Evaluation and media provenance

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

• articleWhat does the Tencent Hunyuan Community License (HunyuanVideo) restrict by territory, user count, derivative models and outputs, and what notices must accompany redistribution?
  Problem. HunyuanVideo (and models built on it) is common in ComfyUI video stacks, but its licence differs sharply from Apache. Conditions. tencent/HunyuanVideo on Hugging Face, revision 6204ad6a (modified 2025-03-06), files LICENSE and Notice; licence card name tencent-hunyuan-community; release date of the model licence Dec 3, 2024; GitHub licence metadata for Tencent-Hunyuan/HunyuanVideo reads NOASSERTION. Method. Territory: the agreement states it does not apply in the European Union, United Kingdom and South Korea, and that use of the Works or any Output outside the Territory is unlicensed (section 5c). Scale: if the licensee's products exceeded 100 million monthly active users in the month before the version release date, a separate licence must be requested and no rights apply until granted (section 4). Output and derivatives: you must not use the Works or any Output or results to improve any other AI model (section 5b); Model Derivatives include distillation and training on s...
• articleWhat licence covers Wan 2.1, Wan 2.2 and Wan2.2-Animate weights and generated outputs, and what duties does the model card add?
  Problem. Deciding whether Wan weights used in ComfyUI workflows can be used and redistributed commercially. Conditions. Hugging Face repos Wan-AI/Wan2.1-T2V-14B (revision a064a6c7, LICENSE.txt present), Wan2.2-I2V-A14B (revision 206a9ee1) and Wan2.2-Animate-14B (revision cb93a225); the VACE-14B card also tags Apache-2.0 (not read in full). Method. The card front matter says license: apache-2.0 for all three revisions read, and the Wan2.1 repository carries the standard Apache License 2.0 text as LICENSE.txt. Each card's License Agreement section states that the models are licensed under Apache 2.0, that the authors claim no rights over generated contents, and that the user remains accountable and must not share content that violates law, causes harm to individuals or groups, disseminates personal information intended for harm, spreads misinformation or targets vulnerable populations. Practical matrix: code Apache-2.0 (GitHub Wan-Video/Wan2.1 and Wan2.2 report Apache-2.0); weights Ap...
• articleHow should an agent track narrative state across scenes so that scenes can be reordered, resumed and checked, given what storylet and Twine engines record?
  Problem. Agents drift and contradict earlier scenes. Conditions. Multi-scene or interactive scripts written with or for tools like ink, Yarn, SugarCube and Harlowe. Synthesis (inference, built from the cited findings, not a source recommendation). (1) Separate durable facts (what the character knows, relationships, inventory) from scratch values, as SugarCube separates story and temporary variables. (2) Record visit history per scene as an ordered list plus counts (Harlowe history/visits; ink and Yarn counts in C3-B542-01/B543-01), and derive 'has seen' checks from it rather than separate flags. (3) Model each scene as a unit with preconditions, content and effects (storylet anatomy) so availability and consequences are explicit and testable. (4) Keep deterministic state outside the model when conditions are LLM-judged. (5) Use mock visits or fixtures to test branches. Failure modes. Stale flags, expired history, unlogged effects. Limitations. This is a design synthesis; none of the...
• articleUsing c2patool, how do I record provenance for a generated asset, and what happens to the manifest when the file is transcoded or converted?
  Problem. Generated images/videos get re-encoded by editors and platforms; agents must know what the signed manifest does and does not survive. Conditions. c2patool commit c367d9c2 (2026-10-23, Apache-2.0, a separate repository from c2pa-rs since release 0.27.15) and c2pa-rs commit 8250bd96 docs. Method. c2patool -m manifest.json -o out.ext signs a manifest into the asset; the output extension must match the source and the tool will not convert file types. Intent flags: --create <IPTC digital source type> (for AI output, trainedAlgorithmicMedia is one listed value; it adds c2pa.created and cannot be combined with --update or --parent); default Edit intent adds the source as parent ingredient with c2pa.opened; --update is for non-editorial changes such as re-encoding or format conversion to an already-signed asset, which must already carry a manifest. --parent supplies the pre-edit file or an ingredient folder made with --ingredient. --sidecar writes a .c2pa file beside the output, --...
• articleWhen I add LPIPS or CLIPScore to a generation regression test, which input conventions, versions and meaning limits must I respect?
  Problem. SSIM is brittle to harmless shifts; teams add learned metrics but misuse them. Conditions. LPIPS reference implementation (PerceptualSimilarity repo, commit 082bb24f, LPIPS version 0.1) and the CLIPScore paper (arXiv 2104.08718v3). Method. LPIPS: lpips.LPIPS(net='alex') is the default and best as a forward metric; net='vgg' is closer to a traditional perceptual loss for optimisation; inputs must be RGB tensors Nx3xHxW scaled to [-1, +1]; output is a distance (lower means more similar). The README asks publications to state the version used. CLIPScore is defined as CLIP-S(c, v) = w * max(cos(c, v), 0) with w = 2.5, reference-free, between a candidate caption and image embedding; the authors computed it with a ViT-B/32 backbone. Practical regression design (inference): record metric name, backbone, version and input scaling in the test config; use LPIPS for image-to-image change (baseline vs new render) with a tolerance calibrated on your own repeat-run noise, and treat CLIPS...
• articleHow does ComfyUI's own repository run image-quality regression tests (baseline directory, SSIM threshold, matching by prompt metadata), and what can I copy or must I fix for my own workflows?
  Problem. Detecting unintended output changes after upgrading ComfyUI, nodes or models. Conditions. ComfyUI v0.42.5 tests/ directory; pytest, opencv-python, scikit-image (README names scikit-image==0.21.0). Method. Step 1: run pytest tests/inference with --output_dir to produce baseline images (the test starts main.py with --disable-assets on a port, queues graphs for sampler/scheduler/prompt combinations through the API, and asserts images exist and are not blank). Step 2: change code. Step 3: run inference again plus tests/compare/test_quality.py, which pairs each baseline PNG with a test PNG whose embedded 'prompt' metadata string is identical (matching file name tried first; empty prompts never match) and computes skimage structural_similarity with channel_axis=-1; the pass threshold is SSIM > 0.95; it writes PASS/FAIL rows, a difference image and a grid image. Failure modes. Metadata match requires identical prompt JSON, so any workflow edit orphans its baseline; only SSIM is im...
• articleWhich older ComfyUI motion-control node packs (DragNUWA, InversedNoise) are stale or risky to install, and what do their dependency pins and weight downloads imply?
  Problem. Tutorials still point to old trajectory-drag and unsampling nodes; an agent must judge installability and risk. Conditions. chaojie/ComfyUI-DragNUWA (commit bd9f53fe, last commit 2024-01-29, MIT) and logtd/ComfyUI-InversedNoise (commit acd77730, last commit 2024-03-31; no LICENSE file found at that commit). Method. DragNUWA's requirements pin old packages (for example kornia==0.6.9, opencv-python==4.6.0.66, pytorch-lightning==2.1.0, black==23.7.0) and install clip from a git URL, so installing into a modern ComfyUI environment can downgrade shared packages; weights are a .pth from Google Drive (a pickle-family checkpoint) or a pruned fp16 safetensors from a third-party Hugging Face repo, with Chinese-mirror wget guidance. It provides trajectory drawing, camera/object motion brushes and optical-flow workflows. InversedNoise's README calls itself experimental/WIP for DDIM/sample inversion (unsampling) in image-to-image and offers no further documentation. Failure modes. Depen...
• articleHow does ComfyUI decide whether to load a checkpoint as safetensors or via torch.load, and what does PyTorch's documentation say about weights_only?
  Problem. A file named .safetensors vs .ckpt/.pt/.pth/.bin/.pkl is the only gate between data loading and pickle deserialisation. Conditions. ComfyUI v0.42.5 comfy/utils.py load_torch_file; PyTorch 2.14 docs. Method. load_torch_file dispatches on the lower-cased filename suffix: .safetensors or .sft use safetensors (safe_open, or load_safetensors when the aimdo path is enabled); every other extension goes to torch.load(..., weights_only=True). folder_paths.supported_pt_extensions lists .ckpt, .pt, .pt2, .bin, .pth, .safetensors, .pkl and .sft, so those pickle-family files are listed in the model dropdowns. The torch.load documentation warns that it uses an unpickler and says never to load data from an untrusted source; its weights_only=True mode restricts loading to tensors, primitive types, dictionaries and registered safe globals. Failure modes. Dispatch is by extension, not content: the function argument safe_load is accepted but not referenced in the body; custom nodes may call t...
• articleDoes ComfyUI core have a built-in, signed allowlist/blocklist for custom-node packs, and how does it measure a pack?
  Problem. Custom nodes are Python imported at startup; teams want enforcement stronger than Manager security levels. Conditions. ComfyUI v0.42.5 contains app/governance.py. In this source tree GOVERNANCE_REQUIRED is False, so initialize() returns immediately: the mechanism is inert unless a build sets it (inference: a 'governed build' with a baked public key). Method. When enabled, initialize() verifies an Ed25519-signed envelope (governance/policy.signed.json, schema 1, size caps 1 MiB/512 KiB, domain separator) whose payload lists activeForms (customNode, nodeId, model, partnerNode), customNodeMode allowlist or blocklist, packs {name, digest}, deniedPacks, disabledNodes, disabledPartnerNodes and models; this build enforces customNode, nodeId and partnerNode, and a policy requiring 'model' makes startup exit. pack_digest computes a blake3 over relative path + per-file blake3 of .py/.pyw/.pyd/.so/.dll/.dylib/.js/.mjs files in sorted order; symlinks and any compiled .pyc/__pycache__ i...
• articleWhat does ComfyUI-Frame-Interpolation offer (RIFE, FILM, GMFSS...), what frame-count limits apply, and where does it download model weights from?
  Problem. Raising frame rate of generated clips and knowing what the node pack fetches at run time. Conditions. Fannovel16/ComfyUI-Frame-Interpolation, commit 26545cc2 (pushed 2026-03-29), version 1.0.11, MIT code. Method. VFI nodes (GMFSS Fortuna, IFRNet, IFUnet, M2M, RIFE 4.0-4.9, FILM, Sepconv, AMT, STMFNet, FLAVR, ATM-VFI, MoMo) take an IMAGE batch of at least 2 frames; STMFNet and FLAVR need at least 4 frames (README suggests FILM with multiplier=4 first for short inputs), STMFNet, FLAVR, ATM-VFI and MoMo are 2x only. clear_cache_after_n_frames limits memory. Multiplier values can be scheduled. GPU ops use cupy by default; a taichi backend (set ops_backend in config.yaml) is experimental for non-CUDA devices. Install is via install.py or install.bat (cupy-wheel is installed by the script). Weights: vfi_utils.py tries a list of GitHub release URLs (styler00dollar, the repo's own releases, dajes) and per-file Hugging Face mirror URLs at run time, calling torch.hub download_url_to_...
• articleWhat does a ComfyUI-Manager / comfy-cli snapshot actually record, and what must I add (model hashes, disabled nodes, non-registry nodes) to restore an environment faithfully?
  Problem. Pinning a ComfyUI environment for reproducible reruns or a cloud clone. Conditions. ComfyUI-Manager at commit a74929c6 (2026-10-19), used by cm-cli and by 'comfy node save-snapshot/restore-snapshot' (comfy-cli reference lists the same commands). Method. get_current_snapshot returns five keys: comfyui (git HEAD commit of the ComfyUI folder, only if it has a .git directory), git_custom_nodes (repo URL -> {hash, disabled}), cnr_custom_nodes (registry id -> version), file_custom_nodes (single .py files with disabled flag) and pips (installed pip packages). Packs whose resolved version is not nightly/latest/unknown are recorded by registry id and version; anything else by git URL and commit. Save with save-snapshot --output file.json or .yaml (yaml nests the data under custom_nodes); restore-snapshot looks in ComfyUI-Manager/snapshots if the path is not an existing file and has pip flags --pip-non-url, --pip-non-local-url, --pip-local-url and --restore-to. restore-dependencies r...
• articleIn an EEG experiment, how did listeners respond to an AI-voiced versus human-voiced newscast, and does it show lower credibility?
  Corrected 2026-10-23 - see C3-B715-02; the original statements below are kept for history and are superseded where the correction says so. Problem. Whether synthetic voice reduces the credibility of narrated information. Conditions. 30 participants (16 female, 14 male, mostly 18 to 40), Chinese-language news items read by an AI voice or a human reporter, with emotional or neutral content; EEG and ERP measures. Measured result. Brain activity responses were greater for the human newscast than for the AI newscast, with differences in beta bands in the left posterior temporal lobe and a content-valence by voice-agent interaction. Source statement vs inference. The authors interpret the larger effect for human voice as possibly reflecting trust in its credibility and fluency; no behavioural credibility rating is the basis of that statement. Limitations. Authors state the sample is small and age and gender were not strictly controlled; one language and one TTS system; EEG differences do ...
• articleHow does comfy-cli / cm-cli work out which custom nodes a workflow needs (deps-in-workflow, install-deps), and what can it get wrong?
  Problem. Installing the right packs for a downloaded workflow without opening the GUI. Conditions. ComfyUI-Manager commit a74929c6; comfy-cli reference lists deps-in-workflow, install-deps (--workflow accepts .json or .png, --deps a dependency spec, --mode remote|local|cache). Method. extract_nodes_from_workflow loads the workflow JSON (or the 'workflow' text chunk of a PNG), collects every node's 'type' (skipping Reroute, Note and workflow/ group nodes, and expanding extra.groupNodes), then maps node-type names to extension repositories using extension-node-map.json from the chosen channel/mode. Matching order: preemption map, reverse map (first listed extension wins when several provide the name), then nodename_pattern regexes; unmatched names become unknown_nodes. Failure modes. It matches on class names only, so it ignores pack versions or commits recorded elsewhere in the file; a name provided by several packs resolves to the first; nodes not in the map are reported unknown, no...
• articleWhat flag set runs ComfyUI as a locked-down headless API server (listen address, TLS, upload size, no custom nodes, offline), and what does the project say about authentication?
  Problem. Exposing a GPU box as an API without enlarging the attack surface. Conditions. ComfyUI v0.42.5 CLI; docs.comfy.org self-hosting and startup-flag pages. Method. Defaults: --listen 127.0.0.1 (bare --listen means 0.0.0.0,::; a comma-separated list is accepted), --port 8188, --max-upload-size 100 MB. Relevant flags in cli_args.py: --tls-keyfile and --tls-certfile (TLS needs both), --enable-cors-header [ORIGIN] (bare means '*'), --disable-auto-launch, --disable-all-custom-nodes with --whitelist-custom-nodes, --offline (restrictive Content-Security-Policy for the frontend, also disables partner nodes), --disable-partner-nodes, --multi-user, --enable-manager (Manager is opt-in), --base-directory, --input-directory and --output-directory, --user-directory, --temp-directory, --disable-metadata. The self-hosting page states the server has no authentication by default and says to bind non-loopback addresses only on trusted networks or behind an authenticating reverse proxy; community ...
• articleWhat does the official workflow JSON schema say about embedding model references and hashes, and when does a saved PNG carry the workflow versus only the API prompt?
  Problem. Recording enough in a file to rerun a ComfyUI result. Conditions. Workflow schema v1 (latest) as published on docs.comfy.org; ComfyUI v0.42.5 SaveImage. Method. The published schema's ComfyWorkflow1_0 requires version, state and nodes, allows extra top-level properties, and defines an optional top-level 'models' array whose entries have name, url and directory as required fields and optional hash and hash_type (a free string, additionalProperties false). Nodes carry mode, inputs, properties (including 'Node name for S&R') and widgets_values (array or object). So an exporter can put model download URLs plus hash/hash_type into the workflow file; nothing in the schema forces a hash. SaveImage always writes the API prompt into PNG text chunk 'prompt' (unless --disable-metadata) and writes the extra_pnginfo entries (normally 'workflow') only if the client sent extra_data.extra_pnginfo. Inference: images produced through a bare API submit contain the API prompt but no UI workflo...
• articleWhat does the Comfy Registry require in pyproject.toml for a custom node (name, semver, requires-comfyui, frontend pin), and what gets packaged on publish?
  Problem. Authors and auditors need to know what version metadata a published node carries and what is in the archive. Conditions. docs.comfy.org registry pages fetched 2026-10-23 (undated). Method. [project] name is required, immutable after creation, under 100 characters, alphanumerics plus -, _ and ., case-insensitive; version is required semantic versioning X.Y.Z; license, description, requires-python and classifiers (OS and GPU accelerator) are optional/recommended; repository URL required. Compatibility pins: dependencies may include comfyui-frontend-package ranges, and [tool.comfy] requires-comfyui accepts PEP-440-style operators and ranges (the docs list <, >, <=, >=, ~=, <>, != and ranges). [tool.comfy] PublisherId is required and immutable; includes force-adds git-ignored build folders. 'comfy node publish' packages every git-tracked file; a .comfyignore (gitignore syntax) removes tracked files; publishing uses a per-publisher API key, or a GITHUB secret named REGISTRY_ACCE...
• articleDoes a more enthusiastic synthetic narrator improve learning from a narrated lesson?
  Problem. Whether adding emotional cues to a TTS narrator helps viewers learn. Conditions. Between-subjects online experiment, n = 244 learners at a large Asian university, listening to Amazon Alexa narrating a multimedia lesson on a denial-of-service attack in a neutral, low-, medium- or high-enthusiasm voice. Measured result. Enthusiastic voices did not raise immediate retention or transfer scores over the neutral voice; medium and high voices were perceived as more enthusiastic; low and high enthusiasm lowered intrinsic and extraneous cognitive-load ratings; the medium voice raised positive emotion from baseline; persona ratings did not improve. Failure modes. Strong prosody can add load for non-native listeners (the authors' argument). Limitations. One 2022 voice product, one topic, university students, immediate tests only, no human-voice control, no words-per-minute figures extracted. Inference (labelled): richer TTS expressiveness is not shown to improve retention, so do not j...
• articleDo listeners rate cloned or generic AI voices as less trustworthy than human voices, and can they tell them apart?
  Problem. Whether an AI narrator voice is judged as real and trustworthy. Conditions. UK listeners rated unfamiliar voices on short stimuli in rating tasks (Experiments 1a and 1b, with generic AI voices and voice clones in separate contexts) and in a combined Experiment 2 (n = 50; 40 human voices, 40 clones, 40 generic AI voices). Measured result. Generic AI voices were labelled human on 41% of trials and voice clones on 58%; AI voices were not hyperreal (generic voices were rated less real and human than human voices; clones about as real as human voices). Both types were rated more dominant; generic AI voices were rated more trustworthy than human voices in Experiment 1a (57.9 vs 53.8, p = .002) while clones were not significantly different (58.9 vs 57.1, p = .133). Failure modes. Realness ratings of human voices depended on the context of other voices heard. Limitations. First-impression traits on single short stimuli, unfamiliar identities, UK listeners; says nothing about longer...
• articleDoes 1.25x to 1.5x playback change quiz scores, satisfaction or how the speaker is judged, in preregistered online experiments?
  Problem. Whether a slightly faster pace changes viewer experience, speaker impressions and learning. Conditions. Three preregistered online experiments: Study 1 a tiramisu tutorial at 1x, 1.25x or 1.5x (final N = 326 after 53% exclusions); Study 2a an economics video at 1x vs 1.25x with a quiz-warning manipulation (N = 313 after 54.4% exclusions); Study 2b the same in undergraduates (N = 246 after 25.2% exclusions). Measured result. Faster playback reliably raised perceived audiovisual distortions and modestly reduced satisfaction; in Study 1 viewers were more satisfied at 1x (M = 4.17) than at 1.25x (M = 3.82, p = 0.011); the speaker was rated more positively at 1x (M = 4.44) than at 1.25x (4.24) or 1.5x (4.20), marginally significant (p = 0.055), and in Study 2b 1x beat 1.25x on speaker rating (p = 0.022). Objective quiz performance did not differ by speed. Failure modes. Very high exclusion rates; self-report scales; one video per content type. Limitations. Speed is a playback mu...
• articleIn a small randomized within-subject test, how much worse is recall of a story heard at 2x speed than at normal speed?
  Problem. Whether double-speed audio costs memory. Conditions. 40 Lebanese university students (aged 18 to 25, convenience sample), each heard two four-sentence stories of 23 seconds, one at normal speed and one at 2x (after a habituation to 2x words), in counterbalanced order, then played Minesweeper for 10 minutes before cued recall. Measured result. Participants recalled about half of the information overall; normal speed gave a 27% improvement in cued recall over time-compressed speech, statistically significant. Students also reported using compressed speech more than average and mostly judged it harmful. Failure modes. Participants admitted rehearsing during the filler game; mood, time of day and location were uncontrolled. Limitations. Very short material, one narrator, one compression ratio (2x), delayed recall only after 10 minutes, no comprehension test, no ideal-pace data; cannot say what a 10 to 20% faster read would do. The authors suggest faster speeds for revision rath...
• articleHow much does speeding up narration or lecture video hurt learning, according to a 2025 meta-analysis, and how reliable is it?
  Problem. Choosing a narration pace (or a playback speed) for instructional or explainer video. Conditions. Evidence is about accelerated instructional video with different teacher appearance, subtitles and subject matter, mostly educational samples in China and the US. Method (what was measured). A meta-analysis of 12 experimental or quasi-experimental papers, converting speed to words per minute and pooling Hedges' g: acceleration raised cognitive load (g = 0.59, nine studies), lowered maintenance test scores (g = -0.41), lowered transfer ('migration') scores (g = -0.50) and lowered learning satisfaction (g = -0.24). Moderators: teacher on screen, subtitles, country and subject changed some effects; the review reports speeds from 1.1x to 2.5x across the pooled studies, and notes the satisfaction drop arose in samples whose original speech was already 183 to 251 wpm. Failure modes. Heterogeneity is large for cognitive load (I2 77.8%) and satisfaction (I2 96.4%); the authors find mis...
• articleCORRECTION to C3-B536-01: is TikTok Symphony's Dreamina Seedance access 'only' unlocked by $5,000 of advertiser spend in 30 days? (Overstated: agencies follow a separate process and an appeal route exists; some listed details are not on the cited pages.)
  Problem. C3-B536-01's title says Seedance access in Symphony 'unlocks only when the linked advertiser ID spent at least $5,000 in the past 30 days (weekly refresh)'. Audit result against the two cited TikTok help pages (re-fetched 2026-10-23). (a) Supported: direct advertisers unlock automatically at $5,000 USD or more in the preceding 30 days; access is reviewed every Monday; the highest-spend linked advertiser ID counts; clips run 4 to 15 seconds depending on mode; daily generation limits scale with ad spend; the tool is described as free. (b) Overstated: 'only'. The same page says 'Advertising agency: access follows a separate process. Contact your TikTok representative to request access', and gives an appeal route for accounts that already meet the spend requirement but cannot generate. (d) Not found on the two cited pages: the Standard, Fast and Mini variant descriptions, the 'Real-human or AI-generated' avatar label, the shoes/hats/sunglasses limitation, and lip sync options. ...
• articleCORRECTION to C3-B241-G73-03: is the 'term and territory defaults disagree' comparison across Wise Music, Atomica and Media Music Now like for like? (No: it mixes a per-quote form for web streams with library standard licences, and one cited page has no term text.)
  Problem. C3-B241-G73-03 says the pages 'disagree on term defaults (no perpetuity vs perpetual vs unlimited)'. Audit result. (a) Supported facts: the Wise Music 2024 web-video form asks for the listed fields, offers one month, one year or other terms up to three years, says it will not grant perpetuity, defaults territory to world unless geo-blocked and asks for 4 to 6 weeks' notice; Atomica says licences state use, territory and term; Media Music Now's standard licence lists any territory and no time limitations. (c) Missing limitations: (1) the Wise form is titled a quote request for audiovisual in-context on-demand streams on websites and social channels, so its no-perpetuity rule is a quote condition for that use, not a general publisher default; (2) the same form states fees start at a minimum of $250 per work, are pro rata and MFN, and are subject to writer or estate approval, which the candidate does not relay and which affects what 'term' means in practice; (3) Atomica says o...
• articleCORRECTION to C3-B241-G73-04: is it true that none of the library pages read says whether temp or preview use is itself licensed? (Overstated: the FirstCom piece and Media Music Now both say something about permission for temp or preview use.)
  Problem. C3-B241-G73-04 concludes: 'None of the four library pages read says whether temp or preview use is itself licensed, or covers festival or pitch screenings.' Audit result. (a) Supported: Media Music Now says temp music can defer full licensing and producers 'come back to us for the fully licensed version'; Marmoset's 2013 post offers lower-resolution MP3 downloads that are not watermarked; FirstCom's Lonnie Sill (IDA, 2014) says an editor temp-scored a whole film from the catalogue before a distribution deal. (b) Overstated: the absence claim is too strong. The FirstCom piece says the editor 'had existing permission to access music from the library website' before temp-scoring, and that a blanket all-media worldwide licence was provided afterwards, which is a statement about the permission basis of temp use under that relationship. Media Music Now calls its downloadable previews 'free' and says temp use can avoid 'full licensing' until the work circulates, which implies a pr...
• articleIs a face-recognition similarity threshold equally reliable for every kind of face (NIST IR 8280 demographic effects)?
  Problem: an identity check that is tuned on one group may over-flag or under-flag others. Conditions: any automated pass/fail on face similarity. Method (NIST's finding): algorithms submitted to NIST's evaluation showed demographic differentials of various magnitudes; false-positive differentials were much larger than false-negative ones; across demographics false-positive rates often varied by factors of 10 to beyond 100, and results depended on the algorithm and on the photo set. Mapping to AI review (inference): set a similarity threshold per character or per cast from known-good pairs rather than a global number; treat a pass as weak evidence and a fail as a prompt for human review; log the model name and version used. Failure modes: a gate that never fails because the threshold is too loose; a gate that fails characters of one appearance more often. Limitations: the report tests 1:1 verification and 1:N on photographs with mugshot, visa and application images, not stylised or g...
• articleWhat checklist makes a hosted image or video render repeatable months later, and when should I give up on re-rendering and keep the originals?
  Problem: hosted models are retired on vendor schedules. Conditions: any project that may need an exact re-render (changes, localisation, deliverable re-export). Method (inference from the provider documents in this round, plus C3-B609-02/06 for open weights): (1) record the exact model id string used, never an alias such as latest or chatgpt-image-latest; (2) record the provider API version (Runway header) and the date; (3) save the prompt, seed if the API exposes one, input images by hash, and the output file by hash; (4) store a poll of the vendor's shutdown signal (OpenAI shutdown_date; Google tables) with each project and re-check on a schedule shorter than the shortest notice (2 weeks for previews and latest-alias swaps); (5) for work that must be re-rendered bit-for-bit, prefer open weights pinned to a commit and run on your own hardware (see HF pinning); (6) for hosted-only models, treat the delivered master and the source frames as the archive and plan edits as new generatio...
• articleWhat do Runway and Replicate version, and does a dated API version header pin the model I used?
  Problem: 'pin the version' means different things per provider. Conditions: Runway API integrations; Replicate model pages. Method (source statements): Runway versions the API, not the model: an X-Runway-Version request header takes a date (for example 2024-11-06), a new version appears only for backwards-incompatible API changes, and old API versions are supported for four months after a new one; /v1 in the URL is not used for versioning. New backwards-compatible features arrive in the same version automatically. Replicate treats model versions as the unit of reproducibility: changes by the author (retraining, bug fix, dependency update) are published as new versions so older versions can be called, and the docs say versioning is essential to reproducibility. Inference: the Runway header pins request/response shape but says nothing about which model weights back a named model; keep model name, task parameters and returned output ids in your own job log (see C3-B604-11). For Replica...
• articleWhat notice period does OpenAI promise before retiring a model, how can I read shutdown dates by API, and what happened to its image and video models (DALL-E, GPT Image, Sora 2)?
  Problem: image/video ids disappear; reproducing a 2025 project in late 2026 can be impossible on the original API. Conditions: OpenAI API users. Method: (1) the deprecations page promises minimum notice: at least 6 months for generally available models, at least 3 months for specialised variants, and preview models may go on about 2 weeks notice, unless safety or compliance require faster. (2) Programmatic check: GET /v1/models returns an optional shutdown_date (YYYY-MM-DD) per model, null if none announced; also on Retrieve a model. (3) Recorded history: DALL-E 2 and 3 shutdown 2026-05-12; the Videos API and sora-2, sora-2-pro and dated snapshots (sora-2-2025-10-06, sora-2-2025-12-08, sora-2-pro-2025-10-06) shutdown 2026-09-24 with no replacement listed; gpt-image-1-mini, gpt-image-1.5 and the alias chatgpt-image-latest shutdown December 1, 2026. Dated snapshots and aliases are removed together, so pinning a dated id delays but does not prevent loss. Failure modes: not every table ...
• articleWhich Google image and video models (Imagen 4, Veo 2/3/3.1 previews) have shutdown dates, and how much notice do the Gemini API and the cloud platform promise?
  Problem: a project made on Imagen 4 or Veo 3.0 may not re-render at all months later. Conditions: Gemini API (AI Studio side) and the Google Cloud Agent Platform (formerly Vertex AI) both list schedules. Method: read both tables. Gemini API: imagen-4.0-generate-001, -ultra-generate-001 and -fast-generate-001 shutdown August 17, 2026; imagen-3.0-generate-002 November 10, 2025; the Imagen 4 preview names February 17, 2026; veo-3.0 and veo-2.0 June 30, 2026; veo-3.1 preview names October 22, 2026 (replacement named: gemini-omni-1.1-flash and, for Veo, the GA models on the cloud platform). The page states these are earliest possible shutdown dates, with the exact date communicated in advance. Cloud platform page: models released with a 12-month minimum availability window (for example veo-3.1-generate-001, retirement November 17, 2026 or later; gemini-3.1-flash-image May 28, 2027 or later; gemini-2.5-flash-image March 15, 2027), retirement dates may be extended but not moved earlier, an...
• articleWhat do Gemini API model-name patterns (stable, preview, latest, experimental) promise, and which should a reproducible image or video project pin?
  Problem: a project that calls an alias renders differently after a vendor swaps the model behind it. Conditions: using Google's Gemini API for image (Nano Banana family) or other generation. Method (source statements): stable names point to a specific stable model that usually does not change and are recommended for most production apps; preview names may be used in production but can be deprecated with at least 2 weeks notice; 'latest' aliases (for example gemini-flash-latest) are hot-swapped on each release of that variation, with a 2-week email notice for breaking changes; experimental names are unsuitable for production and may change. Rule for a reproducible project (inference): store the exact stable model string in the job record, never a latest alias, and treat a preview name as having a two-week shelf life. Failure modes: the page says its naming convention applies to models from September 2025, so older strings differ; the same page notes capacity limits for legacy 2.5 mod...
• articleWhich kinds of discontinuity at a cut make viewers feel a new event has started: space, time or action (Magliano & Zacks 2011)?
  Problem: how far can an AI sequence break spatial or temporal continuity before it reads as a new event? Conditions: narrative film shown continuously. Method (as reported): re-analysis of behavioural segmentation data (Study 1a, 41 participants; 1b, 24 participants) and fMRI data from viewers of The Red Balloon (33 minutes); edits coded as continuous in space/time/action, discontinuous in space or time but continuous in action, or discontinuous in action too. Result: discontinuities in action had the biggest impact on where viewers marked event boundaries; space and time discontinuities had minor effects; edits produced large transient responses in early visual areas, with different patterns for spatial-temporal vs action changes. Mapping to AI shot design (inference): the continuity worth protecting across a cut is the action/goal thread (who is doing what with which object), more than exact location; relocating a character or jumping time while action continues is tolerated, whil...
• articleHow coherent is gaze on professionally cut footage compared with natural video, and what does the upper limit look like (Dorr et al. 2010)?
  Problem: how much agreement between viewers should a designer expect when guiding attention? Conditions: free viewing without a task. Method (as reported in the abstract): eye movements from 54 subjects on 18 high-resolution outdoor videos, variability measured with Normalized Scanpath Saliency extended over time; controls were stop-motion versions of the same content, Hollywood action trailers (probing the upper limit of coherence from deliberate camera work and cuts), and a repetition condition (same movies ten times over two days). Result: up to about 80% of subjects looked at the same region in some segments but variability was usually much greater; gaze on Hollywood trailers was significantly more coherent than on natural movies; static images show onset effects early and idiosyncrasy later. Mapping to AI shot design (inference): synchrony is not a property of 'video' but of composed, cut footage; uncut, wide, AI-generated establishing shots with many equal-salience elements wi...
• articleIf edited film makes viewers look at the same place, how much does individual comprehension still change gaze (Loschky et al. 2015 'tyranny of film')?
  Problem: designers of AI shots assume they can steer attention with framing and motion; does viewer understanding override that? Conditions: rapidly edited, centre-framed Hollywood-style sequence. Method (as reported): a 12-second, six-shot cross-cut James Bond (Moonraker) clip, shown silent at 720x480, 30 fps; one group saw only the clip, the other also saw the preceding 2 min 56 s (Context). Experiment 1 (think-aloud, 98 recruited, 10 discarded as having seen the film), Experiment 2 (event segmentation, 88, 5 discarded), Experiment 3 (eye tracking, EyeLink 2000, 176 participants, 87 per condition, 13 discarded). Results: Context viewers more often drew the predictive inference (91% vs 71% in Exp. 3); gaze showed strong attentional synchrony in both conditions against a chance baseline, with only small differences between them (slightly greater synchrony and lower processing load in the Context group at the first circus-tent shot). The authors read this as more consistent with 'tyr...
• articleWhat does the Smithsonian say about how far to trust its CC0 designation and open-access metadata?
  Corrected 2026-10-23 - see C3-B686-06; the original statements below are kept for history and are superseded where the correction says so. The Smithsonian FAQ says its CC0 designation waives only its own copyright interests, so third-party permissions (e.g. people, trademarks) may still be needed and are the user's responsibility; 'no known copyright restriction' reflects best efforts with available information, and items with limitations can carry CC0 metadata without CC0 media.
• articleWhich ICC profiles are public for testing colour management, and what permission text accompanies each?
  Corrected 2026-10-23 - see C3-B686-04; the original statements below are kept for history and are superseded where the correction says so. The ICC Profile Library says that where ICC owns a profile it may be copied, distributed, embedded, made, used and sold without restriction, but altered versions must drop the original identification and not be misrepresented; the sRGB v4 preference profile page grants use, copying and distribution only if the file is not changed (including its copyright notice). The two texts differ on modification, so one 'ICC profile licence' label would be wrong. Little-CMS's root licence is MIT, with no separate note for its testbed files (producer).
• articlebpy 5.2.2 on PyPI requires Python ==3.13.* and declares licence 'GPL-3.0' with author Blender Foundation; the 5.2 API docs say the module is not on the official download page but a pre-compiled module is available via PIP; blender.org describes Blender as...
  Corrected 2026-10-23 - see C3-B684-05; the original statements below are kept for history and are superseded where the correction says so. bpy 5.2.2 on PyPI requires Python ==3.13.* and declares licence 'GPL-3.0' with author Blender Foundation; the 5.2 API docs say the module is not on the official download page but a pre-compiled module is available via PIP; blender.org describes Blender as GPL-licensed.
• articleThe cuBLAS documentation (fetched 2026-10-21) limits bitwise reproducibility to the same toolkit version, same architecture and same SM count, and excludes multiple active streams and fixed-point emulation; cuBLASXt adds tile-size and CPU-hybrid...
  Corrected 2026-10-23 - see C3-B684-04; the original statements below are kept for history and are superseded where the correction says so. The cuBLAS documentation (fetched 2026-10-21) limits bitwise reproducibility to the same toolkit version, same architecture and same SM count, and excludes multiple active streams and fixed-point emulation; cuBLASXt adds tile-size and CPU-hybrid conditions.
• articlesox_ng (a4a14b4f) is a hard fork of sox-14.4.2 hosted on Codeberg and by default installs as sox_ng/libsox_ng, not replacing 'sox' unless built with --enable-replace, so scripts calling 'sox' may still invoke the...
  Corrected 2026-10-23 - see C3-B691-01; the original statements below are kept for history and are superseded where the correction says so. sox_ng (a4a14b4f) is a hard fork of sox-14.4.2 hosted on Codeberg and by default installs as sox_ng/libsox_ng, not replacing 'sox' unless built with --enable-replace, so scripts calling 'sox' may still invoke the original.
• articleOpenVDB's LICENSE is MPL 2.0 at tag v11.0.0 and Apache 2.0 at tag v12.0.0 (curator GET by tag name), so a premise that OpenVDB is MPL-2.0 is...
  Corrected 2026-10-23 - see C3-B684-03; the original statements below are kept for history and are superseded where the correction says so. OpenVDB's LICENSE is MPL 2.0 at tag v11.0.0 and Apache 2.0 at tag v12.0.0 (curator GET by tag name), so a premise that OpenVDB is MPL-2.0 is version-dependent.
• articleDoes the Smithsonian FAQ summarised in the earlier finding come from the cited URL, and does the FAQ say it gives no accuracy guarantee?
  Problem. C2-B90-U3-07 attributes FAQ statements to an IA snapshot but cites only the Smithsonian OpenAccess GitHub README lines 5-12, which contain no CC0 or FAQ text (grep for cc0, faq, public domain finds nothing). Conditions. Using the Smithsonian CC0 designation as a rights signal. Method. I fetched the IA snapshot of si.edu/openaccess/faq (20250912113456). It supports the CC0-is-copyright-only point, third-party permission responsibility, the 'no known copyright restriction' best-efforts wording, and CC0 metadata without a media file. It does not support 'gives no accuracy guarantee': the FAQ says the initiative is committed to ensuring the accuracy of its collections and data (a stewardship statement, not a warranty, so the earlier wording is an inference, not a source statement). Failure modes. Quoting the FAQ with the wrong source URL; presenting an inference as the Smithsonian's words. Limitations. Terms of Use page, which the FAQ refers to for disclaimers, was not read. Ve...
• articleWhat conditions does the ICC attach to the sRGB v4 Preference profile beyond 'not changed', and how do they differ from the general library licence?
  Problem. C2-B69-W1-04 summarises the sRGB v4 terms as use/copy/distribute provided the file is not changed. Conditions. Redistributing the profile in a tool, dataset or test suite. Method. The ICC page says permission is granted to anyone who acknowledges the file is provided 'AS IS' with no warranty, provided the file is unchanged including the copyright tag, and that ICC's name is not used in advertising or publicity without written permission. The display-class version is stated to have the same terms. The general library statement is different (it permits sale and embedding without restriction; altered versions must not be misrepresented). Failure modes. Treating the sRGB v4 terms as the permissive library licence, or re-saving the profile with a changed header, voids the stated permission. Limitations. This is a restatement of site text, not legal advice. Version and date. registry.color.org pages fetched 2026-10-23.
• articleAre the temp-track, licence-certificate and cue-sheet steps in C3-B241-G73-06's 'documented' checklist actually in its cited pages?
  Correction (c, missing limitation / d) to C3-B241-G73-06. Problem: the checklist labels as 'documented' the steps: keep the licence certificate, treat temp as not-yet-licensed, and deliver cue sheets, performer agreements and executed licences. Re-fetching the three cited pages: no occurrence of 'certificate', 'cue sheet' or 'performer' as deliverables, and 'temp' appears only inside a name. The Guild page is a biography/about page. Documented and checkable: the Wise Music form asks for the portion used, timing, term, territory and budget and says to submit at least 4-6 weeks before posting; Tresona says the price shown is an estimate until publisher approval. Method: relabel the unsupported steps as 'proposal'. Limitations: only three URLs searched; other pages cited by the candidate chain were not re-checked.
• articleWhat does IASA TC-03 (4th edition) specify for analogue-to-digital transfer, replay speed/equalisation, preservation master format and the split between preservation and restoration/access copies?
  Corrected 2026-10-23 - see C3-B645-07; the original statements below are kept for history and are superseded where the correction says so. For preservation transfers IASA-TC 03 (2017) requires replay settings (speed, playback EQ, track format) chosen objectively from knowledge of the original format and documented, forbids subjective 'improvements' such as de-noising in the unmodified transfer, allows playback EQ to be deferred to access copies, and recommends at least 48 kHz/24-bit LPCM, with 96 kHz/24-bit widely used by memory institutions.
• articleWhat does a "Verified" badge on a Hugging Face commit mean for a model repository, and does it cover weight files?
  Corrected 2026-10-23 - see C3-B644-01; the original statements below are kept for history and are superseded where the correction says so. On Hugging Face a Verified commit badge means the commit was GPG-signed with a key uploaded to the account whose email matches the account email; it speaks to commit authorship, not to the integrity or provenance of weight files beyond that commit.
• articleDo the cited IASA page and current iasa-web.org URLs support the TC-03 replay-parameter, deferred-EQ and BWF 48/96 kHz claims?
  Problem. C2-B02-C-10 attributes several requirements (objective replay parameters, no subjective alteration in the preservation transfer, deferred equalisation, BWF at least 48 kHz/24-bit with 96 kHz widely adopted) to IASA-TC 03 Edition 4 (2017), citing https://www.iasa-web.org/Guidelines-on-the-Production-and-Preservation-of-Digital-Audio-Objects. Conditions. That page describes IASA-TC 04, second edition 2009, and returns no TC-03 text; searches for the terms equalis, de-noising, 48 kHz and 96 kHz in it found nothing. Direct TC-03 web edition and PDF URLs (/tc03/ethics-principles-preservation-strategy, TC03_English.pdf, both found via search) returned 404 on 2026-10-23. Method. Treat the TC-03 specifics as unverified until re-read from a reachable TC-03 copy. Failure modes. Misattributing TC-04 (2009) content to TC-03 (2017), or citing numeric formats with no reachable source. Limitations. The claims may well be accurate to the original TC-03 text the earlier worker read; this au...
• articleHow many ISDCF-name claim fields does ClairMeta cross-check against the CPL?
  Problem. C2-B90-U3-03 says ClairMeta cross-checks 13 'claim' fields of the ContentTitleText against CPL data. Conditions. In dcp_check_isdcf_dcnc.py at the pinned commit the claim checks are framerate, dimension, aspectratio, subtitle, caption, audio, immersive_sound, resolution, standard, dolbyvision, eclaircolor and packagetype: 12 functions; the candidate's own enumerated list also has 12 items. Method. Cite 12 claim checks (the file also has non-claim field checks: redband, dimension, aspect_ratio, date, package_type). Failure modes. A reviewer reconciling the count against the code finds a mismatch. Limitations. If the candidate counted a different convention (e.g. including a helper), that was not stated. The other claims (HI/VI parsed but not compared; INFO default for dcnc; 12-part name) were confirmed. Version/date. ClairMeta b3b66a0, fetched 2026-10-23.
• articleWhat would a regression gate for a ComfyUI workflow change look like, and which parts are documented versus my own design?
  Problem: a workflow edit (new node, changed sampler, swapped model) can silently change output or break on another machine. Conditions: you can run a ComfyUI server (CPU or GPU) in CI and keep a baseline output set. Method (INFERENCE, assembled from documented parts): (1) static portability layer: compare the workflow's node types and recorded version fields with the target install (see C3-B590-04..06). (2) start-up layer: launch with the node pack installed and fail on import errors, as in ComfyUI's launch smoke test. (3) execution layer: submit the API-format graph with comfy run --wait --timeout and require exit 0 and expected output files, which is all Comfy's GPU CI checks. (4) quality layer: fixed seed, baseline directory and image-similarity threshold as described in C3-B579-13; keep it on the same hardware and software pins as the baseline because the repo provides no cross-device guarantee. (5) drift layer: scheduled re-check of pins. (6) release layer: publish only after l...
• articleHow can a ComfyUI-related repo detect that an upstream rendering or format assumption drifted, without making every pull request depend on the network?
  Problem: workflows and tooling depend on constants owned by another repository (here ComfyUI_frontend geometry). Conditions: comfy-cli at commit 0b49c62, .github/workflows/litegraph-parity.yml. Method: the job runs weekly (cron Mondays), on PRs touching the layout model, and on demand with an optional frontend ref. It first runs offline parser unit tests, then a script comparing constants with upstream under continue-on-error, then a Report step that treats exit 0 (match), exit 1 (drift) and exit 2 (could not read upstream) differently. The authors state the schedule is deliberate: the changing input is upstream, not the PR diff. Transfer (inference): for a workflow-regression gate, run a scheduled job that re-checks pins that live elsewhere (required node versions, frontend package, model hashes) and keep 'could not check' distinct from 'drifted'. Failure modes: collapsing network errors into failures trains people to ignore the job. Limitations: the specific parity script was not ...
• articleWhat does `comfy node validate` (run before `comfy node publish`) actually check?
  Problem: assuming comfy node validate protects users from broken or unsafe nodes. Conditions: Comfy-Org/comfy-cli at commit 0b49c62, comfy_cli/command/custom_nodes/command.py. Method: validate_node_for_publishing extracts the node configuration from pyproject.toml, exits 1 if configuration cannot be read or project.version is empty, then runs ruff check . -q --select S102,S307,E702 --exit-zero. S102 is exec use, S307 is eval use, E702 is multiple statements on one line. --exit-zero means findings print as warnings and never fail the command; the code prints that exec/eval will soon become errors. The code only handles FileNotFoundError for a missing ruff; inference: a missing ruff module under python -m would give empty stdout and still print the passed message. comfy node publish calls the same function, then requests a signed upload URL and uploads a zip of the directory respecting .gitignore. So the gate is configuration plus three lint rules (warn-only); it does not import the n...
• articleWhat does the Comfy publish-node-action do, and what should I add before letting it publish a custom node?
  Problem: automating Comfy Registry publishing from GitHub. Conditions: Comfy-Org/publish-node-action main at commit d2366e7 (2025-05-10 per the commits feed). Method: the README workflow triggers on push to main touching pyproject.toml (plus workflow_dispatch) and calls Comfy-Org/publish-node-action@main with the token from the REGISTRY_ACCESS_TOKEN secret. The composite action checks out, sets up Python 3.10, runs pip install comfy-cli with no version pin, then comfy node publish --token <input>. The README says comfy-cli publishes the version in pyproject.toml, so the version must be bumped by hand. Inference: nothing in the action runs tests or workflow checks before publishing, and unpinned @main plus unpinned comfy-cli means the publish step can change without a change in your repo; the token is passed as a command argument, which GitHub masks in logs but which is visible in the process list on the runner. Method for a safer gate (inference): add a prior job running your tests ...

More