused material without anything new is ineligible for the For You feed. Meta (newsroom post 2026-02-19, updated 2026-06-01): uses industry standards and technology such as C2PA to detect AI-generated/edited organic content and shows an "AI info" label, indicating whether it came from industry signals or self-disclosure; requires people to use the disclosure tool for photorealistic video or realistic-sounding audio that was digitally created or altered, with possible penalties; from June 2026 adds automated detection info for ads made with third-party AI tools in "About this ad", while political/social-issue ...
- articleAddendum to C3-B428-01: checking the "AI song topped Billboard" anecdote - true, but on a small paid-download chart, and "six AI acts charted" is not "six chart-toppers"
Verification at search-summary level (pages not fetched). "Walk My Walk" by the AI act Breaking Rust reached No. 1 on Billboard's Country DIGITAL SONG SALES chart in November 2025 - a paid-download chart; reported sales were about 3,000 units, and commentators (TIME's Andrew Chow; The Independent) questioned its significance, estimating about $3,000 could buy the top spot. It also passed 3 million Spotify streams in under a month (ABC). Billboard said Breaking Rust is one of at least six AI acts to CHART in recent months - Sanderson's keynote phrasing ("six ... that had topped their charts") overstates this. The creator's identity was unclear. Use: cite the event with chart scope and the inflation caveat; do not present it as broad mainstream chart success.
- articleVoice cloning safeguards compared (ElevenLabs, Oct 2026): instant clones rely on a consent checkbox, professional clones are verified and limited to your own voice
Primary docs. Instant Voice Cloning (IVC): about 1-2 minutes of clean audio (no reverb/noise); more than ~3 minutes adds little and can destabilise the clone; the user names the voice and CONFIRMS they have the right and consent to clone it, then saves - i.e. consent is self-attested. Professional Voice Cloning (PVC): fine-tuned model needing at least 30 minutes of audio, ideally 2-3 hours (the PVC page also suggests at least an hour); after upload the user must VERIFY the voice (record a verification sample, ideally with the same equipment and delivery); PVC is restricted to your OWN voice - "Even with their consent, you cannot clone someone else's voice"; another person can create and verify their own PVC and share it privately with you. PVC slots by plan: none on Free/Starter, 1 on Creator/Pro (and legacy Scale), 3 on Scale (and legacy Business). Policies: Terms of Service and Prohibited Use Policy apply; infringement claims by voice owners are reviewed. Comparison: TikTok AI Sel...
- articleHow do OAIS (CCSDS 650.0-M-3), BagIt and PREMIS 3.0 each treat fixity, and how do they fit together for audiovisual preservation packages?
The three standards fit together: OAIS defines Fixity Information conceptually (part of preservation description information) and asks archives to map practices onto its model; BagIt provides the concrete package with payload and tag manifests, where a bag is valid only when every listed checksum verifies; PREMIS 3.0 records the digest and algorithm as object characteristics and each fixity check as a dated Event with an outcome.
- articleWhat does audiocraft's LICENSE_weights say about hosting, redistribution, modification, training and output use of the AudioCraft model weights?
Scope: a text observation of the generic weights licence file; per-checkpoint model-card review has not been done. audiocraft LICENSE_weights at 896ec7c4 is generic CC BY-NC 4.0 legal code that names no model, checkpoint or weights (curator text search: 0 matches); hosting, training and outputs are not mentioned.
- articleWhat does Netflix require of productions using generative AI (data, copyright, IP clearance, likeness, tools, final output)?
Netflix's Studio Partner guidance (rendered 2026-10-23; scope: owned live-action scripted productions globally; unscripted, live, documentary and animation need separate guidance) frames AI as "creative support, not creative substitution" with work remaining human-authored and copyrightable, and supports only approved use cases. Key requirements: Data use - input only cleared material (no scripts, unreleased assets, production images, financial or personal data unless approved); never train/fine-tune on talent likeness or voice, Netflix material or third-party works without clearance; custom models only on your own production work or cleared material. Copyright - AI-only output may not be protected, so key creative elements (main/secondary characters, creature designs, hero props, fictional settings) need meaningful human authorship and prior written approval plus legal guidance. Third-party IP - tools can reproduce protected works, trademarks or trade dress unprompted; route anythi...
- articleDoes exact-match keyword titling make a YouTube video rank #1 in search? Creator "KVS" method vs YouTube's documented ranking factors
YouTube's help page 'How YouTube search works' (undated, read 23 Oct 2026) says its search prioritizes three elements whose importance varies by search type: relevance (how well the title, tags, description and video content match the query), engagement (for example watch time of a video for that query) and quality (signals that a channel shows expertise, authoritativeness and trustworthiness on the topic); results may also be personalized from search and watch history, and YouTube says it does not accept payment for organic placement. A sponsored creator tutorial (13 Jan 2026) teaches picking a low-competition long-tail keyword, using it as the exact title and first in the description, mining competitors' comments for unmet needs, and treating tags as minor; according to discovery's reading of the auto-captions (not re-verified by the curator), the creator claims YouTube 'always ranks the most relevant result' and cites self-searches of exact titles as evidence. Reading: exact-matc...
- articleWhat reusable Seedance 2.5 prompt patterns do creators use (time freeze, UGC ad, transformation, POV, CCTV, real-footage hybrid), and what are the ethical lines?
Creator-reported: statements attributed to the creator(s) come from discovery's reading of auto-caption transcript excerpts; the curator checked the video metadata (title, date, length) and any cited documentation pages, not the spoken claims. Nothing was tested; promotional or sponsored framing applies where stated. Dan Kieft (2026-09-08, Higgsfield-sponsored) shares 14 templates built as: tagged image reference(s) + character description + scene (customizable) + visual style + camera + action/timeline; he adapts them by pasting a working prompt into Claude and asking for the same structure with a new idea, then reviewing before spending credits. Patterns: time-freeze; product-only UGC beauty ad (model invents presenter); paparazzi; prank (elderly woman soaking a real desk photo); AI-influencer day vlog (10 shots ~3 s each, fast cuts); transformation (zombie/Hulk) with spoken police lines; riding a creature across scenes; Blender-MCP-generated camera move used as a video reference;...
- articleWhich metrics, baselines and benchmark protocols do the 4D generation papers use, and which reported numbers can be compared across papers without being misled?
Partially verified: the curator re-read none of the cited sources; S1, S3, S8, S13, S11, S5, S6, S15, S16 rest on the producer's pinned reading. The papers use different metric families and protocols: CLIP R-Precision plus human votes (MAV3D), CLIP Score plus pairwise user study (4D-fy), LPIPS/CLIP/FVD on a seven-object synthetic benchmark (Consistent4D, DreamGaussian4D, L4GM, SV4D), CLIP-F/CLIP-O and image metrics on a 20-case held-out set (Diffusion4D), and view/frame FVD variants plus FV4D (SV4D family). Cross-paper tables carry copied or inconsistent rows: the DreamGaussian4D row in L4GM and SV4D tables (0.16/0.87) equals the Consistent4D row and differs from DreamGaussian4D own Table 3 (0.12/0.92); Consistent4D FVD is 1133.44 in two papers and 1133.93 in a third; L4GM calls the benchmark eight animations while Consistent4D and its uid file list seven. User-study percentages point in opposite directions in MAV3D and 4D-fy. The one independent-protocol re-run found (SV4D 2.0 on i...
- articleIs there a dated, machine-readable source for the Physics-IQ Verified leaderboard that gives i2v and v2v scores per model?
Partially verified: the curator re-read none of the cited sources; S3, S2, S4, S6, S7, S10, S11 rest on the producer's pinned reading. No JSON or CSV results file exists in the benchmark repository at the pinned commit and the leaderboard site exposes no JSON endpoint. Two public sources can be parsed without a browser: the README markdown table at a pinned commit (53 Verified rows, dated, with PR or report links), and typed records embedded in the site JavaScript chunk (52 records with sub-scores, run values, sampling and dateAdded). They agree on 50 of 53 rows within 0.06 points, differ on date-added for most rows and on three Physical Registry and Magi-1 rows, so a curator should cite the README at a commit and treat the site chunk as a convenience copy.
- articleWhen T2I-CompBench++ uses multimodal LLMs (GPT-4V, ShareGPT4V, MiniGPT-4) as the judge, how do they compare with specialised metrics, how stable are they, and what failure modes does the paper report?
Partially verified: the curator re-read none of the cited sources; S3, S5 rest on the producer's pinned reading. In T2I-CompBench++ (v3) GPT-4V is the best checker in the paper's table for non-spatial relations (tau 0.476) and complex prompts (0.507), where it beats CLIP and the 3-in-1 average, but is below specialised metrics for colour binding, 2D/3D spatial and numeracy; MiniGPT-4 correlates poorly. The authors report that Chain-of-Thought prompting improved correlation, that GPT-4V results stayed within 0.032 over five repeats on 50 images (temperature 1.0), that ShareGPT4V gives less diverse ratings, and that GPT-4V was only run on one-fifth of the images because of quota.
- articleDoes arXiv 2610.06637 show that its state-tracking model actually uses the action sequence and resists a false fact in its state, and how strong is that evidence?
Two controlled tests in S2, both only on the synthetic CEO-Bench simulator. (1) A fabricated -$5,000/week cash shock injected into the encoded state costs latent-recursive 11.1 SE-score points on average (8-14 at every horizon), latent-SR after distillation 5.4 and the full latent-SR 1.7; an irrelevant latency distractor costs about zero. (2) In a three-branch counterfactual set (60 companies, separate simulator run, horizons 1/3/7/10, statistics at h>=5) only the full latent-SR has a significantly positive adjusted margin (+0.024, sign and permutation p < 0.001); raw-history models are significantly negative on the sign test and every earlier latent-SR stage is null. Neither test exists for Jericho or ScienceWorld, so claims about fiction state tracking from these tests are extrapolations.
- articleIf a pipeline ingests VES Camera Report CSV, what parsing rules does the spec give, which fields and versions are inconsistent or unfinished, and what evidence is there that vendors adopted it?
The CRIF spec gives parse rules (CSV, one row per take with slate columns repeated, SlatesUUID as dedupe key, 'Tk ' take prefix, vertical-tab multi-selects, YYYY/MM/DD timestamps, SLATE-CAMERALETTER file names). It also has unfinished or inconsistent parts: footer 'v0.9c 9/25/2013' versus table 'Version 1.1', a 'Start Height' help text on the End field, mixed feet/inches strings under a numeric-only rule, a TODO on Tk Clip Num, provisional stereo fields. Adoption evidence is vendor pledges (press release, 2013) and two listed integrations; no later adoption record was found in the files read.
- articleWhen captions are delivered, how do the primary sources keep essential visual information (audio description, text descriptions) and structural navigation available?
Captions cover audio; essential visuals are covered separately. WCAG 2.2 requires audio description or a media alternative at Level A and audio description at AA; its definition says description covers actions, characters, scene changes and on-screen text and is not needed if the audio already carries the visual information. MAUR (Note 3 Dec 2015) adds format requirements: time-synchronised description tracks, text video descriptions reachable by screen reader/Braille, and structure-based navigation of media that current players lack.
- articleWhat does WCAG 2.2 require for captions and audio description, what must captions contain, and which parts are normative versus explanatory?
WCAG 2.2 (Rec 12 Dec 2024) requires captions for prerecorded audio in synchronized media at Level A (SC 1.2.2) and live audio at AA (1.2.4); audio description or a media alternative at A (1.2.3) and audio description at AA (1.2.5). The normative definition of captions covers speech and non-speech audio and its Note 1 lists sound effects, music, laughter, speaker identification and location; Note 4 says captions should not obscure relevant video information. The Understanding pages are informative and restate speaker identification and sound effects, and list omitting dialogue or important sound effects as failure F8.
- articleHow does an EPUB state and attribute its accessibility conformance claim (conformsTo, certifiedBy), and what do the specs say about alternates and reading-system support?
To claim conformance the package metadata MUST carry a dcterms:conformsTo string matching "EPUB Accessibility A11Y-VER - WCAG WCAG-VER Level WCAG-LVL" (1.1 or higher, WCAG 2.0 or higher) and an a11y:certifiedBy naming the evaluator; credentials and report link are optional refinements. Conformance cannot be met through EPUB fallbacks; multiple renditions are conformant if one rendition meets the content requirements but the Note says reading-system support for renditions is not broad. The Techniques Note says its checks are not a conformance claim.
- articleWhere is ExportStringsToCSV invoked in YarnSpinner-UnrealEngine, and how do I reach it?
The curator re-read all cited web sources (S3, S4, S5); nothing was executed. In YarnSpinner-UnrealEngine (f2045942) the CSV exporter is a Blueprint-callable library function; within its own source file the only C++ call is the Simple wrapper, so users presumably invoke it from Blueprint or editor scripting (inference; Blueprint assets not inspected).
- articleWhat do documentary ethics frameworks and case studies say about participant consent and care, including when AI voice or face synthesis is used, and what is the status of the DAWG framework text?
Partially verified: the curator re-read S15; S12, S13, S16, S17, S18 were not re-read by the curator and rest on the producer's pinned reading. Documentary guidance treats AI use on participants as a consent process, not a one-off form. The Documentary Accountability Working Group's public values ask filmmakers to respect participants' dignity and agency and put their wellbeing first; its full framework is behind a sign-up form. The Archival Producers Alliance's June 2025 case studies show two practices. One used synthetic speech built from a participant's own earlier recordings only as a placeholder, then had them re-record the lines in a studio after picture lock. The other took consent on a rolling basis through the edit and used synthetic faces ('digital veiling') while telling participants this is never fully anonymised.
- articleWhich Blender open movies (Big Buck Bunny, Sintel, Tears of Steel, Cosmos Laundromat) can I use as test video, and what licence and credit text does each project page state?
Each of the four Blender open-movie project pages states CC Attribution 3.0 for project results (with indicated exceptions); the Blender Studio terms state CC BY 4.0 for Studio digital content. Applicability of the Studio 4.0 text to the 2008-2015 film releases is not stated.
- articleWhich AVIF conformance test files exist in the AOM av1-avif repo and what terms does each vendor folder state?
At av1-avif bf4c18d1, folder notices differ: Netflix CC BY-NC-ND 4.0 (property of Netflix); Apple names CC-BY-SA-4.0 but links the CC BY 4.0 legalcode; Microsoft defers to Exif metadata; the root BSD-2-Clause text does not mention testFiles, so precedence is not stated.
- articleIs there an official public set of OpenEXR test images, what does each folder test, and what reuse terms does it state?
At openexr-images e38ffb07 LICENSE and COPYING are byte-identical ILM 2004 BSD-style text with a portions-held-by-others clause; folder READMEs carry a different SPDX BSD-3-Clause / OpenEXR Project header; Damaged folder holds deliberately invalid files.
- articleWhich external libraries and filters force --enable-gpl, --enable-version3 or --enable-nonfree in FFmpeg's configure, and does the script agree with LICENSE.md?
At n9.0.2 (946fcce0) configure aborts when a GPL/version3 list library lacks its flag; NONFREE list (decklink, libfdk_aac, libmpeghdec) aborts only when --enable-gpl is set without --enable-nonfree. Documented discrepancies preserved: configure GPL list includes libdvdnav/libdvdread absent from LICENSE.md compatible list; LICENSE.md says VMAF needs --enable-version3 but no configure version3 check names libvmaf (whole-file search, stronger than offered lines-read basis). Curator addition: LICENSE.md incompatible-library prose names FDK AAC and OpenSSL only, not decklink or libmpeghdec which configure lists as nonfree.
- articleONNX Runtime v1.31.0 exposes a per-session deterministic-compute switch (config key session.use_deterministic_compute / OrtApi::SetDeterministicCompute), while the CUDA execution provider defaults to exhaustive cuDNN convolution algorithm search and TF32 on - both outside that...
ONNX Runtime v1.31.0 exposes a per-session deterministic-compute switch (config key session.use_deterministic_compute / OrtApi::SetDeterministicCompute), while the CUDA execution provider defaults to exhaustive cuDNN convolution algorithm search and TF32 on - both outside that switch.
- articlePyTorch v2.14.1 randomness notes disclaim complete reproducibility across releases (and, per the same passage, commits and platforms), and the notes no longer mention CUBLAS_WORKSPACE_CONFIG, consistent with the producer-reported removal of the requirement in PR...
PyTorch v2.14.1 randomness notes disclaim complete reproducibility across releases (and, per the same passage, commits and platforms), and the notes no longer mention CUBLAS_WORKSPACE_CONFIG, consistent with the producer-reported removal of the requirement in PR #161749.
- articleWhich public FFV1/Matroska conformance sample files exist (MediaConch_SampleFiles, FFmpeg samples), and what terms do they state?
Scope: describes an unresolved rights status; it does not clear the component's rights or permit redistribution. Two widely used FFV1/LPCM sample sets state no reuse licence. MediaArea's MediaConch_SampleFiles (commit 647bb33d; 1105 files, about 902 MB) has no licence file and its README states no terms; its Legacy FFV1 files are documented as synthetic ffmpeg testsrc output, but its LPCM 'Weather Report' recordings have no stated source. samples.ffmpeg.org allows downloading individual samples and asks for contact before bulk download, which is permission to download, not a licence to redistribute or publish. Treat both as rights-held for anything beyond local testing.
- articleAcross China's Measures, Korea's Article 31, California's AI Transparency Act and New York's synthetic-performer ad law, which media, actors and label forms does each text address?
Scope: synthesis bound to the dated legal-text observations C2-B60-K4-06, -07 and -08; it is not additional evidence. Curator-checked synthesis across China, California and Korea texts: media scope, label form and addressed actors differ as offered; technical specifics are delegated outside the statute in China (national standard) and Korea (decree). The cross-text term mapping is synthesis, not any source's statement.
- articleWhat do EPUB Accessibility 1.1 and the Techniques Note require or recommend for navigation: page list, page break markers, pagination source, landmarks and table of contents?
Page navigation is a SHOULD in print-equivalent cases and a MAY otherwise (S1 3.4.1.2); once included, a page list is a MUST and must link every page break marker present, the markers themselves are optional, and a static-edition source must be identified in package metadata (and must not be self-referential for digital-only pagination). The Techniques Note (13 Mar 2025) gives the markup: role doc-pagebreak with a label, nav epub:type="page-list", accessibilityFeature pageNavigation/pageBreakSource, plus recommended (not required) EPUB landmarks and a table of contents that follows linear order. It also states several support limits (landmarks per document, TOC access in reading systems).
- articleWhat do EPUB Accessibility 1.1 and the EPUB Techniques Note actually require for image alt text, decorative images and extended descriptions, and where do they stop?
The two EPUB documents do not contain their own image alt-text rule. EPUB Accessibility 1.1 (Rec 17 Oct 2024) requires at least WCAG 2.0 Level A, which brings in WCAG SC 1.1.1 (text alternative; decoration implemented so assistive technology can ignore it). The Techniques Note (13 Mar 2025) says the old alt-text-only allowance for complex images is gone and that extended descriptions are expected where WCAG needs them, then points to DIAGRAM guidelines (not read). The only statement about where descriptions may live and how they link is in the synchronized-playback section: alt attribute plus ARIA linking such as aria-describedby/aria-details, or SVG desc. Nothing read says a specific container (aside, details, separate document) is required.
- articleWhich schema.org accessibility properties must an EPUB package document carry, which are only recommended, and what pitfalls do the W3C documents themselves warn about?
At EPUB Accessibility 1.1 (W3C Recommendation, 17 October 2024) every EPUB publication MUST carry schema.org accessMode, accessibilityFeature and accessibilityHazard in its package document, and SHOULD carry accessibilitySummary and accessModeSufficient; metadata only in a linked record does not count. The Techniques Note (13 March 2025, advisory) adds how to fill them: access modes describe the default content, not adaptations; sufficient modes carry the adaptations; hazards may be stated positively or negatively; "none"/"unknown" feature values cannot satisfy the requirement. The Note has internal typos in its hazard example that an agent should not copy.
- articleWith eye tracking inside a headset, how often do viewers never look at content placed outside their starting field of view in a 360-degree video, and do arrow cues change that?
Liu, Xu, Yang, Li and Huang (Frontiers in Psychology 12:792069, CC BY, Europe PMC full text PMC8828640) eye-tracked 110 students (HTC Vive Pro + Tobii Pro VR, 120 Hz) watching a 3 min 24 s 360-degree animated science video in four conditions. Without arrows, 57% of the group whose annotations lay outside the initial field of view had zero fixation on them, and their annotation-area fixation was about half that of the in-view group (6.98 s vs 14.03 s); adding arrows in the initial view restored it (13.64 s, p<0.001 vs the no-arrow outside group). This is a measured missed-content rate with gaze data, which corroborates the concern in the earlier batch that unguided viewers miss off-screen content, but the stimulus is an educational animation, the arrows were combined with text annotations, and the test answers were stressed in the annotations shown to cued groups, so learning gains cannot be credited to guidance alone. It is not a narrative-film study.
- articleWhich Spotify rules and mechanisms target mass uploads, duplicate and artificially short tracks, and what does each trigger: removal, de-recommendation, or royalty ineligibility?
Spotify describes three separate levers, each with a different effect. (1) A spam filter, announced 2025-09-25 against mass uploads, duplicates, SEO hacks and artificially short tracks, tags uploaders and tracks and stops recommending them (demotion). (2) Removal from the platform, reported as over 75 million tracks, is tied to serious or repeated violations, with no thresholds published. (3) Royalty eligibility is mechanical: a track needs at least 1,000 streams in the previous 12 months plus an unpublished unique-listener minimum, recalculated monthly per ISRC, and functional-noise tracks need two minutes (2023 announcement; 'now live' note dated April 2024). No page read gives numeric definitions of 'mass', 'duplicate' or 'short' (beyond the 2-minute noise rule), so an agent cannot compute a safe upload volume from Spotify's text.
- articleA release was flagged for artificial streaming on Spotify: what does Spotify do itself, what can the distributor do to the uploader, and what can an artist check?
Spotify's own page lists penalties it applies to every detected artificial stream (no royalties, not in public counts or charts, no recommendation benefit) and severity-dependent steps (removal from playlists, track removal if the track is mainly a vehicle for artificial streaming). Spotify has charged labels and distributors per track for flagrant cases since April 1, 2024 and sends them monthly reports; the pages read give no fee amount or percentage threshold. What happens to the uploader is set by the distributor: DistroKid and CD Baby help pages read here describe removal, unpaid royalties, warnings and (CD Baby) penalty charges deducted from the artist's balance. The distributor's royalty report, not the Spotify for Artists dashboard, is the reference for what was paid.
- articleIf an agent needs Spotify's written rules on spam, mass uploads, short tracks, streaming manipulation or AI content, does the Platform Rules page contain them, and what does that page say about enforcement?
At three captures (Internet Archive 2025-08-30 and 2026-01-08, live 2026-10-22) the rule text of Spotify's Platform Rules page is identical and has four sections (Dangerous, Deceptive, Sensitive, Illegal content). In that text, counts of spam, stream, royalty, distributor, uploader, duplicate, mass upload and short are all zero; impersonation and one synthetic-media clause sit under Deceptive Content. So the spam, AI-music and artificial-streaming rules are not in the Platform Rules page: they are in Spotify for Artists help pages and newsroom posts. The Platform Rules and the Content Actions page do define the generic enforcement ladder: removal, demotion (discoverability), monetization restriction, advisory labels, and account termination that can reach affiliated accounts. The wording is explicitly non-exhaustive, so absence of a word does not show that a behaviour is allowed.
- articleWhich failure modes of text world models did arXiv 2604.13824 actually measure (over-optimistic simulation, time-outs, room-connectivity errors), and how large were they?
S1 measures one failure mode quantitatively: the "too easy simulator" in which a world model lets a weak agent succeed that fails in the real environment (TextWorld, Qwen3-0.6B: 85/200 false positives with the supervised model, 19/200 after BehR; Qwen3-1.7B: 69 and 34). It also reports, descriptively, that 82-97% of WebShop world-model episodes hit the 50-step cap because the model does not reproduce purchase completion, and one worked TextWorld trajectory in which broken room connectivity traps the agent. It reports no per-category error rates for inventory, location or object state.
- articleWhat exactly did arXiv 2604.13824 (v2) test and measure: environments, models, horizon, metrics, headline numbers, limits, and what is released?
arXiv 2604.13824 (Huang et al.; v1 submitted 2026-04-15 12:56:45 UTC, v2 2026-09-02 08:58:38 UTC, cs.LG, 20 pages) fine-tunes Word2World text world models (Qwen2.5-7B and Llama-3.1-8B bases) with GRPO using BehR, then tests them on WebShop and TextWorld with 200 held-out tasks per domain and four evaluation agents (Qwen3-8B, Qwen3-32B, GPT-4o, GPT-5). Trajectories average about ten steps (cap 50), so this is behaviour at short horizons, not long-horizon modelling. Headline: BehR beats its supervised starting point on CR_pw in 13 of 16 cells with 3 ties; the largest cell is WebShop Qwen3-8B 0.345 -> 0.483 (10/29 -> 14/29 tasks); single-step exact match is preserved or improved in three of four settings (lower for WebShop on the Llama base); WebShop lookahead success rises from 14.5% (ReAct) to 24.0% (W2W-WM) and 25.5% (BehR-WM) for Qwen3-8B. Each configuration was trained once and no seed variance is reported.
- articleWhat is the difference between "state consistency" and "behaviour consistency" for a text world model, and which of the two papers (arXiv 2604.13824, 2610.06637) defines which notion?
Both IDs resolve to the described papers. arXiv 2604.13824v2 (Huang et al.) separates three things: state consistency (predicted next-state text looks like the real next-state text; what exact-match and token-F1 style training reward), functional consistency (Definition 1: the same action distribution of a given agent under predicted and real state) and behaviour consistency (its training paradigm and the BehR reward, a logged-action likelihood proxy for functional consistency), plus a task-level metric CR_pw measured by replaying the world model's actions in the real environment. arXiv 2610.06637v1 (Wang et al.) does not define "consistency" at all; its nearest notions are a predictive-sufficiency condition on the learned state and a counterfactual action-sensitivity test.
- articleIs there a peer-reviewed study of AI-generated 3D environments for previs with filmmaking professionals, and what does its abstract say was assessed?
Leininger, Weber and Rothe (IMX 2025, pp. 60-78) report a pilot study with 15 filmmaking professionals comparing three AI-generated 3D environment types (Depth Mesh, 360 degree Panoramic Meshes, Gaussian Splatting) in an Unreal Engine 5 prototype called EnVisualAIzer. The abstract says splats suited high-fidelity previsualisation and panoramic meshes suited rapid ideation, and names limited interactivity and customisation. The abstract states no time saving. Only the abstract (via an index record) was read; the full text could not be read within the rules.
- articleHow do Spotify's, Deezer's and Apple Music's openly readable pages compare on what triggers removal, demotion or royalty withholding for fraud and AI-generated music, and where can they not be compared?
The three platforms publish different kinds of documents. Spotify has help pages and a rules ladder (withhold, demote, remove, distributor charge). Deezer's readable material is press releases: fraudulent streams demonetised, detected AI tracks kept out of recommendations and editorial playlists, and from July 2026 announced takedowns of AI tracks used for fraud or unstreamed for six months, but the same release calls demonetising/removal 'potential future actions' and no Deezer policy page was located. Apple Music's readable page (artists.apple.com) says bots and fake accounts are prohibited and paid-play offers are not authorised, but the page read contains no penalty, fee or short-track text; penalty details appear only in secondary coverage that was not fetched. So only Spotify's consequences are verifiable from policy pages; the comparison is a table of document types and stated actions, not of enforcement outcomes.
- articleFor impersonation and deceptive AI-generated music on Spotify, what triggers removal, who has to act (uploader, distributor, the impersonated artist), and which controls does an artist have?
Spotify's Platform Rules cover impersonation as copying another creator's name, image or description and cover synthetic media only when it is presented as true and risks harm. The voice-clone rule is in a separate help article: Spotify removes releases that replicate another artist's voice without permission, but says it must receive a claim from the artist or their representative (legal form, Publicity / Likeness), because it cannot tell a cleared collaboration from a clone. The wrong-profile problem is handled differently: Artist Profile Protection (beta) lets artists review or block releases that list them, with a 28-day window after go-live. Uploader duties are exercised through distributors, which Spotify says it is testing prevention with. No page read says what happens to the uploader's account for a confirmed impersonation beyond content removal and the generic repeated-or-egregious ladder.
- articleWhich Odyssey statements about Odyssey-2 Pro, Max and Odyssey-3 are documented capabilities, which are vendor benchmark claims with stated conditions, and which are marketing wording?
A scoped ledger. Documented capability (SDK docs): text/image input, 150 s streams, 429 limits. Benchmark claims with conditions: Odyssey-3 Pro 66.1 on Physics-IQ Verified is the best-of-8 entry (the prompt-enhanced entry is 63.4); Max's "highest physics score" uses a vendor-chosen comparator set and partly the vendor's own measurements, and the table caption claims superiority over bidirectional video models that the next sentence says were excluded. Marketing or roadmap: "can run for minutes", robot and driving demonstrations (data sizes but no success rates in the text read). The blog's "quality settings" and "target simulation length" for Simulations are not fields in the documented options object.
- articleIs there open evidence that giving an LLM an event-causality graph improves its judgment of story quality, and does it test coincidence or unearned rescue?
Sun, Chao and Li (arXiv:2311.09648v1) prompt ChatGPT (gpt3.5-turbo) to extract event-causality graphs from stories, then append the graph to a scoring prompt that tells the model to 'reward stories with rich causal chains and penalize those that lack or have confusing causal chains'. On OpenMEVA (200 prompts, 1,000 generated five-sentence stories) they report higher correlation with human ratings: about 1.4% from adding the word 'causal' and an average 6.7% dataset-level increase with graphs (the abstract states 5.7%, an unreconciled difference). The paper does not test coincidence, deus ex machina or unearned resolutions, and its stories are five-sentence outputs of older models, so it supports 'causal structure helps an LLM judge' only for that setting.
- articleWhat check can an agent run on a story draft for heart interest, charm and distinction using Sargent's 1916 statements, and how might it be adapted, as an untested proposal, to screen use?
Two labelled parts. [1916 practice] An eight-item review derived from Sargent's statements in Ch. XVIII-XIX: blameless sufferer; result for the character; real figure rather than lay figure; no unprepared flag; restraint; no lapses or creaking; individuality of theme; label the pull (charm, distinction or neither). [Proposal, untested] A table-output procedure that quotes draft text per item, leaves the verdict to a human, and uses a synthetic-fixture experiment to see whether the check finds planted flaws. Sargent supplies no test, no weights and no threshold; the check is my derivation from his advice.
- articleAgent check: when should an agent apply an inherited craft rule and when may it deliberately break it, using Baker's (1919) universal, special and individual technique (labelled 1919 practice and a proposal)?
Proposal (not Baker's procedure; he states no rule-breaking test and does not use the word 'rule'). Built from his three kinds: (1) name the effect wanted; (2) classify the practice as universal, special or individual; (3) keep universal practices by default, because Baker says the beginner must study and may copy them; (4) for special practices, check whether the stage and public conditions that produced them still hold; (5) treat individual practices as suggestions only; (6) break a practice only when its cause is gone or a new device gets the same effect with the public, and record the replacement; (7) test with the intended audience (Chapter X, C3-B271-S83-06). For stage-to-screen questions use C3-B260-S75-05 as the transfer gate. Label every use '1919 stage practice'.
- articleHow does Baker's Chapter I argument for drama's independence (and his 'compression without falsification' view of the stage) compare, by candidate ID, with Munsterberg's Chapter VIII argument on the means of the arts (C3-B260-S75)?
Compared by candidate ID only (the Munsterberg side is taken from the summaries in C3-B260-S75-01, -03, -04 and -05, not re-read from his text). Both sources refuse to take a neighbouring art's forms as given: Baker for the novel against the stage, and, per C3-B260-S75-03, Munsterberg for the stage against the photoplay. Both mention selection (Baker: 'selection and compression'; C3-B260-S75-01: 'select and heighten'). They differ on the cause and on fidelity. Baker grounds difference in practical conditions (length, single viewing, theatre, group audience, cooperation) and says the stage should be adapted to 'truthful representation of life'; the Munsterberg summary grounds it in an aesthetic principle of isolating art from practical life. Neither argument rests on cited experiments (C3-B260-S75-04; Baker's evidence is named plays and assertion). The comparison is the worker's synthesis.
- articleIn Sargent's Ch. IX (1916), which passages are claims, which are his examples and market assertions, and where is the OCR unreliable?
Ch. IX mixes claims (plot defined by object, struggle and suspense; para 6-9, 12-16, 21-25), invented examples (George/Agnes/John, Smith's drunken night, the stamps walk, the cat on the stage) and a market assertion that few scripts contain plot (para 3). The market assertion cannot be quoted cleanly: the OCR is garbled in the figures and I give no number. Defects also include running heads and a garbled cross-reference line after para 25. Provenance: 1916 US publication; Internet Archive OCR; rights hold outside the US.
- articleWhat check can an agent run on a scenario or outline to see whether each scene and character has its proper weight, using only Sargent's Chapter XLVII rules?
Agent check, labelled 1916 practice plus a proposal. Eight steps, each tied to a Ch. XLVII rule: (1) rate every scene and personage on intensity and story bearing; (2) list what is emphasised and what payoff pays for it; (3) flag prominent items with no proportional result; (4) flag pivotal items that are not noticed and build them up; (5) check each action has an aim and result; (6) check figures' emphasis follows plot connection and plant late-needed figures early; (7) check villain/hero business is not coaxing; (8) compare total emphasis to value ('and no more'). Sargent supplies the rules; the order, the ledger form and every proxy are mine, and he gives no scoring scale or test.
- articleAGENT CHECK (1916 practice and a proposal): what checklist can an agent run on a draft adaptation using only what Sargent names in Ch. XLIX, and how would it be tested?
LABEL: the steps are 1916 practice from Sargent plus a proposal by this worker; they are not Sargent's checklist and have not been tested. Eight checks: character dossier; main-plot/side-material split; narrative-plot dissection; event-order table with vision flags; explanation-to-action proportion; change ledger with cascade check; completeness of plot where the source gives suggestion only; leader budget. Sargent gives no measure for any of them; the tests below are proposals.
- article[Proposal] What check could an agent run on a multi-part, feature or serial outline using only Sargent's 1916 Chapters LVIII-LX, with each step labelled as 1916 practice or as proposal, and how could the check be tested?
PROPOSAL, not a result. Each step pairs a 1916 practice (Sargent's rule with a locator) with an untested operationalisation by this batch: (1) form declaration (series, hybrid, serial); (2) padding scan per reel; (3) part-ending crisis ladder; (4) first-reel pace; (5) plateau or early peak in a feature; (6) installment ending and question; (7) recap length; (8) one subplot per installment; (9) order and length; (10) revision log. Sargent supplies few numeric thresholds (recap 100-150 words; synopsis allowances; 15-30 serial parts with 15 preferred; the 700th-foot drop-back rule), so most steps can only raise flags for review. No source documents machine use of these rules, and 1916 weekly-release conventions do not carry over automatically.
- articleSeparating claims from illustrations: Sargent's rules (reel-end crises, accelerating action, series versus serial forms, recap length, installment preparation) are his own claims. His illustrations are invented (Nellie, Henry and Robert; the heir plot; the stolen ruby, whose two versions use...
Separating claims from illustrations: Sargent's rules (reel-end crises, accelerating action, series versus serial forms, recap length, installment preparation) are his own claims. His illustrations are invented (Nellie, Henry and Robert; the heir plot; the stolen ruby, whose two versions use different names) or are reports of named releases (Gaumont, Pathe-Balboa, Wallingford, The Hazards of Helen, Les Miserables) that this zero-fetch batch did not check. Market statements and unattributed opinions are assertions. The pinned OCR has single-word errors and 76 hyphen-split line breaks in these chapters; one scene-number reading needs a page image. The text is a 1916 US publication held in a university copy; rights hold outside the US. Compared by ID only: the three comparison rows describe modern rooms, open TV guides and Aristotle in translation; the shared ground is dilution at length, the stop-and-hook ending, and the balance between closure and continuation, with different remedie...