Evaluation and media provenance

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

• articleAcross IP-Adapter, FaceID, InstantID, PuLID and PhotoMaker, where is the licence line between code and weights, and what blocks commercial use?
  Problem. A pipeline that bakes an identity tool into a commercial workflow can inherit non-commercial terms from a dependency. Conditions. Any use of the five tools in client or product work. Method. Check three layers separately: (1) tool code licence; (2) tool checkpoint terms; (3) face-analysis or base-model dependency. From the files fetched: IP-Adapter repo code Apache-2.0; IP-Adapter-FaceID checkpoints: research-only because InsightFace models are non-commercial; InstantID code Apache-2.0 but insightface models non-commercial and InstantID checkpoints research-only; PuLID code Apache-2.0 with FLUX use bound to the FLUX.1-dev licence and face-encoder terms not stated in the files read; PhotoMaker Apache-2.0 with listed third-party exceptions; InsightFace code MIT but pretrained models and data non-commercial, with a 2025-11-24 licensing route for buffalo_l and inswapper. Decision rule (inference): for commercial output, treat any pipeline that downloads InsightFace antelopev2/b...
• articleFor FLUX, how do I set PuLID's 'timestep to start inserting ID' and true-vs-fake CFG to trade identity fidelity against editability?
  Problem. Identity-conditioning methods can overwrite style, lighting and composition. PuLID aims to keep the original model's output as unchanged as possible apart from the face. Conditions. FLUX.1-dev with PuLID-FLUX v0.9.1 (better ID fidelity than v0.9.0, about 5 points on the authors' similarity metric); runs on a 16 GB GPU with fp8 per the README (needs PyTorch >= 2.4.1 for flux-dev-fp8). Method. (1) 'timestep to start inserting ID': 0 inserts from the first step; earlier insertion gives higher ID fidelity but lower editability; later insertion lowers fidelity, raises editability and disturbs the base model less. Authors suggest 4 for realistic images and 0-1 for stylised images; lower it if similarity is too low. (2) CFG: FLUX.1-dev is guidance-distilled, so the usual guidance_scale (e.g. 4) is 'fake CFG'. Authors recommend fake CFG for photoreal scenes (better aesthetics and facial naturalness) and switching to true CFG if ID fidelity is unsatisfactory, or in stylised scenes w...
• articleHow do I tune InstantID's two weights, what does the keypoint ControlNet lock, and what does its licence permit?
  Problem. Storyboard or character stills need one face identity from a single photo while text still controls scene and style. Conditions. SDXL; InstantID checkpoints (ControlNetModel IdentityNet + ip-adapter.bin); insightface antelopev2 for the face embedding. Method. (1) Extract the largest face with FaceAnalysis(name='antelopev2', det_size=(640,640)); take face_info['embedding'] and draw_kps(face_image, face_info['kps']). (2) Call the pipeline with image_embeds=face_emb, image=face_kps, controlnet_conditioning_scale=0.8, ip_adapter_scale=0.8 (README example). (3) Tuning from the README: raise both scales for more similarity; if over-saturated, lower ip_adapter_scale first, then controlnet_conditioning_scale; lower ip_adapter_scale for stronger text control. (4) With LCM-LoRA use about 10 steps and guidance_scale in [0,1]. Failure modes. Only the largest face is used (no multi-person); inference (labelled): because IdentityNet is conditioned on five facial keypoints from the refere...
• articleWhat do the IP-Adapter-FaceID variants (FaceID, Plus, PlusV2, Portrait) condition on, which parameters tune them, and can I use them commercially?
  Problem. Keeping one actor's face across generated stills needs an identity signal stronger than a CLIP style/content embedding. Conditions. SD 1.5 or SDXL generation where you have one or a few clean face photos. Method. (1) FaceID uses a face-recognition embedding extracted with insightface (FaceAnalysis name='buffalo_l', det_size=(640,640), faces[0].normed_embedding) instead of a CLIP image embedding, plus a LoRA to improve ID consistency. (2) FaceID-Plus adds a CLIP image embedding for face structure; PlusV2 makes that structure weight controllable: in the model-card code pass shortcut=True and s_scale (example 1.0) to generate(). (3) FaceID-Portrait takes multiple face images (default n_cond=5, num_tokens=16), needs no LoRA and no ControlNet. (4) Typical call parameters shown: num_inference_steps=30, width=512, height=768, seed=2023 with DDIMScheduler and a Realistic Vision base. Failure modes. The card itself says generalisation is limited by training data, base model and face...
• articleWhat does openly readable empirical work on suspense say about time limits (deadlines, countdowns), and what experiment would settle it?
  Negative result with scope: in the open-access Lehne and Koelsch (2015) model paper, the whole text contains no occurrence of deadline, time limit, time pressure, countdown, ticking or clock; the authors state there is a scarcity of empirical research on deferring the resolution, that the model lacks a temporal component, and they list open questions such as an optimal temporal distance to the outcome. The nearest open evidence is a repository abstract (Balint, Kuijpers, Doicaru 2017) reporting that non-diegetic delay lowers felt suspense while diegetic delay does not. No experiment that manipulates a stated deadline was found in this batch's searches, so claims that a ticking clock raises suspense rest on craft statements (see -01, -02), not on measured effects.
• articleWhen captions, subtitles and audio description are added to an IMF or DCP, which profile or format carries each, what do the open validators check, and what is lost when the same text is moved between formats or into an archive container?
  Timed text crosses a profile boundary at each handoff: IMSC 1.3 is a text-only TTML profile; the Photon IMF validator requires UTF-8 timed text and accepts PNG image and OpenType font resources, checking encoding and resource types rather than rendering; DCP subtitles follow ST 428-7; and platform profiles can be narrower (Netflix accepted IMSC 1.1 only for Japanese at observation). libdcp's verifier contains no hearing-impaired or audio-description checks. Correction: ST 428-7:2014 is publicly readable from pub.smpte.org (read in CA-B249); only ST 2067-2 remains unread here.
• articleWhat does Ofcom Rule 2.2 require of a UK-licensed broadcaster, how does Ofcom's guidance limit it, and what has Ofcom decided where footage or a portrayal was not what the programme presented it as?
  Partially verified: the curator re-read S1, S3, S4; S2, S5 were not re-read by the curator and rest on the producer's pinned reading. Ofcom Rule 2.2 forbids factual programmes from materially misleading the audience; the guidance aims it at misleading that risks harm or offence, not inaccuracy as such. In ITV's Exposure (2012), video-game footage presented as genuine IRA film and a mislabelled riot clip breached the rule despite limited harm and an apology, with Ofcom criticising weak authentication. Reconstructions are named in privacy Rule 8.19 (and discussed in fairness findings).
• articleWhich fields in the FrameTimelineEvent trace data does Perfetto require and read, so I can tell whether a captured trace is usable?
  Perfetto's trace processor parses expected/actual display-frame and surface-frame start events and frame-end events; it drops display-frame events missing cookie, token or pid and surface-frame events (both expected and actual) missing cookie, token, display_frame_token or pid, counting each as frame_timeline_event_parser_errors.
• articleWhat do I put in a Perfetto trace config to capture Android FrameTimeline jank data?
  To capture FrameTimeline jank data, a Perfetto config needs the data source android.surfaceflinger.frametimeline (no further fields shown) plus at least one buffer, on Android 12 or later.
• articleFor speech and video datasets, how often does the dataset licence disagree with the terms of the underlying sources, and where can I see the per-dataset records?
  Corrected 2026-10-23 - see C3-B643-05; the original statements below are kept for history and are superseded where the correction says so. The Data Provenance Initiative's 2025 audit (arXiv 2412.17847v2) reports, for 3916 datasets, that 25%, 33% and 32% of text, speech and video datasets carry non-commercial licences but that most of their source content carries non-commercial restrictions through source terms (author-reported); its released AMI record lists the dataset licence as CC BY 4.0 while the GitHub licence inferred from a converter repository (AMICorpusXML) is MIT, and Papers with Code reports Unspecified.
• articleDo over 80 percent of speech content carry source restrictions in the Data Provenance Initiative multimodal audit? Corrected: the body reports 78 percent for speech (99.8 text, 99 video); the 'over 80% in each modality' wording comes from the introduction.
  Problem. The earlier finding repeats that over 80 percent of content in each modality has restrictions in its sources. Conditions. arXiv 2412.17847v2 (19 Feb 2025). Method. The introduction says over 80 percent of content from each modality carries undocumented restrictions, but the results text gives non-commercial source restrictions on 99.8 percent of text quantities, 78 percent of speech and 99 percent of video. The 25, 33 and 32 percent licence figures, the 3916 dataset count and the 79, 55 and 65 percent main-text inconsistency figures were re-checked and match. Failure modes. Quoting 'over 80% in each modality' for speech overstates the paper's own body figure by about two points; cite the per-modality numbers. Limitations. The two statements may use slightly different denominators; the paper does not reconcile them. Version and date. v2 fetched 2026-10-23.
• articleWhat licence does OpenToonz v1.8.0 carry in its LICENSE.txt? It contains the BSD 3-Clause (modified BSD) text, applying to files outside the thirdparty and mypaint-brush directories; it is not only a notes file.
  Problem. C2-B14-D2-12 says the OpenToonz LICENSE.txt at v1.8.0 contains only notes about third-party directories and that the full BSD wording was not seen. Conditions. Licence inspection from the raw file. Method. Read the whole raw LICENSE.txt: after the third-party notes and DWANGO copyright it includes a section headed BSD 3-Clause 'New' or 'Revised' License with the full conditions and disclaimer; the README says files outside thirdparty and mypaint brushes are under the Modified BSD License. Failure modes. Treating OpenToonz as having no stated licence text would wrongly block use. Limitations. Third-party directories carry their own licences; this finding is not legal advice, and the licence covers the software, not user-created animation. Version/date. Tag v1.8.0; fetched 2026-10-23.
• articleWhat legal effect do the Commission's Article 50 guidelines claim, and what does the signatory table show for audio-capable providers?
  Partially verified: the curator re-read S8; S10, S9 were not re-read by the curator and rest on the producer's pinned reading. The Commission's Article 50 guidelines (2026-07-20) are non-binding and treat an adequate code of practice as a predictable way to demonstrate compliance and a possible fine-mitigating factor, not a substitute for the Act.
• articleWhich AI-generated or AI-edited images from Microsoft Designer (and Canva) does the vendor say carry Content Credentials or an AI label?
  Partially verified: the curator re-read S11, S12, S14; S13 was not re-read by the curator and rests on the producer's pinned reading. Microsoft states C2PA content credentials for Designer in three overlapping wordings (editing features; AI-generated images; images edited with generative AI) without a field list, so coverage of, for example, text-only edits is unclear; Azure OpenAI applies credentials automatically to DALL·E and GPT-image-1 outputs and documents the manifest fields (producer reading for the field list).
• articleFor EU audiovisual accessibility, which text governs the broadcaster's captions/AD, which governs the player or platform, and where does a technical standard like EN 301 549 fit?
  Partially verified: the curator re-read S1; S2, S3, S4, S7 were not re-read by the curator and rest on the producer's pinned reading. Two instruments split the work: the AVMSD (2010/13/EU as amended 2018) targets media service providers' own services via Member States, while the EAA targets services and terminals that give access to them (including EPGs).
• articleWhat do EAA Annex I requirements say concretely for AV access services, EPGs, AV consumer terminals and e-books, and what do the non-binding Annex II examples add?
  Partially verified: the curator re-read S1; S2 was not re-read by the curator and rests on the producer's pinned reading. Annex I requires services providing access to AV media to transmit access services (SDH subtitles, audio description, spoken subtitles, sign language) fully, with adequate quality, synchronised and under user control, and to make EPGs accessible; it sets no numeric quotas or quality metrics.
• articleWhat do SAG-AFTRA's contract bulletins (2024) and the 2026 TV/Theatrical MOA say about digital replicas of voice and likeness: definitions, notice, consent, pay and strikes?
  Partially verified: the curator re-read S12; S10, S8 were not re-read by the curator and rest on the producer's pinned reading. Under the 2026 SAG-AFTRA MOA (1 July 2026 to 30 June 2030), creating an employment-based digital replica needs at least 48 hours' notice and consent, scans need an articulable business reason, and post-production alterations for noise reduction, timing, pitch or tone and clarity (among others) do not need replica consent. This matters for dialogue cleanup on SAG-AFTRA-covered pictures.
• articleFor a writer on a WGA-covered project, what do the WGA's own public texts say about AI-generated material, being required to use AI, disclosure and credit?
  Partially verified: the curator re-read S1, S2; S3, S4 were not re-read by the curator and rest on the producer's pinned reading. For MBA-covered work, the WGA summarises four AI rules (AI is not a writer and its output is not literary material; companies may not require writers to use AI; companies must disclose AI-generated material they furnish; such material is not source material for credit or separated rights), while the binding Article 72 of the 2023 MOA is narrower and conditional (definitions, company policies, consent before use, semi-annual meetings). The WGA page also advises refusing AI transcription of pitch meetings.
• articleWhich VBench-2.0 dimensions can I run on my own clips (for example previs shots), what do they need, and which only work on the standard prompt suite?
  At the pinned repo commit, only five of the 18 dimensions accept custom videos: Human Anatomy, Human Identity, Human Clothes, Diversity (needs 20 clips per prompt) and Multi-View Consistency. The other 13 (including Camera Motion, Instance Preservation, Motion Rationality and all Controllability and Physics state-change dimensions) assert-fail on custom input and need the standard prompt suite, which means sampling the model under test on fixed prompts (3,860 videos for a full run by my count). Requirements are heavy (CUDA 11.8, flash-attention build, several third-party weights, some from Google Drive/Dropbox).
• articleWhat does VBench-2.0 actually measure (five categories, 18 dimensions), how is each dimension scored, and how is a single Total Score built?
  VBench-2.0 (arXiv v2) scores generated video on 18 dimensions grouped into Human Fidelity (3), Creativity (2), Controllability (7), Physics (4) and Commonsense (2). Scoring mixes generalist judges (LLaVA-Video-7B captions or yes/no questions, Qwen2.5-7B-Instruct as text judge) with specialist detectors (ViT human-anomaly detectors, ArcFace, CoTracker heuristics, a Qwen2.5-VL-3B LoRA anomaly detector). The paper reports per-dimension scores for four systems and no aggregate; the equal-weight Total Score is defined only in the repository README and used by the leaderboard code.
• articleWhich public test-media sources give a licence statement per file, which only per collection, and which should an L10 experiment avoid unless it checks the individual file?
  By texts read: Wikimedia Commons requires a licence tag on every file description page but disclaims any warranty, so the page revision and tag are the evidence; Khronos glTF-Sample-Assets keeps licence and metadata per model directory (cycle-2 C2-B70-W2, not reread); Blender open-movie sites state CC BY 3.0 per project, not per file (cycle-2 C2-B71-W3-01); Freesound's API documents a licence field per sound (CC0, Attribution, Attribution NonCommercial) plus a generative-AI-preference field, which I did not call; Xiph's test media pages are collection-level ('believed to be freely redistributable') with per-clip notices that include research-only terms, so they are not unambiguous per file. Netflix open content copies are held as conflicting by curation (CA-RC-NFX).
• articleWhat must an agent record alongside a ComfyUI workflow so it can be re-run faithfully later?
  The two JSON forms leave out most of what decides the output. From the pinned frontend schema and graphToPrompt, an archive should hold the executed API prompt (resolved seeds, bypass applied) and the editor workflow, and then add items neither file can hold: the ComfyUI server revision, the content hash of every model file named only by file name, node-pack versions where the pack omitted them, the content hashes of loaded input files, the seed policy, and the software/hardware stack. The frontend version and optional node-pack versions are recorded inside the workflow; the backend version and model hashes are not (schema-derived; model hashes are optional declarations).
• articleWhich training datasets does each of Real-ESRGAN, FILM, RIFE and Practical-RIFE itself name, and where is that stated?
  Real-ESRGAN names DIV2K, Flickr2K and OutdoorSceneTraining (OST) in its paper and DF2K plus OST in docs/Training.md. FILM names Vimeo-90K as its main training dataset (README) and as training data (paper). RIFE names Vimeo90K in its paper and README. The Practical-RIFE README at the pinned commit names no training dataset, so its data lineage is unknown from that file. REDS is not named in the four model repositories read.
• articleHow can an agent check a shot description or generated frame for internally consistent focal length, camera distance and perspective?
  A numeric consistency check can be built from two pinned statements: for a rectilinear lens, field of view follows 2*atan(sensor_diagonal/(2*focal_length)), which reproduces the Zeiss figures (15 mm: 110 degrees, 24 mm: 84, 35 mm: 63 on a 43.2 mm image circle; recomputed by this worker, see finding F1), and for a fixed subject size the distance scales with focal length (Zeiss). The agent compares the stated lens with the described framing (how much of the scene) and the stated distance (subject size, background apparent size). No source tests this on generated frames; the application to AI-generated frames is a proposed hypothesis with a test below.
• articleWhat terms and scope do the public TD-MPC2 code, checkpoints and datasets state, and what is left open?
  TD-MPC2 has the clearest statements of the three planning systems: the repository README says the project is MIT-licensed (LICENSE: MIT, Copyright Nicklas Hansen 2023) with third-party code under its own licences; the Hugging Face model card says in prose '**License:** MIT' for the 324 checkpoints (12 multi-task from 1M to 317M, 312 single-task at 5M), and the dataset card says MIT for the 30-task and 80-task replay-buffer datasets (545M and 345M transitions). The card's out-of-scope note says checkpoints are not expected to generalise to unseen tasks. Not stated: redistribution of simulator-derived data, model_training on outputs, and the terms of the underlying simulators (DMControl, Meta-World, ManiSkill2, MyoSuite) and of MuJoCo assets.
• articleBefore I put AI-generated Foley into a delivery, what per-component rights and provenance checks do the sources imply an audio post agent must make?
  A synthesis from the pinned licences and cards: (1) record model, checkpoint and revision; (2) read the weights licence (CC BY-NC, Tencent community licence with Territory limits, or conflicting statements) against the delivery's commercial status; (3) check output clauses (Tencent: no ownership claim but no use to improve other AI models and no use outside the Territory); (4) check each auxiliary component (VAE, Synchformer, vocoder, CLIP) for its own licence; (5) note training-data statements the card disclaims (MMAudio) or omits (HunyuanVideo-Foley paper); (6) log provenance marks (Google V2A research applies SynthID; Veo 3 card cites SynthID) and any required notices; (7) escalate conflicts to a human rights owner. This is a checklist proposal, not legal advice.
• articleWhat do the authors of video-to-audio models say the models get wrong, so I know what not to ask of them?
  Documented failure modes cluster: speech and lip sync (MMAudio 'unintelligible mumbles'; Google V2A 'uncanny lip-syncing'; Movie Gen does not support voice), dense, small or occluded motion (Movie Gen: tap dance, footsteps), unfamiliar sound concepts (MMAudio: 'RPG firing'), unrequested music (MMAudio), dependence on clean input video (Google V2A), and long-form or precise-timestamp reasoning (ThinkSound's MLLM). These come from the model authors' own statements; none was reproduced.
• articleI have a silent or AI-generated video clip and want a sound track: what kinds of video-to-audio systems exist, what do they take in and put out, and which of them can I actually run?
  Primary sources split into measured open systems (MMAudio, HunyuanVideo-Foley, ThinkSound: video plus optional text in, audio out, weights on Hugging Face), a measured closed system (Meta Movie Gen Audio, 13B, 48 kHz, up to 30 s per pass plus extension, described as research-only), and vendor showcases (Google V2A blog, Veo 3 joint audio) that give no sample rate, duration or sync metric. Sample rate and clip length differ by system; only the open ones can be inspected or run, and their licences differ (see G033 candidates).
• articleBeginner: what minimum evidence should an agent require before telling a production that a world model 'worked' on a job, and which found cases meet it?
  PROPOSAL (synthesis, not a tested standard): require eight items before saying a world model worked on a job: system and version, production date, task, inputs, outputs used in the final cut, a comparison baseline that was actually run, who measured and who wrote the claim, and what was not measured. Mapping the cases found: none of the generative cases meets items 1, 5 or 6; the only case with a run alternative (DOSENMUTTER) is not a world-model case. Until a case meets the checklist, an agent should say 'vendor reports' and not 'it worked'.
• articleWhat acceptance check could an agent run before relying on a streaming video or world model for a long take (proposal, not a tested protocol)?
  No source gives a ready acceptance test. Combining measures the papers use gives a proposal: generate at the longest length you need, score per-segment brightness/imaging quality and a motion measure, compute first-vs-last delta, run a copy-first-clip control, and separately test whether the model returns to an earlier place (memory), which the papers say these methods do not provide. This is synthesis, untested.
• articleWhat do the raw wpt.fyi report.json records say for the WebVTT vertical-text and ruby reftests at WPT 5d23b743bb487d18c51ea0f016b4b22c0895a813, and can a failure message tell you what is wrong with the rendering?
  Partially verified: the curator re-read none of the cited sources; S11, S12, S13, S14, S15, S16 rest on the producer's pinned reading. Across the four engines at WPT 5d23b743, the six WebVTT vertical-text and ruby reftest files give 24 results with exactly one pass (Firefox, bidi_ruby.html); all five vertical-text reftests fail everywhere.
• articleWhat does the Oral History Association say to do before, during and after a recorded interview (consent, preparation, narrator review, representation), and what is its stated scope?
  Corrected 2026-10-23 - see C3-B688-04; the original statements below are kept for history and are superseded where the correction says so. The Oral History Association's practice (adopted October 2018) runs consent as a process. Discuss the project and record the agreement before recording, prepare an open-ended guide, and respect a narrator's right to refuse topics and restrict access. Get a signed release, ideally at the end of the interview, and where possible let the narrator review and approve the recording or transcript before use, including deleting, restricting or redacting parts. Anyone using interviews should not manipulate the narrator's words and should give context. These are oral-history norms, which differ from journalistic practice.
• articleHow do design-tool colour-harmony rules (Procreate handbook example) compare with O'Donovan et al.'s large-dataset findings, and what limits do the authors state?
  Corrected 2026-10-23 - see C3-B688-07; the original statements below are kept for history and are superseded where the correction says so. Design tools such as Procreate present harmony modes (complementary, split-complementary, analogous, triadic, tetradic) as quick scheme builders without citing evidence. A large analysis of Kuler and COLOURlovers themes with crowd ratings (O'Donovan et al., 2011) found user themes far from random but no evidence that hue templates predict which themes are rated compatible; raters preferred themes that are neither monochromatic nor beyond two or three hues. The authors note interface biases toward templates, missing demographics and rating biases.
• articleWhere do oral-history norms for narrator review and documentary filmmakers' own reported interview practices diverge, and what do filmmakers report about repeating interviews and releases?
  Corrected 2026-10-23 - see C3-B688-04 and C3-B688-05; the original statements below are kept for history and are superseded where the correction says so. In a 2009 study of 45 long-form interviews (41 US directors or producer-directors), documentary makers said most subjects sign releases giving them complete editorial control; some show subjects material early, but the final decision stays with the filmmaker, and they reported overriding a signed subject's request when they judged the subject to have more social power. One director regretted pressing a subject to repeat a painful interview. The Oral History Association, by contrast, expects narrator review and agreement before an interview is made public and cites a documentary controversy over exposure risks. The two fields state different defaults (producer synthesis, supported).
• articleWhat do Google's public pages say SynthID can detect, how robust it is, and under what terms can its code or detector be used?
  Partially verified: the curator re-read S21, S22; S19, S20, S23 were not re-read by the curator and rest on the producer's pinned reading. Of Google DeepMind's SynthID family, only the text watermark has public code: an Apache-2.0 reference implementation that is explicitly not for production, with a production version in Hugging Face Transformers. DeepMind's image, video and audio watermark claims (robust to cropping, filters, compression) give no numbers, and per the producer's reading Google's developer docs say text detection is probabilistic and weakened by rewriting or translation.
• articleIs sox_ng actually confirmed vulnerable to CVE-2026-111157 (MS ADPCM WAV heap overflow), and did the maintainer commit to a fix? As of 2026-10-23 the public thread shows a reproduced crash but also unresolved doubt about whether the report matches sox_ng code.
  Problem. Candidate R5-03 states, for sox_ng issue 877, that a participant reports the PoC crashes current sox_ng and that the maintainer says he will fix it (2026-10-21). Conditions. Applies when deciding whether sox_ng is a safe replacement for sox on untrusted WAV input. Method. Re-read the issue and all comments via the Codeberg API on 2026-10-23. Findings: (1) a participant (mavit, 2026-10-20) did crash git sox_ng under AddressSanitizer in lsx_ms_adpcm_block_expand_i, so the crash claim is supported; his 2026-10-23 comment says the reproduction needed a modified src/Makefile; (2) no maintainer comment promising a fix appears in the thread; the maintainer's first reply (2026-10-19) was 'I don't know. Is it?' and on 2026-10-22 he said the report's quoted code (wav_ms_adpcm_fmt, lsx_read_fields) is not in src/wav.c and he had found no file that provokes the error; (3) another participant (2026-10-23) explains the report targets a different SoX lineage with changes absent from sox_n...
• articleDo O'Donovan et al. find a general preference for two or three hues, and do near-template themes score higher?
  Problem. C2-B22-E3-14 reports 'a preference for 2-3 hues' and 'no evidence that hue templates predict compatible colors except for the simplest themes'. Correction (overstated/missing dataset split). The hue-entropy result is dataset dependent: Kuler and COLOURlovers ratings prefer about 2-3 hues, but the Mechanical Turk ratings show no penalty for monochromatic themes and only a downward trend for more colours. Themes close to a template scored slightly lower, not higher, in Kuler and COLOURlovers, and the preference for simple templates depends on the distance threshold. Limitations. Read from pdftotext of the 2011 paper. Version/date. SIGGRAPH 2011.
• articleWhen a REUSE.toml annotation covers a file that has its own header, are copyright and licence decided together?
  Problem. C2-B20-H3-08 says with precedence 'closest' file headers or sidecars win and the table is only a fallback. Correction (missing limitation). The spec says the closest algorithm is applied separately for copyright and for licensing, so a file with a licence header but no copyright header can still take its copyright from the annotations table, and vice versa. Conditions. REUSE spec 3.3, annotations tables with default precedence. Limitations. Behaviour of the reuse tool itself was not tested. Version/date. Spec 3.3, 2024-11-14.
• articleDoes Honest Truths show documentary makers hold complete editorial control, with no narrator review?
  Problem. C2-B17-A3-05 contrasts OHA narrator approval with filmmakers' releases that 'usually grant complete editorial control' and 'some show subjects the film early'. Correction (missing limitation). The report says the release terms are usually dictated by insurers, and that filmmakers, perhaps because the terms were not their own, often gave subjects more leeway than the release, including a right to change their minds or see material in advance. Gordon Quinn's quote is about showing the film before it is finished. The contrast with oral history is real but narrower than 'some'. Also, the Pollard anecdote: he felt bad about pushing a subject to a breakdown, yet the moment stayed in the film. Limitations. Interview study of 45 interviews, 2009. Version/date. September 2009.
• articleDoes the OHA say narrator review and approval before release is required?
  Problem. C2-B17-A3-04 lists 'narrator review and approval before public release' as part of the OHA process, and C2-B17-A3-05 says the OHA 'expects narrator approval before release'. Correction (overstated). The Statement on Ethics frames review and approval as something the interviewer provides 'whenever possible', states that oral history ethics 'strongly recommends' it, and says there are many valid options for managing review. It is a strong norm, not an unconditional requirement. Limitations. Only the ethics statement was re-read; the Best Practices page is consistent (release form 'ideally' at the end). Version/date. Pages fetched 2026-10-23; PDF editions state 'Adopted October 2018'.
• articleAre DnR v3 training data and its model weights public, and under what licence? (correction to R1B-03)
  Problem. Cycle-1 candidate R1B-03 says the DnR v3 training set record and model implementation were not located and only a Basque test set (CC BY-SA 4.0) was found. Conditions. Matters when choosing a dialogue/music/effects separator and its training data. Method. Zenodo now lists (a) a 'Divide and Remaster v3: Multilingual Training Set' record (id 12713709, created 2024-12-20, licence cc-by-sa-4.0) whose description warns the Zenodo copy is corrupted and points to Netflix and Hugging Face mirrors, and (b) a 'Model Weights' record for the DnR v3 paper (id 12701995, cc-by-sa-4.0). Failure modes. Treating v3 as unavailable, or downloading the corrupted Zenodo copy. Limitations. The mirrors were not opened, the weights were not inspected, and CC BY-SA 4.0 on the record does not settle the licences of the underlying LibriVox/FSD50K/other sources. Version/date. Checked 2026-10-23.
• articleWhich formats can c2pa-rs sign today, and is Ogg Vorbis/Opus support a pull request or an issue? (correction to P3-02)
  Problem. Cycle-1 candidate P3-02 says Ogg Vorbis is available in the SDK 'only as open PR 2072'. Conditions. Applies to anyone tracking c2pa-rs Ogg support. Method. Re-fetch of github.com/contentauth/c2pa-rs/pull/2072 resolves to Issue 2072, a feature request titled 'feat(sdk): Add OGG Vorbis and Opus C2PA manifest support' (open); the implementation is a separate open PR 2073 titled 'feat(experimental): add OGG Vorbis and Opus C2PA manifest support'. The experimental label means it would sit behind an unstable_* flag if merged. Failure modes. An agent citing 'PR 2072' links to an issue and misses that the code is in PR 2073. Limitations. Release 0.91.5 (23 Oct 2026, a mov type-check fix only per its changelog) now supersedes 0.91.4, the version the candidate was written against; no handler change seen. Version/date. Checked 2026-10-23.
• articleDoes the SOFA 'Files' catalogue record a licence or terms field per database, and how consistent is it?
  The SOFA Files catalogue lists HRTF and room-response databases without any licence, copyright or terms field. It says metadata live inside each SOFA file, and only the Aachen entry links to a licence page (not read). Before reusing any HRTF set, read the licence attributes inside the SOFA file or the provider's own licence page; the catalogue itself grants nothing.
• articleWhat licence texts do the YarnSpinner core, Unity, Godot and Unreal repositories carry at the pins used?
  At the pins read, YarnSpinner core and the older YarnSpinner-Unreal repository carry MIT licence texts, while the newer YarnSpinner-UnrealEngine plugin carries the custom Yarn Spinner Public Licence v1.0 with a competing-product restriction, a no-AI-training clause and restrictions on redistributing unmodified source; 'Yarn Spinner is MIT' does not describe the newer Unreal plugin.
• articleHow machine-readable is per-asset licence and credit metadata in the Poly Haven API, the Sketchfab Data API, OpenGameArt pages and Kenney pages, and what must an importer parse out of free text?
  Poly Haven's API gives per-asset authors and descriptions but no licence field (CC0 is stated site-wide; only 3 of 521 model descriptions mention it), so an agent must take the licence from site terms. Sketchfab exposes a structured licence object on model detail, but unfiltered search results for non-downloadable models carry a null licence, so licence presence depends on downloadability and endpoint. Author credit is a structured field only on these two services among those surveyed.
• articleWhat target formats and signal-capture practices does IASA-TC 06 (2019) set for video digitisation, including frame-level fixity?
  IASA-TC 06 (2019) gives several acceptable preservation targets for analogue videotape (AVI/QuickTime with FFV1 or v210, Matroska/FFV1, MXF uncompressed or JPEG 2000) with growing institutional consensus on Matroska/FFV1, notes AS-07 became SMPTE RDD 48 in 2019, describes frame-level fixity in sidecars (ffmpeg framemd5/framecrc) or embedded, and treats VTR alignment with TBC, proc amp, waveform monitor and vectorscope as part of signal capture.
• articleHow do font files carry licence information (OpenType name IDs 13/14, OFL text and Reserved Font Names), and what do OFL documents say about conversion, WOFF and subsetting?
  OpenType name IDs 13 (License Description) and 14 (License Info URL) are free-text licence fields (page dated 2024-05-31). OFL-FAQ 1.1-update7 treats WOFF/WOFF2 conversion as unmodified only if the font data is unchanged apart from compression and metadata is omitted or carried unaltered, and treats webfont subsetting as modification.
• articleWhat is the ISDCF digital cinema naming convention (DCNC) structure, and how is it checked by tools?
  ClairMeta (commit b3b66a0) encodes the ISDCF naming convention as rule set '9.6' with fields in order: title (up to 14 characters), content type, aspect ratio, language (audio-subtitle, optional CCAP/OCAP), territory-rating, audio type, resolution, studio, date, facility, standard and package type. Reading the code (not run) shows two problems. Its language patterns accept only two- or three-letter codes, but the current ISDCF register has 91 four-letter codes, so names using them would probably be flagged. Its resolution pattern (^2K|4K$) is loosely anchored, so strings that merely start with '2K' or end with '4K' would pass. This is a tool's reading of the convention, not the convention itself.
• articleRights warning pattern in creator tutorials: celebrity likeness and third-party characters used as test subjects, sometimes via hosts with looser filters than the model maker's own app
  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. Three of these tutorials use protected material as casual test content: a close-up portrait of a named actress that Google's own Gemini app refused but a third-party host produced ("because of the the restrictions you have inside of Google, I couldn't make that image") (R2); a Pokemon character composited into a scene (GPT Image 2 complied; Midjourney refused under community guidelines) (R1); and an alien adapted from a commercial animated series as a design reference (R4). Implications: (1) the same underlying model can have different safety filters on different hosts, so "it worked on platform X" is not permission; (2) right-of-publicity/likeness and character copyrig...
• articleMedia business literacy from an author channel: why "bestseller" is everywhere (list segmentation, print-run skew) and what a 2026 studio adaptation deal looked like from the IP holder's side
  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. R4 (Sep 2026, 3.4M views - the channel's top in-window upload): the common claim that lists are easily gamed is mostly wrong (BookScan gives sales counts; NYT, USA Today and BookScan rank within a few places of each other). Real reasons "every book is a bestseller": many segmented lists (NYT splits YA/adult, formats, fiction/non-fiction, graphic novels; USA Today is unsegmented; Amazon sub-sub-category lists mean less); the frequency illusion; and above all print-run skew - a mega-seller run of ~400,000 copies vs ~10,000 lifetime hardcovers for a non-bestseller means most copies on shelves are bestsellers. Legit tactic: release in a low-competition month (February vs No...

More