Evaluation and media provenance

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

• articleSame creator, two video-model shootouts six weeks apart, opposite verdicts on Gemini Omni Flash: test conditions (resolution cap, start-frame use, prompt specificity) decide rankings more than the models
  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. Two Higgsfield-hosted comparisons by Youri van Hofwegen. Aug 2026 (R2): five models at each model's MAX settings (Seedance 2.0 native 4K 15 s, Sora 2 Pro Max 12 s 1080p, Kling 3.0 4K, Happy Horse 1080p, Gemini) on a talking-head lip-sync test and a breakdance motion test, scored 1-10. Verdict: Seedance top on both; "Gemini is the real surprise" top-two while cheapest (30 credits vs Kling 90, Happy Horse 68, Sora 153, Seedance 330 - Seedance partly because of native 4K); Kling 3 lip sync "pretty bad" (5/10); Sora bottom. Sep 2026 (R1): four models BLIND-labelled A-D, all capped at 10 s, 16:9, 720p with audio "otherwise this becomes a test of which model has the biggest r...
• articleHow a top AI-tools channel ranks video generators (Nov 2025-Jul 2026): same-prompt rounds, audio as a scoring factor, and "cost per second of usable video" rather than credits per generation
  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. Youri van Hofwegen's ranking format (condensed reads; videos use OpenArt or Higgsfield as the test host and link to them). Method: identical prompt across models per round (universal realism scene, multi-character fight, dialogue), audio quality counts in the score (poor audio drops a model even if visuals are good), and a final VALUE round. His key value rule: compare how much usable video each generation yields, not credits per generation - a cheaper call that returns a shorter clip can cost more per second (R1); in Jan 2026 he contrasted a top realism model costing ~3,000 OpenArt credits with a near-equal alternative at ~400 credits per 10 s (R3); in Nov 2025 he rate...
• article"Face swap any video" tutorials and permanent digital personas: what the workflow enables and the consent/disclosure lines it crosses
  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. Youri van Hofwegen (2026-02-19, R4) demonstrates replacing his own face in recorded video with a generated person (Nano Banana Pro for the identity frame, then a video tool with lip sync), auto-captioning, and then making that synthetic person "perform" trending dances and scripted ads so a "permanent digital influencer" can work "for free". Assessment: swapping one's OWN performance onto a fictional face is legal in many places but still realistic synthetic media that platforms require to be disclosed (TikTok C3-B398-01; YouTube altered/synthetic disclosure), and EU AI Act Art. 50 may require visible deep-fake disclosure for professional/monetised use (C3-B426-01). Swa...
• articleRunning synthetic "AI influencer" accounts (Feb-Jun 2026 creator practice): consistency methods, a local dashboard, and the disclosure obligations creators tend to skip
  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 says he runs about five AI-influencer pages, some with millions of impressions a month, and built a free local dashboard that stores nine personas (each in a niche, e.g. a yoga instructor) and one-click generates character sheets, close-ups and "feature" sheets, then scenes for posting (R3). Consistency methods (R2): detailed prompts for skin texture and imperfections to avoid a plastic look; generate a background grid with the character as reference and upscale a chosen panel; edit with "keep the exact same person, face, skin texture, expression, pose, background, proportions and camera framing" and change only clothes or pose; pose transfer by giving two ref...
• articleA blind creator test (Jul 2026): Seedance 2.0 4K won most rounds, Grok Imagine best dialogue, Kling 3.0 "80% of the quality at a quarter of the cost"
  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 (Higgsfield-sponsored, 2026-07-05). A teammate generated clips so he ranked them blind: Seedance 2.0 at 4K high bitrate, Kling 3.0 at 4K high bitrate, Grok Imagine at 720p, Google Omni at native 720p (upscalable). Tests: image-reference POV dialogue, text-to-video, motion control (Grok lacks it), editing (Omni expected to win; he ranked Seedance first, Omni second, Kling third), dialogue/lip sync (Grok first, Seedance second, Kling third), multi-reference (Seedance first). Overall Seedance took the most golds, but he notes it is among the most expensive; Kling 3.0 offers 4K and 60 fps cheaply - "80% of the quality with 1/4 of the cost" - and some shots were ea...
• articleCreator licensing claims vs provider terms: a Mureka tutorial promises "100% royalty-free... commercial license... no copyright strikes" while Mureka's terms are narrower
  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. Mikkel Lassalle's Mureka video (2026-03-03, R8) tells viewers that all generated tracks are "100% royalty-free and come with a commercial license" and that "you don't have to worry about any copyright strikes". Mureka's own terms as fetched (C3-B413-01, R13) say: paid-tier users own Output; FREE-tier Output is owned by Mureka; the general service licence is for personal, non-commercial use; Mureka does not guarantee copyright registrability and the user bears legal risk; using Mureka's distribution service grants Mureka an exclusive licence. No term read guarantees freedom from copyright claims (Content ID or otherwise). Lesson for curation and creators: a sponsored tut...
• articleWhat do later sources add or criticise about applying the twelve animation principles beyond 1987 3D keyframing?
  Partially verified: the curator re-read S15; S13, S16 were not re-read by the curator and rest on the producer's pinned reading. A systematic review of 27 human-robot interaction studies (arXiv v4, 2019; arXiv non-exclusive licence) frames robot motion with the twelve principles and its count table shows no reviewed study for Timing, Staging, Solid Drawing or Appeal, while Secondary Action and Straight Ahead/Pose to Pose were most used.
• articleWhat human rating criteria does HANNA use for story quality and how consistent are raters?
  HANNA defines six human criteria (relevance, coherence, empathy, surprise, engagement, complexity) on a 5-point Likert scale and annotates 1,056 stories from 96 WritingPrompts prompts (human story plus 10 systems) with 3 MTurk raters each. Averaged-rater reliability (ICC2k) is only fair to moderate (0.28-0.56), the criteria are weakly to moderately intercorrelated (Kendall 16%-62%, mean 40.7%), and human stories scored highest on every criterion.
• articleHow do public specs label music and effects stems, and what rights information accompanies stems versus cue sheets?
  Delivery specifications label stems by what they contain, not by who owns them: AS-11 X9 uses MCA tags such as Music and Effects (ggME) and Audio Description (ggAD), EBU R 123 calls the M&E bed International Sound, and Paramount requires Atmos beds and objects separated at least into dialogue, narration, music and effects. None of the stem rules read carries rights metadata; identifiers and ownership (ISRC, ISWC, society, owner and shares) live in music cue-sheet formats such as CISAC AVR+.
• articleDoes AES77-2023 set different loudness targets or peak limits from AES TD1008, and what can I read for free?
  No open source read states the numeric targets in AES77-2023, which is a paid standard. A trade report says AES77 is based on TD1008 and does not recommend loudness range or device playback targets; TD1008's own figures (-18 LUFS speech, -16 LUFS track-normalised music, -14 LUFS for the loudest album track, at most -1 dBTP at the codec input) are public and already reviewed. Whether AES77's numbers equal TD1008's cannot be confirmed from open sources.
• articleDo the AMIA open-workflows list and OSSArcFlow materials define AV normalisation or conformance checking steps?
  The AMIA open-workflows README is a curated list of links (Archivematica case studies, OSSArcFlow, RAWcooked DPX workflows, MediaConch, NYPL specifications) and states no normalisation targets or conformance rules itself; AV normalisation and conformance parameters have to be taken from tool-level sources such as Archivematica's format policy registry or MediaConch policies (negative result for this list).
• articleWhat exactly does WCAG 2.2 SC 1.4.3 require for text contrast, and how do I apply it as a check to AI-generated posters, title cards and thumbnails?
  Problem. Generated backgrounds are busy, and text laid over them often fails contrast. Conditions. WCAG 2.2 Understanding document for SC 1.4.3 (Level AA); it governs web content, so for posters and thumbnails it is a benchmark, not a legal rule. Method (documented). Text and images of text need a contrast ratio of at least 4.5:1, except large-scale text at 3:1; the Understanding text says 18 point text or 14 point bold text is judged to be large; computed values are not rounded (4.499:1 fails 4.5:1); text that is part of a logo or logotype is exempt; the document notes images of text do not scale well because they pixelate. Relative-luminance formula details are in C3-B066-G25-04 and C2-B22-E3-03. Checks on AI graphics [inference]. (1) After setting type (C3-B593-03), sample the background pixels under each glyph box, not the average of the image, and compute the ratio for the lightest and darkest background pixel against the text colour; the worst case must pass. (2) Gradients and...
• articleHow much do pixels differ across GPUs and browsers for the same WebGL or WebGPU scene, and what tolerance do conformance tests allow?
  A 2012 study rendered one simple antialiased WebGL scene on 270 crowd-sourced machines: the images looked identical but formed 50 distinct pixel-level renders, mostly consistent within a graphics-card model (the text also mentions 51 groups). Current WebGPU conformance tests assume small differences: integer formats compare exactly by default, float and normalised formats can use ULP or fractional thresholds, a canvas colour-space reftest allows up to 2 levels per channel, and WGSL leaves the rounding mode open when quantising. So pixel-exact equality across devices should not be expected; no recent cross-device measurement was found.
• articleWhat did Smith & Henderson's eye-tracking study find about continuity editing rules (match action, gaze match) and unnoticed cuts, and what does it mean for cutting AI shots?
  Problem. Continuity rules are taught as craft; is there evidence that they hide cuts? Conditions. Smith & Henderson (2008): participants watched seven feature-film excerpts (five minutes each, DVD-quality) and pressed a button when they saw an edit; eyes were tracked; edits were hand-coded for adherence to the rules, including Between Scene, Within Scene, Match Action and Gaze Match. Method (reported). Edits constructed according to the Continuity Editing Rules gave greater edit blindness than edits not adhering to them; a quarter of edits joining two viewpoints of the same scene went undetected, a third when the cut coincided with a sudden onset of motion. The Match Action rule is stated in the paper as a smooth cut between two viewpoints of the same action achieved by coinciding the cut with a sudden onset of motion. Only 8.8% of edits coincided with a saccade, and the authors conclude that most unnoticed cuts come from inattentional blindness as viewers follow the narrative. For ...
• articleWhat do the VibeVoice authors say about long-form multi-speaker generation: length, speaker count, evaluation, limitations and misuse risks?
  Problem: choose a model for podcast or multi-participant audio. Conditions: dialogue transcripts with speaker labels. Method (VibeVoice technical report): a 7.5 Hz continuous speech tokenizer plus an LLM with next-token diffusion; up to 90 minutes in a 64K context window with a maximum of four speakers. Evaluation (authors'): 8 long transcripts totalling about 1 hour; WER via Whisper-large-v3 and Nemo, speaker similarity via WavLM, and MOS from 24 annotators on realism, richness and preference, each listening to about six hours; comparisons include ElevenLabs v3 alpha and Gemini 2.5 Pro TTS. Stated limits: English and Chinese only; no non-speech audio (background noise, music, sound effects); no overlapping speech; deepfake and disinformation risk; the authors do not recommend commercial or real-world use without further testing and describe the model as research and development only. Failure modes: unexpected audio for other languages; long output drift is not separately measured i...
• articleHow does Sesame's CSM use conversation context for speaker consistency, and what are its documented limits and misuse prohibitions?
  Problem: keep two speakers consistent across turns in generated conversation. Conditions: CSM-1B base model on a CUDA GPU, with Llama-3.2-1B and CSM-1B access. Method (README): CSM generates RVQ audio codes from text and audio inputs using a Llama backbone and a small audio decoder producing Mimi codes. With no prompt it uses a random speaker identity; to keep voices consistent, pass previous turns as Segment objects (text, speaker id, audio) as context, because the README says CSM sounds best when provided with context. The example call sets max_audio_length_ms=10_000, so long outputs have to be built turn by turn. The release is a base model, not fine-tuned to any voice, and it generates audio only (an LLM must write the text). Non-English is weak despite some capacity. Misuse: the README explicitly prohibits impersonation or fraud (mimicking real individuals without explicit consent), misinformation or deception, and illegal or harmful use. Failure modes: voice drift when context...
• articleWhat does Fréchet Audio Distance measure, how is it computed, and what did its original paper show about correlation with human judgments?
  Problem: judge a set of generated or enhanced audio clips without a clean reference per clip. Conditions: a reference set of good audio and an evaluation set (the sample-size findings below show N matters). Method (Kilgour et al., 2019): extract embeddings with a pretrained audio classifier (VGGish) from both sets, fit a multivariate Gaussian to each, and compute the Fréchet distance between them; lower is better. The paper proposed FAD as a reference-free metric for music enhancement and validated it against artificial distortions; it reports a correlation of 0.52 with human perception ratings versus 0.39 for signal-to-distortion ratio, -0.15 for cosine distance and -0.01 for magnitude L2 distance. The authors note VGGish ignores phase and uses mel-scaled bins, so some distortions could go undetected. Failure modes: it compares distributions, so it cannot judge a single clip; it depends on the embedding and reference set; 0.52 is a moderate correlation. Limitations: designed for en...
• articleDoes KAD fix FAD's Gaussian and sample-size problems, and how large must N be?
  Problem: choose a distribution-distance metric for small evaluation sets. Conditions: audio generation experiments with limited generated samples. Method (Chung et al., 2025): FAD assumes Gaussian embeddings, which often fail (the paper shows non-Gaussian VGGish embeddings), and its finite-sample bias scales as O(1/N), so FAD decreases as N grows and can be manipulated by changing N. Kernel Audio Distance uses Maximum Mean Discrepancy (distribution-free, unbiased) and the kadtk toolkit. Measured by the authors: on Clotho 2 with N from 100 to 3,839 the FAD deviation roughly halves each time sample size doubles, while KAD stays near its asymptotic value; on DCASE 2023 Task 7 Foley systems, KAD reaches a Spearman correlation of up to -0.93 with human ratings against -0.80 for FAD's best case. Failure modes: KAD has larger variance at small N; music-trained embeddings (MERT, CLAP-laion-music) correlated less well than others on Foley. Limitations: authors' own comparison, Foley sound no...
• articleIf I report FAD or CLAP for generated audio, what must I state and why do I still need a listening test? (inference)
  Problem: automatic scores look objective but hide settings. Conditions: any report of FAD, KAD or CLAP score. Method (inference from the round's papers): report the embedding and checkpoint, the reference set, N (number of clips and frames), whether bias was corrected, and the prompt set; do not compare FAD across papers with different embeddings or N; use per-song FAD to find outliers rather than as a verdict; use CLAP score only as a text-adherence proxy and name the checkpoint. Add a listening test with hidden reference and anchors (see C3-B570-15 for MUSHRA) because: original FAD correlated only moderately with human ratings (0.52); VGGish correlated poorly with music quality ratings in the Gui study (n=10 raters); and the human-preference study found model-level metrics can misrank individual models. Also keep automatic MOS predictors in context (C3-B570-14). Failure modes: optimising to a metric (Goodhart); reporting a single number. Limitations: no source gives a minimum N or...
• articleI have a finished master: what are the handoffs to a cinema/streaming package, to captions and audio description, and to an archive, and which public tool or text checks what at each one?
  Synthesis supported by reviewed steps: the chain from master to archive has distinct handoffs (package to IMF or DCP; adding timed text and audio description; validation; archive packaging), each with its own primary text and open checker, and a validator's pass covers only what it implements (Photon aims to validate IMPs for listed applications; libdcp and ClairMeta check DCP structure; BagIt 'valid' only matches manifest checksums). ClairMeta profiles can downgrade or silence checks, so record the profile with any pass. No single tool spans the chain.
• articleWhat rights and attribution metadata does IIIF Presentation carry and how do rights, requiredStatement, provider and metadata differ?
  IIIF Presentation 3.0 splits rights information three ways: 'rights' holds one Creative Commons or RightsStatements.org URI (or a registered extension) and is informative and machine-oriented, possibly shown only as an icon; 'requiredStatement' holds label/value text that must be displayed when the resource is displayed or used; 'provider' gives structured data about who supplied the content. Institution-specific wording therefore goes in requiredStatement or metadata, and a manifest without 'rights' asserts nothing about status. Presentation 2.1 used 'license' (URI) and 'attribution' (text that must be shown) for the same split.
• articleWhat do Ofcom's Section One rules 1.11 to 1.16 say for children's programmes, and is there a decided case on Rule 1.13 or 1.15?
  Under the Ofcom Broadcasting Code (January 2019 edition), easily imitable harmful dangerous behaviour must not appear in programmes made primarily for children without strong editorial justification (and not pre-watershed without editorial justification), and offensive language must not be used in programmes for younger children except in the most exceptional circumstances (Rules 1.13, 1.15). Ofcom found a 2008 pre-watershed stunt-clip programme in breach of 1.13 for insufficient editorial justification; applying that case to scripted children's drama is an inference. The May 2026 consultation on a Tier 1 Standards Code for large on-demand services keeps existing rules as sufficient for protecting young audiences and adds protections for under-18s who appear in programmes.
• articleWhat practical steps and contract wording does Equity (UK) publish for performers asked to be digitally scanned or voice-captured, and what legal framing does it give?
  Equity (UK) advises performers to: ask five specific questions before engagement (face/body scanning, voice capture, synthesised performance, data mining and related uses, each with a description of intended use); limit any capture to one named production; negotiate an AI-training carve-out; refuse in writing at contract stage if unwilling; not sign scanning permissions on set; have scans deleted once the production is complete; and keep up to date and share the advice. Equity says no UK law is written explicitly for AI and points to copyright, performers' rights and data protection. Its vision statement sets eight cloning principles, from consent for past, current and future performances to engagement under collective agreements. These are union advice and proposed clauses; the pages report no outcomes.
• articleWhat did the primary drama-management papers actually evaluate, and how can an agent test a branching story or drama manager with what they report?
  Partially verified: the curator re-read none of the cited sources; S1, S3, S2, S7, S4 rest on the producer's pinned reading. The inspected papers evaluate very differently. Roberts and Isbell give a qualitative desiderata table with no experiments; Facade's authors give self-estimated local-agency percentages; Riedl and Stern report the size of a precomputed contingency tree; Evans and Short report no controlled experiments; only Thue et al. ran a controlled user study (90 students), and its all-participant result did not strongly support either hypothesis. Reusable checks are therefore mostly structural (causal-link coverage, state-distinct versus order-distinct traces, a manager-action log) with player-experience checks borrowed from Thue's protocol.
• articleHow did DreamGen use video world models to create synthetic training data for robot policies, and what did it measure?
  Partially verified: the curator re-read none of the cited sources; S10 rests on the producer's pinned reading. DreamGen (NVIDIA et al., arXiv v2, June 2025) fine-tunes an image-to-video model on a target robot, rolls it out from an initial frame plus instruction, recovers pseudo-actions with an inverse-dynamics model or latent actions, and trains a policy on these 'neural trajectories'. Reported results: on a humanoid with GR00T N1, 14 novel behaviours 11.2% -> 43.2% and 13 novel-environment tasks 0.0% -> 28.5% (policy trained only on neural trajectories plus a single-environment pick-and-place baseline); RoboCasa solely-synthetic training reaches 20.6% average over 24 tasks; policy performance rises log-linearly with number of neural trajectories. Costs and limits: 240k RoboCasa samples took 54 hours on 1500 L40 GPUs; initial frames are provided manually; the benchmark's automatic evaluator can hallucinate.
• articleWhat do the NVIDIA Cosmos paper and docs actually document for synthetic data generation and agent evaluation, versus only propose?
  Partially verified: the curator re-read none of the cited sources; S5, S7, S8, S9 rest on the producer's pinned reading. The Cosmos paper (v3) lists four uses of a world foundation model (policy evaluation, policy initialization/training, planning, synthetic data) and says it reports no empirical results for them. What it does measure are model-level checks: 3D consistency, physics alignment and post-training results for camera control, robotic manipulation (action-conditioned and instruction-conditioned video prediction) and driving. The docs (observed 2026-10-22) describe workflows with inputs/outputs: Cosmos-Transfer2.5 for two augmentation workflows (Simulations to Photorealism; Scale World State Diversity) with computed control maps; Predict2.5 action-conditioned post-training on Bridge end-effector poses; Reason1 best-of-N rejection sampling with a score = fraction of artifact-free evaluations. The docs show example videos and procedures but no downstream policy or accuracy re...
• articleWhat generation metadata is lost when a generated image is re-encoded, converted or uploaded, and how should an agent preserve it?
  Partially verified: the curator re-read none of the cited sources; S8, S9, S2, S1, S4, S3 rest on the producer's pinned reading. Embedded settings live in ancillary PNG text chunks (or EXIF/tags) that decoders ignore, so they exist only while a PNG-aware tool copies them. The PNG specification says ordinary image editors usually discard unrecognized information; the tools' own code writes nothing when their disable switches are set and writes different containers per format. Practical rule (synthesis): extract and store the metadata together with a hash of the original bytes before any resize, convert, screenshot or upload, because the sources inspected give no way to recover it afterwards. Behaviour of specific hosting sites was not verified.
• articleWhich benchmark checks whether a video or world-generation model follows a camera path and keeps a scene consistent, and does it measure physical laws?
  Partially verified: the curator re-read none of the cited sources; S3, S10, S11 rest on the producer's pinned reading. WorldScore (Stanford; arXiv v2 Nov 2025) evaluates 'next scene' generation: given a scene image, a next-scene prompt and a camera layout (a trajectory plus one of 8 text camera moves), it scores controllability (camera, object, content), quality (3D, style, photometric consistency, subjective quality) and dynamics (motion accuracy, magnitude, smoothness). It tested 20 models, including 13 video models; video models had weak camera control (best video 40.22 on camera controllability) and 3D scene generators beat them on static worlds. It does not define a physical-law metric, so it informs camera and consistency questions only.
• articleWhat do the three C2PA conformance-programme agreement forms (Generator Product, Validator Product, Certification Authority) say, as text, about disclosure of information, term and termination, warranties and liability?
  The curator re-read all cited web sources (S30, S32); nothing was executed. The three C2PA conformance agreement forms (v1.0, 25 June 2025) have no fixed term: the applicant can terminate on at least ten days' written notice and uncured breaches (thirty days) allow termination. The C2PA provides its specifications and policies as is, caps its total liability to the applicant at US $100 (the applicant's indemnification obligations excepted), and publishes a Conforming Products List; the certification-authority form has the applicant warrant its information to the best of its knowledge. A reading of the text only, not legal advice.
• articleDo the Objaverse-trained 4D generators (L4GM, SV4D, SV4D 2.0, Diffusion4D) say whether their training data overlaps the Consistent4D benchmark objects, and what should an agent assume when comparing their scores?
  Partially verified: the curator re-read none of the cited sources; S14, S13, S9, S5, S15, S16, S8 rest on the producer's pinned reading. The Consistent4D author states that six of the seven benchmark objects are in Objaverse and recommends filtering them from training. A set intersection I ran (synthesis, not executed code from the repositories) finds 6 of the 7 benchmark uids in a 37,789-ID list of animated Objaverse-1.0 objects published in the Diffusion4D repository, matching that statement. SV4D states it excluded the benchmark objects; L4GM v1 states that it renders all animated Objaverse 1.0 objects and evaluates on the benchmark, and the passages searched contain no exclusion statement; SV4D 2.0 v3 contains no such statement in the strings searched. Overlap is therefore an open question for L4GM, and possibly SV4D 2.0, on this benchmark, not an established fact: L4GM filters videos by motion, and the L4GM training list is not in its repository.
• articleWhat does the safetensors format define (header layout, __metadata__, dtypes, alignment) and how can an agent read a model file's header without loading weights?
  Corrected 2026-10-23 - see C3-B595-05; the original statements below are kept for history and are superseded where the correction says so. A .safetensors file is: 8 bytes (little-endian u64 N), N bytes of UTF-8 JSON header, then a raw byte buffer. The header maps tensor names to {dtype, shape, data_offsets[begin,end]} (offsets relative to the start of the byte buffer) and may contain a '__metadata__' string-to-string map. At the inspected commit the library accepts the dtypes listed in tensor.rs (BOOL through F64/U64, BF16, several FP8 and sub-byte types), caps the header at 100,000,000 bytes, requires contiguous gap-free offsets, and its writer sorts tensors by dtype alignment and pads the header with spaces to a multiple of 8. An agent can read the header with a few lines of standard code (no ML library) and learn tensor names, shapes, dtypes and metadata.
• articleHow can I tell a ComfyUI workflow (UI/graph) JSON from an API-format prompt JSON, and what differs in ids, links, widget values and version fields?
  Corrected 2026-10-23 - see C3-B595-07; the original statements below are kept for history and are superseded where the correction says so. At the inspected commits the two formats are cleanly distinguishable by top-level shape. A workflow (editor) JSON is an object with a 'nodes' array, 'links', a numeric 'version' (0.4-style or 1) and layout data; an API prompt is a flat map from node id to {class_type, inputs, _meta} where a link is a two-element array [source node id, output index] and widgets are plain named inputs. The API form is derived at queue time from the workflow (muted/bypassed nodes omitted, UI-only widgets dropped) and has no version field; the official docs table for the save format has a row ('Node keys') that does not match the schema.
• articleWhat do primary sources say, measured or claimed, about topology, UVs, polygon budgets and PBR correctness of generated assets for games, previs or VFX?
  Corrected 2026-10-23 - see C3-B588-05; the original statements below are kept for history and are superseded where the correction says so. No inspected source measures retopology quality (edge flow), UV distortion or engine polygon-budget compliance, and none tests PBR values against ground-truth materials of a held-out asset. What exists: SF3D reports triangle counts for one mesh per method (SF3D 27.4K vs TripoSR 32.1K, OpenLRM 662K) and a box-projection UV unwrap in 0.5 s, with remesher options (triangle, quad split to triangles); TRELLIS.2 reports PSNR/LPIPS on rendered PBR maps for reconstruction, offers GLB export with remesh and a decimation target (README example 1,000,000 faces, 4096 texture), and states small holes occur; SF3D and SPAR3D state material estimation is homogeneous or sometimes suboptimal; TRELLIS (first release) states baked-in lighting and no PBR; Hunyuan 2.0 describes a low-polygon stylization module in its hosted Studio. Production claims are the vendors'.
• articleWhat do open documents define for camera noise and dynamic-range test methods (ISO 15739 public scope, Imatest's documented definitions, EMVA 1288), and what linearity assumptions do they make?
  Dynamic-range figures depend on the method: Imatest reports DR at several scene-referenced SNR thresholds (20, 12, 6, 0 dB), and EMVA 1288 splits into a Linear release and a General release for non-linear cameras or cameras whose preprocessing changes temporal noise; any DR claim must name its method, threshold and encoding handling.
• articleWhat quality criteria, numeric thresholds and reporting duties do Ofcom, CRTC, ACMA and the FCC publish for captions, and which of them set numeric accuracy or latency values?
  CRTC 2019-308 sets a live-caption lag limit of six seconds averaged over the program and a NER accuracy rate of 98 for English-language live programming (with a transition period from 1 September 2019 to 29 February 2020) and annual reports by 30 November; it is a mandatory numeric standard, unlike ACMA's observed 6-second level per producer.
• articleDoes the official Unity registry confirm com.unity.recorder 5.1.7, and does the mirror's package.json at df67fd77 match the registry metadata for that version?
  Corrected 2026-10-23 - see C3-B598-03; the original statements below are kept for history and are superseded where the correction says so. The official UPM registry JSON (GET, observed 2026-10-22T12:21Z) lists com.unity.recorder 5.1.7 as dist-tags.latest, with registry time 2026-08-03T10:58:47.510Z and dist.shasum 0b23bb7a1fb4a062b8640ef5b1d174ca9b070467 (SHA-1 hex; no integrity/sha512 field in any of 42 versions). The mirror's package.json at commit df67fd77 equals the registry's 5.1.7 entry on every shared metadata field (name, displayName, version, unity, dependencies, description, keywords, samples, documentationUrl, _upm.changelog); the mirror file adds repository, upmCi and relatedPackages, the registry adds dist. This corroborates that the mirror's metadata is consistent with Unity's registry, not that its .cs bytes equal the archive.
• articleWhat scores do named systems (Sora 2, Veo 3.1, Gen-3, Hailuo, Cosmos, SVD, LTX-Video and others) get under each geometric-consistency metric, and which of those numbers can legitimately be compared?
  Corrected 2026-10-23 - see C3-B588-06; the original statements below are kept for history and are superseded where the correction says so. GeCo-Eval (80 static-scene prompts, 4 videos each, with motion columns) is the only inspected benchmark that scores Sora 2 and Veo 3.1; it puts Sora 2 and Veo 3.1 close together (object-centric fused 0.069 vs 0.090, indoor 0.169 vs 0.233, outdoor 0.265 vs 0.217; lower is better) with Veo moving more. WorldScore gives Gen-3 68.31 and Hailuo 67.18 on its 0-100 reprojection axis; SGC scores Cosmos 0.0722 and Latte 0.3226; the MEt3R paper has SVD at 0.032. Metrics, scales, prompts, clip lengths and motion differ, so the numbers form a table of separate experiments, not a ranking; the only visible overlap is model family names.
• articleHow do I write a SWHID that cites an exact revision, file path and line range, and which qualifier combinations are valid?
  A qualified SWHID adds ;key=value qualifiers to a core id. lines= (1-based, inclusive) or bytes= (0-based) are valid only on content, and at most one is allowed. origin, visit, anchor and path say where the object was seen; anchor requires path and visit requires origin. The canonical order is origin, visit, anchor, path, lines/bytes. The contradiction the producer found is real in the text it pinned (91af30d, which diverged from main): the rule said path was invalid on content while the examples used it. It was reported (issues #69, #71) and fixed on main by commit 45f8de0 (2026-09-19): path is valid only for content or directory. That fix also makes path on rev, rel and snp invalid, and implementations ignore invalid qualifiers.
• articleWhich model file formats can run code when loaded in ComfyUI, and what protections exist at the pinned versions?
  Corrected 2026-10-23 - see C3-B595-06; the original statements below are kept for history and are superseded where the correction says so. Pickle-based model files (.ckpt, .pt, .pth, .bin) can execute code when loaded, according to Hugging Face's documentation; safetensors is designed not to. ComfyUI at commit 46aad33e loads .safetensors/.sft files with safetensors.safe_open and everything else with torch.load(..., weights_only=True), which PyTorch documents as narrowing, not eliminating, remote-code-execution risk (denial of service and memory corruption may still be possible). ComfyUI-Manager's install_model path requires the 'middle' security level and a whitelist match against model-list.json, and for non-.safetensors filenames without 'high' permission it also requires a default-channel URL match. The ComfyUI frontend's missing-model download allowlist (huggingface.co, civitai.com and civitai.red prefixes plus a few exact URLs) still permits .ckpt, .pt and .pth suffixes. Huggin...
• articlePer component (tokens, raw video, weights, code) do the 1X cards, the GitHub README/LICENSE and the 1X blog agree on terms, and which uses does each text actually address?
  Corrected 2026-10-23 - see C3-B588-08; the original statements below are kept for history and are superseded where the correction says so. For each component every text read gives the same label: tokenised data Apache 2.0 (HF card, GitHub README schema block, 1X blog), raw video CC-BY-NC-SA 4.0 (raw card front matter, tokenised card prose, 1X blog), GENIE weights apache-2.0 (card front matter/API), code Apache-2.0 (LICENSE text). Different labels on different components are not a conflict. No text addresses hosting, redistribution, modification, training, commercial or output use except by the label itself; stated uses are purposes (experimenting with robotics data, challenge participation) and a contest rule. No rights conflict is filed.
• articleWhat did Schloss and Palmer (2011) measure about colour-pair preference and harmony, and what do they say limits generalisation to palettes and to design practice?
  Schloss and Palmer (2011) show that people judge colour pairs on three distinct questions: how much they like the pair, how harmonious it is, and how much they like the figure colour against its ground. Across 992 two-colour concentric-square pairs, liking and harmony both rise with hue similarity (the best harmony model, with hue similarity, coolness and desaturation, explains 67.3% of variance), while liking of the figure colour rises with hue contrast. The authors caution about other layouts and about combinations of three or more colours.
• articleDoes the Unity registry record for com.unity.recorder carry any integrity (sha512) field in its 42 versions, or only dist.shasum?
  Correction to C3-B222-G68-01 (nature: overstated absence claim). Original: no integrity/sha512 field in any of 42 versions. Audit result (registry JSON fetched 2026-10-23): version 5.1.7 still has only dist.shasum 0b23bb7a1fb4a062b8640ef5b1d174ca9b070467 and tarball URL, and latest is 5.1.7, as stated. But 7 of the 42 version entries (4.0.0, 4.0.0-pre.5, 4.0.0-pre.4, 4.0.0-pre.2, 3.0.3, 3.0.2, 2.5.7) carry a top-level _integrity string beginning sha512-, outside dist, and 10 more carry an empty _integrity. So the accurate statement is: no dist.integrity field in any version, no _integrity on 5.1.7, but historic entries expose leftover npm-style _integrity values. Treat them as publisher-supplied metadata of unknown reliability. Version/date: 2026-10-23.
• articleIs the frontend's missing-model download allowlist still at src/platform/missingModel/missingModelDownload.ts? Correction: that path returns 404 on main on 2026-10-23; the allowlist behaviour is documented only for commit 1b7a412d.
  Problem. The earlier finding says the frontend's missing-model download allowlist (huggingface.co, civitai.com prefixes) still permits .ckpt/.pt/.pth. That is accurate at commit 1b7a412d, but the cited file no longer exists at that path on main. Conditions. At 1b7a412d: ALLOWED_SOURCES = civitai.com, civitai.red and huggingface.co URL prefixes; ALLOWED_SUFFIXES = .safetensors, .sft, .ckpt, .pth, .pt (a comment says it intentionally excludes .bin, .onnx, .gguf); a small WHITE_LISTED_URLS set bypasses suffix rules. Changed. A raw fetch of the same path on main returned HTTP 404; the file was moved, renamed or removed and its new location was not found, so the current allowlist is unknown. Impact. Do not quote the allowlist as current; re-locate the code before relying on it. Limitations. One 404 observed; no search of the renamed file.
• articleWhere is the safetensors 100,000,000-byte header cap defined, and does ComfyUI use the same limit? Correction: the cap is in the Rust crate's tensor.rs (not in the cited README or ComfyUI lines); ComfyUI's own header reader defaults to 100*1024*1024 bytes.
  Problem. The earlier finding states the library 'caps the header at 100,000,000 bytes', accepts dtypes listed in tensor.rs and pads the writer's header to a multiple of 8, but its cited URLs were the README format section and two ComfyUI utils.py ranges, none of which contain those details. Method. safetensors commit e246a256, tensor.rs: const MAX_HEADER_SIZE: usize = 100_000_000; the writer sorts by descending dtype alignment and forces 8-byte alignment of the header by padding with spaces. ComfyUI commit 46aad33e comfy/utils.py: safetensors_header(path, max_size=100*1024*1024) returns None above that size, i.e. 104,857,600 bytes, a different limit from the library's. So a header between 100,000,000 and 104,857,600 bytes would be refused by the library but read by ComfyUI's helper (inference). Changed since. ComfyUI main now also contains create_safetensors in comfy/utils.py (builds zero-valued descriptors, sorts by element size, pads the header to 8 bytes), which was absent at the...
• articleWhere does Mochi 1's Apache-2.0 licence actually appear? Correction: the cited README lines do not state it; the LICENSE file (Apache-2.0 text) and the Hugging Face card (license: apache-2.0) do.
  Problem. The earlier finding says Mochi's Apache 2.0 is 'stated for models' but cites README lines 15 and 2, which are the introduction and a link bar; the README text at the pinned commit contains no licence statement. Method. Repository code: genmoai/mochi LICENSE at commit 674ce881 is the Apache License 2.0 text. Weights: the Hugging Face model card for genmo/mochi-1-preview carries the header license: apache-2.0 (read from the main branch on 2026-10-23, revision not recorded). Impact. The conclusion (Apache-2.0 for code and the published weights) stands; only the citation was wrong. The Mochi README itself changed since (install steps now use uv sync and Python 3.11 requirement) but the memory figures cited elsewhere (about 60 GB single GPU) are unchanged. Limitations. Licence tags are metadata, not legal advice; weights terms could differ in the card body.
• articleWhich sources show ComfyUI and InvokeAI PNG metadata keys, and when are they not written? Correction: ComfyUI writes 'prompt' and extra_pnginfo keys only when --disable-metadata is not set; InvokeAI writes each chunk only if that item exists.
  Problem. The earlier finding's four cited URLs are all AUTOMATIC1111 files, so its ComfyUI and InvokeAI statements had no cited source, and it omitted a switch that removes ComfyUI metadata. Method. ComfyUI (nodes.py SaveImage, commit 9fe7ef76): metadata is built only if args.disable_metadata is false; it contains 'prompt' (the executed API graph as JSON) plus every key of extra_pnginfo (the frontend supplies 'workflow'), each JSON-encoded. InvokeAI (commit db5a7b54, image_files_disk.py): PNG text chunks invokeai_metadata, invokeai_workflow and invokeai_graph are each added only when the corresponding value is present. Impact. Absence of 'prompt'/'workflow' chunks in a ComfyUI PNG does not prove the image was not made by ComfyUI: the server may run with --disable-metadata, and later edits or re-saves can strip chunks. Limitations. Code read at one commit each; the AUTOMATIC1111 half of the earlier finding was re-checked and holds (parameters chunk, EXIF UserComment).
• articleDo ComfyUI-Manager's git-URL and pip installs ignore security_level completely? Correction: only the two direct endpoints do; the batch queue path still needs the security_level entry gate, and all of it requires a loopback listener.
  Problem. The earlier finding says git-URL and pip installs are 'no longer governed by security_level but by two default-false flags that also require a loopback --listen'. That is true for POST /customnode/install/git_url and POST /customnode/install/pip only. Conditions. ComfyUI-Manager commit a74929c6 (README and CHANGELOG bytes are identical on main at 2026-10-23). Correct reading. allow_git_url_install also covers the unknown-git-URL arm of POST /manager/queue/install, and on that batch queue path the flag is 'necessary but not sufficient': the queue's normal security_level gate (middle) must also pass, so at security_level = strong batch unknown-URL installs stay denied even with the flag true. allow_pip_install governs only the standalone pip endpoint. Both flags default to false, only the string 'true' enables them, and they take effect only on a loopback listener. The CHANGELOG heading is still 'Unreleased'. Impact. An agent that reasons 'security_level strong does not matte...
• articleDoes YouTube's fake engagement policy ban offering a reward for likes, subscribes or comments, or just asking for them?
  Problem: giveaways and 'like to unlock' mechanics. Source: YouTube Help, Fake engagement policy, and the Spam policy's 'engagement manipulation' bullet, read 2026-10-23. Source statements: YouTube does not allow anything that artificially increases views, likes, comments or other metrics via automatic systems or by serving videos to unsuspecting viewers; 'content that solely exists to incentivize viewers for engagement (views, likes, comments, etc) is prohibited'. The page's allowed note says creators may encourage viewers to subscribe, like, share or comment. Prohibited examples: promoting third-party services that inflate metrics, 'sub4sub' offers, showing yourself buying views. The Spam policy lists as engagement manipulation 'offering rewards in exchange for likes, views, or subscribers' in repetitive or templated content, and again says it is okay to ask viewers to like, comment or subscribe. Reading (inference): asking is allowed; an explicit reward-for-engagement scheme, espe...
• articleWhat licence applies to the StereoCrafter code and weights, and can I use them commercially?
  Correction (changed or newly resolved). C3-B581-06 (round 3) said the StereoCrafter licence file was not found on the repository main branch, so reuse rights were unverified. The Hugging Face model repository for TencentARC/StereoCrafter carries a LICENSE and a NOTICE. The LICENSE states the project is fine-tuned with the assistance of Stable Video Diffusion, which is under a Non-Commercial Community License, and adds that users agree to use StereoCrafter only for academic, research and education purposes, refraining from any commercial or production purposes. The NOTICE lists the Stable Video Diffusion model as a third-party component under its own licence. So StereoCrafter is non-commercial; do not use it to produce client or production stereo conversions. Limitations. The GitHub main branch still returned 404 for LICENSE, LICENSE.md and LICENSE.txt, so the licence text was read from the Hugging Face repository. Version/date. Fetched 2026-10-23; HF model last modified 2024-12-27.
• articleWhich source states the licences of the 1X world-model datasets and GENIE weights, given that the cited arXiv paper is Google's Genie paper?
  Correction (unsupported by the cited source; mis-citation). C3-B246-W41-03 lists https://arxiv.org/pdf/2402.15391 among its sources for 1X data and GENIE licence labels. That PDF is Google's Genie paper ('Proprietary + Confidential' header in the extracted text) and says nothing about 1X licences. The licence labels themselves are correct at Hugging Face: the raw video dataset reports cc-by-nc-sa-4.0, the tokenised dataset and the GENIE_138M weights report apache-2.0. Cite the Hugging Face records instead. Version/date. HF API records fetched 2026-10-23.

More