Evaluation and media provenance

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

• articleWhat language, text-direction and Unicode-mapping information can a PDF carry, and what do PDF/A and PDF/UA validators actually check about it?
  ISO 32000-1 (PDF 1.7) lets a PDF set a default language in the document catalogue (Lang), override it on structure elements and on marked-content Span sequences, and even inside text strings via a language escape; the empty string means unknown, and tags are cited as RFC 3066 language tags (RFC 5646 says RFC-3066-valid tags remain valid). Direction is a separate Tagged PDF layout attribute (WritingMode). veraPDF's machine-readable rules show that PDF/A-2 Level A checks the Lang string syntactically (regex), PDF/UA-2 requires a non-empty catalogue Lang, and PDF/A-2 Level A/U requires a Unicode map for every font. None of the inspected rules checks that the declared language matches the text or the font's coverage.
• articleWhat exactly does Münsterberg claim about parallel action, forward glances/imagination and spectator versus character emotion, what limits does he put on those claims, and which later items in the catalogue touch them?
  For parallel action Münsterberg claims the spectator does not experience intercut scenes as successive ('omnipresence', a chord), yet he himself notes psychologists dispute whether attention divides, and warns that too many changes (170 in ten minutes in one Carmen) produce 'nervous jerking'. For forward glances he claims film can act as imagination acts, and for emotion he separates feelings that mirror the character's from feelings that differ (fear for the unaware child), adding that makers had hardly used the second kind. Later items located touch these only at the edges: cross-cut shots marked as new events in one clip, a suspense model that separates felt from perceived tension, and Kuleshov-type context effects on rated valence. No located item tests flash-forward reception. All are 1916 theory versus second-hand adjacent records.
• articleHas later empirical work in the catalogue or cycle-3 batches tested Münsterberg's claims that the close-up depicts the viewer's attention and the cut-back the viewer's memory?
  No catalogue or cycle-3 item located by c3dup tests the two mappings directly. Near neighbours measure different endpoints: edit detection (the JEMR 2008 edit-blindness study, 7 participants), event-boundary marking (Zacks 2010; Magliano and Zacks 2011), and comprehension/load of flashbacks versus chronological order (Sevenants and d'Ydewalle 2011, n 30 and 8). Smith's attentional theory of continuity shares Münsterberg's attention emphasis but is itself a theory with cited/illustrative gaze evidence. Münsterberg's mappings therefore remain 1916 theory; related findings are adjacent, small and mostly single-film. All cross-references are second-hand records, not re-read papers.
• articleWhat information would a reader need before treating Wayve's GAIA-3 'predicts relative policy performance' and GAIA-4 'outcome fidelity' statements as a validated correlation with real-world driving, and which of it is missing?
  SYNTHESIS from the pages read. GAIA-3 and GAIA-4 describe validation designs (GAIA-3: simulated interventions compared with on-road experiments plus LiDAR alignment; GAIA-4: outcome, closed-loop and component fidelity on a world-on-rails replay of logged scenes) but give no population, statistic or uncertainty. A reader needs at least: the policies compared, the scenarios and their selection, the real-world metric used, the agreement statistic with intervals, who ran it, and whether the model was trained on the logged scenes it is tested on. None of these is in the page text; the only published GAIA evaluation protocol (GAIA-2 paper) tests visual and conditioning fidelity, not policy-ranking agreement. GAIA-4 states its replay keeps other road users on logged paths, which limits what an outcome agreement can show.
• articleDoes Wayve publish a technical report or paper for GAIA-3 or GAIA-4, and which GAIA versions do have one?
  In the sources inspected, Wayve's GAIA hub page lists two downloadable papers (GAIA-1, 29 Sep 2023; GAIA-2, 26 Mar 2025) and none for GAIA-3 or GAIA-4. The GAIA-3 (2 Dec 2025) and GAIA-4 (3 Aug 2026) blog pages contain no arXiv, PDF, DOI, GitHub or Hugging Face links, and an arXiv title-field search at 10:21Z found GAIA-1 and GAIA-2 but nothing titled GAIA-3 or GAIA-4. So the only evidence for GAIA-3/4 an agent can read is the company page text; the GAIA-2 technical report is the latest GAIA paper found. This is a bounded negative result, not proof that no report exists anywhere.
• articleWhat pre-air checklist for scorebugs, lower thirds and data graphics can be assembled from the broadcaster, regulator and sports-body texts, and how strong is each check's evidence?
  Synthesis only. Seven checks can be traced to the inspected texts: (1) list every element where colour carries meaning (BBC mobile audit step); (2) add a non-colour cue such as label, logo, tick/cross (BBC Trust remedies; fans' proposal); (3) run the palette through a CVD simulator (named by the Trust as used in 2015; model unspecified); (4) test text/background contrast (WCAG AA on web and app; TV only "good contrast"); (5) greyscale or monochrome screen (BBC broadcast guidance; UEFA, World Rugby for kits); (6) check the rendered output on real or virtual lighting and small screens (Trust; World Rugby); (7) do not rely on narration. No inspected source gives a numeric pass/fail threshold for TV graphics, a validated order of the checks, or a test of any scorebug, so this list is a proposal with an experiment sketched for later.
• articleWhat open-access empirical evidence exists on how people judge a narrative's plausibility, and what does it imply for checking an outline?
  The one openly readable empirical study located here that manipulates narrative structure and measures perceived plausibility is Canter, Grieve, Nicol and Benneworth (2003, University of Huddersfield repository copy). It used invented suspect testimonies (a possible homicide and a possible burglary), reordered clauses relative to a Stein-and-Glenn canonical sequence, and inserted 'criminal anchors' (stereotype-activating details), then asked student/community raters for plausibility on a ten-item scale. Reported findings: the anchors lowered perceived plausibility; sequence effects depended on the scenario in one study and were significant in a second, with sequence and anchoring contributing about equally to variance in the last study. The study is about legal testimonies, not fiction, and does not manipulate coincidence; participant counts are inconsistent inside the paper. I found no openly readable experiment that manipulates coincidence in fiction (bounded negative).
• articleDo other baselines in these tables drift like DriveDreamer (version, rounding, row choice, quoted versus recomputed), and do the citing papers' own results change between arXiv versions?
  Yes for some, not all. DriveGAN has the same v1/v2 split as DriveDreamer (inside DriveDreamer's own paper) and later tables read use only the v1 pair. MagicDrive appears with at least five different values in the tables read (FID 16.20 and 16.59 in its own table; 19.1/218.1 in UniMLVG, a pair not found in MagicDrive v1/v7 or MagicDrive-V2 v4; FVD 218.12 in MagicDrive-V2; FVD 221.90 in DiVE; FID 46.18 recomputed by Delphi). Drive-WM is printed at FID 15.8 in its own text, DriveDreamer-2 and Delphi, but 15.2 in UniMLVG v1-v3 (cause not stated). Panacea 16.96/139 appears as 16.9/139.0 in DriveDreamer-2 and 16.96/139.0 in Delphi; DrivingDiffusion 15.83/332 appears rounded to 15.8 in DriveDreamer-2 and GenAD. Version drift of a citing paper's own result occurs in UniMLVG (8.8/60.1 in v1 and v2, 5.8/36.1 in v3, baselines unchanged); no change was found between the versions read for GenAD, DriveDreamer-2, Delphi (v1 vs v3), MagicDrive-V2 (v1 vs v4) or MagicDrive (v1 vs v7 for its MagicDriv...
• articleWhich objective metrics (PESQ, STOI, DNSMOS, SI-SDR) and listening tests do the DeepFilterNet papers and the metric papers report, what do those numbers measure, and how far can they stand in for post-production acceptance of cleaned dialogue?
  The DeepFilterNet papers report PESQ, CSIG, CBAK, COVL and STOI on Voicebank+Demand and, for DFN2, DNSMOS P.835 on the DNS4 blind set; neither the DFN2 nor the DFN3 paper (v1 HTML read) reports a listening test. DNSMOS P.835 is a non-intrusive predictor trained on crowd P.835 ratings of suppressor output, accurate per model (PCC 0.94-0.98) but weaker per clip (0.71-0.83). Intrusive metrics need a clean reference, which production dialogue lacks; SI-SDR ignores gain error. In URGENT 2024 two systems topped DNSMOS/NISQA yet scored low on MOS. No source read links any of these metrics to a post-production acceptance threshold.
• articleDoes ElevenLabs publish an Enterprise Music term sheet or Enterprise-specific music media-rights text beyond Music Terms Exhibit A (updated 2026-10-09)?
  No separate public Enterprise Music term sheet was found. The Enterprise-specific music text that is public is (a) two Enterprise columns inside Exhibit A of the Music Terms (Enterprise Music Lite and Enterprise Music, each "Contact Us for Details"), where Lite carries the same film/TV/radio/Studio Games exclusion as the self-serve plans and Enterprise Music lists all online and offline commercial use, and (b) short summaries on the /music page, the API page and in the docs (finetune screening, marketplace custom licensing). Limits, resale and library terms for Enterprise are "Custom", so the governing text is an order form that is not public. The v1 Terms URL linked from the API page returned 404; an archived copy (Nov 2025) gives "Contact us for details" for every Enterprise row.
• articleWho keeps the consent evidence and the cloned-voice data in ElevenLabs' and Microsoft's texts, and what can the speaker or customer delete?
  ElevenLabs' Creative Terms put consent records on the customer (five years or the legal period, shown on request); the Terms let the user request deletion of User Voice Models and opt out of training use of Content, with effect only going forward. Microsoft's Product Terms let Microsoft retain a copy of each Synthetic Voice and acknowledgement audio under a customer-granted licence; Microsoft's talent disclosure says it keeps the recorded statement as long as needed for security and integrity and may keep a copy of custom neural voice models; the Code of Conduct requires apps serving external users to let them remove their voice. Texts compared at dated versions; a synthesis of vendor statements, not a legal analysis.
• articleBefore an agent uses a free 3D, texture, HDRI or sound asset in a production or dataset, what per-asset licence evidence can it record, which libraries expose a machine-readable licence field, and which must be captured from pages?
  A synthesis proposal built from the observed library structures. Sketchfab exposes a licence object per model; the Poly Haven and ambientCG API responses inspected carry no licence field, so their licence evidence is a site-wide page; Freesound extends each sound's licence to its metadata and keeps uploader AI preferences per user; Sonniss and Sketchfab say the version in force at download governs; BBC limits uses by purpose. The record therefore needs the asset id, licence label and URL, the licence and site-terms texts with hashes and dates, the version rule, AI-related flags, and per-component status, all with access time. This is a proposal, not a tested procedure and not a legal conclusion.
• articleWhat do openly readable empirical studies report about how readers or models track a character's changing traits through a story, and what remains unmeasured?
  No open-access experiment that directly manipulates character change and measures reader perception was found (bounded to the searches listed). What open papers do report: PersoNet (ACL 2023) builds a dataset of readers' notes about characters' traits at a reading position and finds that humans who know the book history reach more than 70% accuracy versus about 45% for the best model, and that oracle history traits are limited because characters' personalities evolve; BookWorm (EMNLP Findings 2024) finds models handle static aspects (role, personality) better than dynamic ones (events, relationships); TimeChara (2024) shows role-playing LLMs break point-in-time consistency. These are NLP benchmark results on proxies, not direct tests of perceived transformation.
• articleWhat fields should an agent record as per-file licence evidence for a test asset so that a reviewer can re-check it later?
  A proposed record, derived from the three inspected source types and labelled synthesis: asset identifier and source URL; the statement locator and its class (file tag, folder notice, collection page, API field); statement revision or Last-Modified and sha256 of the statement bytes; the asset's own hash if the source reports one (Commons sha1) without downloading; the exact licence text or tag copied (excerpt under 25 words); author and credit line; scope caveats (collection vs file, 'some may impose additional restrictions', research-only conditions); and an explicit 'not stated' for each unaddressed right (hosting, redistribution, modification, model_training, commercial use). The Xiph and Commons examples show why each field is needed.
• articleHow could an agent numerically check a graded frame's levels against EBU R 103 and BT.709/BT.2100 without a hardware scope?
  A proposed procedure assembled from the standards: decode to the stated bit depth and range, drop pixels outside the active picture, low-pass filter with R 103's separable filters (horizontal 1,2,3,4,3,2,1 over 16; vertical 1,2,1 over 4), compute the share of R,G,B,Y samples outside nominal, preferred and total bands, and flag out-of-gamut only when more than 1% of the image exceeds the preferred band. Colour balance and hue checks follow from per-channel percentiles and Cb/Cr angles. The method is the worker's synthesis; a small stdlib script reproduces the band arithmetic on synthetic data, and no real frame, tool or scope output was run or compared.
• articleDoes the official page of each dataset (DIV2K, Flickr2K, OST, Vimeo-90K, REDS) state any licence or terms of use at all?
  Two of five datasets state terms on the pages read: DIV2K says it is made available for academic research purpose only (copyright with the original owners), and REDS says CC BY 4.0. The Vimeo-90K project page and repository README, the Flickr2K section of the EDSR/NTIRE2017 README, and the OST project page and README state origin and download links but no terms. Absence is bounded to those pages: the Vimeo-90K dataset subpage returned 403, and no separate terms page was found for Flickr2K or OST.
• articleWhat do public funder/distributor delivery documents say about the E&O certificate and chain-of-title papers that must be delivered?
  The one public delivery document inspected that states E&O requirements is a New Zealand Film Commission deliverables list (public funder, not a streamer or broadcaster): an insurance certificate and on request policy terms showing E&O matches the financing agreement term, coverage and deductible; the funder recorded as additional insured; no change without 30 days prior notice; plus legal opinion on chain of title and licences/releases/contracts, and music cue sheet. For named streamers and broadcasters (Netflix, BBC, PBS) no public E&O delivery text was found; limits quoted for them in secondary pages were not used. The lawyers' 2017 text describes a typical distribution agreement as $1M/$3M, deductible up to $25,000, three years.
• articleWhat exactly did the WHAM paper measure for persistency, with what values, sample sizes, labelers and limits?
  Nature Table 1 (observed 2026-10-22) gives the share of 1.6B-WHAM generations labelled as persisting an edited-in element: with one conditioning frame, Powercell 58%, character 45%, Vertical Jumppad 58%; with five frames, 86%, 85%, 98%. The protocol is small and narrow: 600 videos (480 for Powercell/character, 120 for Jumppad), no-op actions, 10 sampled videos per edited input, each judged only over the first ten generated frames (1 s) by seven of the authors, blinded to input length, with 90% agreement on the 300 double-labelled videos (stricter label kept). The 10-frame condition appears only as an Extended Data figure (values not extracted), and the text says persistency rose from 1 to 5 frames but not from 5 to 10.
• articleHow robust are open-access findings about readers tracking goals, and what can an agent check in a draft to make a character's goals inferable and their shifts visible?
  Open-access evidence available here is thin. A 2004 recognition-probe experiment (84 retained readers) shows goal restatement and replacement change goal accessibility; a 2019 CC BY picture-story study shows explicit goal shifts raise the odds of perceiving an event boundary (odds 3.53 times) and lengthen viewing time; a 2014 CC BY review confirms the Trabasso-style superordinate/subordinate goal hierarchy but says few studies vary goal structures. The core papers named in the question (Suh and Trabasso 1993; Lutz and Radvansky 1997; Magliano and Radvansky 2001) could not be read openly (paywall or 403), so robustness cannot be judged from replications. A draft check can be proposed: list goals, tag each as stated, inferable from action, or absent, and mark every goal shift and restatement; this check is the worker's synthesis.
• articleWhich fear measures do the open horror studies use (self-report, continuous ratings, skin conductance, startle EMG, fMRI) and what can each tell an agent about a script or cut?
  Inspected studies use: continuous per-moment fear ratings with fMRI and between-viewer synchrony (Hudson, N = 37 plus replication); skin conductance, ECG and nine-point retrospective ratings (Chun, N = 50); continuous expectation ratings plus retrospective startle/distress/tension/enjoyment ratings (Acosta and Sears, unreviewed); and focus-group discussion (Kershaw). The jump-scare paper says classic startle measures (heart rate, skin conductivity, facial EMG) have used probe tones, not film stingers. For an agent without participants, only the self-report scale design is reusable.
• articleHow can an agent grade a horror/thriller script or cut for fear-building and say which claims behind each check are measured versus craft lore?
  A usable audit has six checks, each tagged with an evidence grade from the inspected sources: anticipation phase present before a scare (measured fMRI and self-report, partly second-hand); obscurity without confusion (craft lore, plus indirect uncertainty experiment); sound cue use and whether it is a learned genre cue (qualitative); repeated-cue risk (measured for re-viewing, not for within-film repetition); darkness (lore); jump-scare physiology (not yet measured). This is a synthesis; no inspected source validates the audit itself.
• articleWhat is the difference between a jump scare and slow dread in horror, and what has actually been measured about each?
  A jump scare is a sudden image plus a loud sound (a startle); dread is sustained anticipation before a threat arrives. Open sources support a two-part structure: a measurable anticipation phase that raises later reaction, and an acute reaction at the scare. The strongest measured evidence is one fMRI study of feature-length horror films with self-rated fear; the jump-scare-specific physiology is described by its own authors as not yet studied.
• articleWhen I need to clone or convert a real person's voice with a hosted provider, what consent proof does each provider ask for, in the first place?
  As observed 2026-10-22 the providers read ask for three different things. Microsoft (personal and professional voice), Google (Chirp 3 Instant Custom Voice) and OpenAI (recording-based custom voices) each require a separate recorded consent statement, matched to the speaker of the sample, and gate access to approved customers. Microsoft and Google prescribe the wording; OpenAI has the caller read one of its listed phrases. ElevenLabs asks for a confirmation (instant clone) or a recorded verification of your own voice (professional clone). Resemble's terms say consent may be required and must be verbal if so, but its voice-cloning pages are now archived historical documentation with no consent step.
• articleGiven what the open evaluation tools measure, how should an agent evaluate a TTS or voice-conversion take for narration or dubbing?
  This candidate is a synthesis (a proposal, not an executed result) built from candidates 01 and 02. No source read defines a dubbing or narration acceptance test. The sources support a layered check: an ASR round-trip for intelligibility, a speaker-embedding cosine against the actual reference audio for voice identity, and a small blind listening comparison for naturalness, expressiveness and timing, with counts and the comparator reported. Timing, lip-sync, emotion and artefacts are not covered by any tool read here and need listening.
• articleHow do two open zero-shot TTS papers run their listening tests (CMOS and SMOS), how many raters do they say they used, and what is missing for judging whether a score difference is real?
  Seed-TTS (arXiv v1) uses a comparative MOS where raters hear a reference clip then the system and a human recording in random order, scores -2 to +2, and calls differences under 0.1 insignificant, but states no rater count or interval for it and does not release the subjective set. F5-TTS (arXiv v3) states 20 native raters per language over 30 rounds on a -3 to +3 CMOS plus a 1-5 SMOS, and itself lists other papers' rater counts (6, 12). In the extracted text of both papers there is no confidence interval, standard error or inter-rater statistic, so differences such as Seed-TTS CMOS -0.07 versus -0.08 cannot be assessed from the reports.
• articleI have a character say something indirect ("It's really cold in here"): what do experiments say makes an audience read it as a request, and what makes them not?
  In a story-reading experiment, the same sentence ("Man, it's really cold in here") was read as a request far more often when the speaker could be inferred not to know about an obstacle (a broken heater) than when the speaker knew. Conventional forms ("Can you turn on the heater?") are described as easy; unconventional ones need inference about what the speaker knows. People differ widely in how much they use that knowledge. An fMRI study of spoken question-reply dialogues reports mentalising-network involvement for indirect replies. Limits: written second-person vignettes read by online participants, not film or drama dialogue.
• articleWhat are the WCAG 2.2 contrast-ratio and relative-luminance formulas and thresholds (1.4.3, 1.4.11), and what in them applies to text burned into video?
  WCAG 2.2 (Recommendation, 12 Dec 2024) defines contrast ratio as (L1 + 0.05) / (L2 + 0.05) with sRGB relative luminance L = 0.2126 R + 0.7152 G + 0.0722 B (linearised with the 0.04045 piecewise function). 1.4.3 requires 4.5:1 for text and 3:1 for large-scale text (18 pt or 14 pt bold); 1.4.11 requires 3:1 for UI components and graphical objects. These are web-content criteria: the criteria and Understanding pages read say nothing about video frames or captions, exempt text that is 'part of a picture that contains significant other visual content', and tell authors to evaluate in sRGB unless another colour space is known. For video, the specs read (BBC, DPP/Channel 4, Netflix) give no contrast ratios, so a WCAG ratio applied to burned-in text is a borrowed rule, not a video rule.
• articleHow can an agent check a frame, caption or lower-third placement against action-safe and title-safe rules, and what edge cases must the check state?
  No source read provides a safe-area test procedure, so this candidate is a synthesis from the stated rectangles. A check needs (1) a named rule set and raster, (2) rectangle coordinates with a stated numbering base, (3) the text or graphic's bounding box measured at delivery resolution, and (4) a policy for motion, 4:3 protection and per-spec force. Sources supply the rectangles and three subtleties: R 95 counts pixel 0 as the first sample; ATSC/BBC express subtitle limits as percentages (origin 5% 5%, extent 90% 90%); Netflix lets motion text leave the box if it is readable inside at some point. An automated layout/caption checker for generated frames is a proposed hypothesis with a test, not a documented use.
• articleHow do primary motion-generation papers define foot skating, penetration, floating and jitter, with what thresholds and units?
  Definitions differ by paper. PhysDiff (v3): Penetrate and Float are distances from the ground to the lowest body mesh vertex with 5 mm tolerance, Skate is mean horizontal displacement of foot joints in contact in two adjacent frames, Phys-Err is their sum, all in mm. GMD (v3) and Yi et al. (v1) use a skating ratio: fraction of frames where a foot slides more than 2.5 cm while foot height is under 5 cm. Body-momentum (v1) measures jitter as the derivative of acceleration in world coordinates (10 m/s^3) and foot sliding in mm during contact. MDM v2 uses a foot-contact loss in training, not an evaluation metric.
• articleWhich internal-consistency checks can an agent run on a cost report (arithmetic, sign rules, actual vs committed, final-report ETC) using rules that agencies state in writing?
  NZ On Air's April 2026 guidance states several checkable rules: estimated total = actual + unpaid invoices + estimate to complete; variance is shown in its own column (not as negative ETC); negative figures must not appear in estimate to complete; actual cost excludes committed or estimated figures; the final report's estimate-to-complete column should be zero except for final-drawdown items (normally PCO); rows mirror the contracted budget with sub-totals; final reports must be full. Telefilm Canada adds that substantive variances need notes per category. The check list below is a synthesis of those statements, not an agency-issued validator.
• articleWhat do open papers measure for monocular and few-camera markerless capture against reference systems, what is the evidence for commercial tools such as Move AI, and what does it imply for animation production use?
  Open results are research-method numbers, not commercial-tool numbers: on Human3.6M (marker-based) a monocular lifting model reports 46.8 mm MPJPE from detected 2D keypoints (37.2 mm from ground-truth 2D); on in-the-wild 3DPW WHAM reports 57.8 mm; on EMDB global trajectories WHAM reports 354.8 mm W-MPJPE100, 6.0 % root-translation error and 4.4 mm foot sliding. A 2022 clinical review gives 40-60 mm joint-centre differences and up to 30 degree ankle-angle differences for monocular methods, and OpenCap (two or more phones) reports 4.5 degree joint-angle MAE against skin-marker capture, with the authors noting that skin-marker error itself limits what can be shown. No open independent validation of Move AI or Flow Studio mocap against marker-based ground truth was found (bounded Europe PMC search), and Move's own accuracy page gives no numbers. Implication (synthesis): treat markerless output as animation reference or blocking requiring cleanup of root drift and foot contact, not as ma...
• articleWhat simple stakes and suspense check can I run on an outline beat using what these models and experiments say?
  Combining the structural-affect precondition (a significant consequence for a character, then a resolution) with the escape-route accounts used by Dramatis, Suspenser and the Xie-Riedl prompts gives a five-question per-beat check: what is at stake, what outcome does the reader expect, what escape routes are visible, how likely does each seem, and what blocks them. The sources support the pieces; the checklist as a whole is a synthesis and has not been tested by these authors.
• articleHow should upscaled footage be judged: what do primary sources say about full-reference versus no-reference metrics and their limits?
  Blau and Michaeli prove that lower distortion (full-reference measures such as PSNR/SSIM) and better perceptual quality are at odds for any distortion measure, which is why GAN-based upscalers can score worse on PSNR while looking better. Real-ESRGAN reports only the no-reference NIQE and warns such metrics do not track human preference. The VMAF docs describe a full-reference metric trained on subjective data for stated viewing conditions, say it captures gains from enhancement operations, and provide a NEG mode that disables enhancement gain for codec comparisons; its FAQ shows upscaling the distorted clip before comparison. No inspected source states a validated protocol for judging AI-upscaled finishing footage; broadcaster specs require QC but no metric thresholds in the passages read.
• articleHow do the audio-capable world-model papers measure that sound stays in sync with actions, camera motion and events, and what do they leave unmeasured?
  Three different measurement designs exist. AVWM compares the audio-energy vs motion-energy cross-correlation lag and peak of generated audio with that of ground-truth audio against the true visual trajectory (simulator data, 0.15 s steps); HelixBench (1,015 clips) scores DeSync on an Event subset and a Grounding-DINO-based three-region stereo score for camera-aligned panning; MaineCoon reports an unexplained "AV-Al" column where it is behind Ovi. Each protocol scores windows or short rollouts; none of the texts read measures the delay from a user action to the sound onset.
• articleIn an FFmpeg zscale + tonemap HDR-to-SDR chain, which HDR metadata is read, kept, stripped or left stale, and where would an agent silently lose dynamic metadata?
  Reading FFmpeg at 5fd305b8 (code only; nothing executed): the HEVC decoder exposes HDR10+ and Dolby Vision as frame side data, but vf_tonemap derives its peak only from MaxCLL, then mastering-display max luminance, then a fixed default (10.0 = 1000 nits for a non-PQ-tagged frame), and never reads dynamic metadata. vf_zscale removes every COLOR_DEPENDENT side-data type (CLL, MDCV, HDR10+, Dolby Vision RPU/metadata) whenever the output transfer or primaries differ from the input. In FFmpeg's own documented chain the first zscale (to linear) therefore strips the metadata before tonemap runs, so the detected peak falls back to the default unless peak= is set. libx265.c has no HDR10+ handling at this pin, so a re-encode drops it silently; x265 needs --dhdr10-info.
• articleWhat do ITU-R BT.2408 and BT.2390 say about HDR-to-SDR down-mapping, do they cover dynamic metadata, and where is the BT.2390 EETF now?
  BT.2408-8 section 5.2 says display-light down-mapping is usually preferred, that the monitoring down-mapper must match the one used to make the SDR transmission, and that a plain 0.5 linear scale does not preserve the look on a 100 cd/m2 display. Neither BT.2408-8 nor BT.2390-12 mentions dynamic metadata or trims (keyword counts of 0 for 'dynamic metadata' and '2094'); they are static, display-mapping guidance. The hermite-spline EETF that tools label 'BT.2390' appears in BT.2390 editions -5, -7 and -8 but in this extraction not in -9 to -12; BT.2408-8 Annex 5 carries it. This extends C2-B05-B-08 (BT.2100/2408/2446 reference-white and BT.2446 methods) without repeating it.
• articleIn plain terms, what does benign-violation theory predict for a joke or gag I am writing, and where does the open literature say the theory stops working?
  Open sources state the theory as: something is funny when it seems wrong (a violation) and also okay (benign). Use it as a checklist of two appraisals and of severity and distance, not as a proven rule: the open sources read here report that severity, social distance, power and culture shift what counts as benign, that some jokes with non-benign violations were still rated funny, and that no independent non-moral test was readable.
• articleWhere do Wilensky's 'story points' and Labov's evaluation/reportability disagree about what makes a story worth telling, and which agent checks follow from each?
  Wilensky's 1980 Berkeley memorandum treats a story's point as content: a spelled-out problem for a character (goal conflict, competition or lost subsumption) plus a solution component. He gives point prototypes but says his psychological predictions are untested. Labov defines the point as the most highly evaluated action and says interest has no absolute standard, so reportability is relative to the audience. The sources disagree on form versus content and on whether interest can be fixed independently of the audience. The Wilensky text read is the 1980 memorandum, not the 1983 BBS article, which was paywalled.
• articleWhen one model is scored under several action-following protocols (Oasis and Matrix-Game across MineWorld, Matrix-Game, GameCraft and Yume papers), do the scores agree, and are the protocols comparable?
  They do not agree. Oasis keyboard accuracy is 0.77 in Matrix-Game 1.0, 0.73 in Matrix-Game 2.0, and F1 0.41 (precision 0.49, recall 0.44) in MineWorld's own table. Matrix-Game 1.0 scores 0.95 keyboard/0.95 mouse in its own paper, yet in GameCraft's tests (150 mixed images, 12 actions) it has the highest translation error and the second-highest rotation error of five methods (RPE trans 0.18, rot 0.35 versus 0.08 and 0.20 for GameCraft) and the lowest user-study action-accuracy rank (1.63 of 5), and in Yume-Bench its human-rated instruction following is 0.271 against Yume's 0.657. The sources differ in test set, domain, estimator and rater, so these are four separate protocols, not four measurements of one quantity. The differences are consistent with domain shift (Yume says Matrix-Game struggles to generalise to real-world scenes) but no source isolates the cause.
• articleWhat do Chamfer distance, F-score, CLIP score, FID/FD, ULIP/Uni3D and user-preference numbers in image-to-3D papers actually measure?
  The papers use four metric families. Geometry accuracy against a ground-truth mesh: Chamfer distance and F-score on sampled point clouds (TripoSR, SF3D, SPAR3D, TRELLIS, TripoSG for its VAE). Distribution quality of renders: Frechet/kernel distance or FID variants (TRELLIS FD/KD; Hunyuan CLIP-FID and CMMD). Prompt alignment: CLIP cosine similarity of renders to the prompt, and ULIP/Uni3D scores between a sampled point cloud and image or caption (Hunyuan, TRELLIS.2). Human or model preference: authors-run user studies (TRELLIS, TRELLIS.2, Hunyuan 2.0) and, for TripoSG, an LMM (Claude 3.5) judge. Geometry accuracy needs a ground-truth mesh; the alignment and distribution scores do not.
• articleWhat checks could an agent apply to a draft episode ending, and which are supported by measurement versus craft advice?
  Synthesis only. From the sources read, an agent can run a checklist: (1) name the open question the ending leaves; (2) check it is a turn near the end of the act (Silberman, WGF); (3) check for a bridge into the next episode (Alias); (4) label any claim of retention as unmeasured; (5) note arousal side effects (Wirz). No source supplies a validated test of an ending; the checks are proposals.
• articleWhat counts as binge-watching in the research, and what do the studies actually measure?
  A 2020 open systematic review (28 articles, 2013-2020) reports no single definition of binge-watching: studies range from 1-3 to 3+ or more episodes in one sitting, and the review table lists mostly online surveys. The review separates an entertainment perspective from a negative-outcome perspective. It does not test cliffhangers, so it cannot support or refute claims about episode endings.
• articleHow is "resolution satisfaction" or narrative closure defined and measured in open sources, and what check can an agent run on whether an ending answers what the story raised?
  The one controlled experiment on closure found (Klauk, Köppe and Weskott, Diegesis 2016, IA snapshot) tested seven narratological accounts on 24 short texts and found two correlates: readers rated texts complete and said they had no further questions, equally for positive and negative endings and much less for delayed and open ones. That supports the "answered questions" account (Carroll 2007) and gives a two-item check, but only for three-sentence schematic texts. Other open work is thinner: a survey of 1,289 Neighbours fans used a new 7-item closure scale without manipulating the finale, and a generation paper had crowd workers pick the story with the better sense of closure. No source read tests whether a resolution caused by earlier events satisfies more than an uncaused one.
• articleHow could an agent check an ensemble outline for orphaned characters, unbalanced lines and unmotivated convergence, and what in the sources supports each check?
  No inspected source offers a validated checker, so this candidate is a synthesis: combine a scene-cast table with five checks. (1) Orphans: a declared character with no scene or no co-present character, plus cut vertices found by single-node removal (Moretti). (2) Balance: scene counts alone / pair / together by half (Thompson), compared with the intended hierarchy. (3) Neglect gaps: longest run of scenes without a line (Bost et al.), with an intent note because gaps can be deliberate. (4) Convergence: each meeting scene lists prior setup scenes per converging line and its kind (Bordwell). (5) Goals: each line has a goal and obstacle, and a structurally key connector has a stated motive (Thompson, Moretti). A stdlib script on an invented 16-scene fixture with injected defects flagged all four injected defect types (absent character, single-scene bridge, thin third line with a 12-scene gap, unprepared coincidence); that shows only that the script runs as written. Thresholds (2.0 rati...
• articleHow should intercut parallel storylines and temporarily neglected characters be represented in a character network over time?
  Bost et al. argue that whole-narrative aggregation and fixed time slices both mislead for TV plots where scenes alternate between parallel storylines; their narrative smoothing keeps a relationship in force while the characters are off screen and measures importance per scene. They show examples (a secondary character's subplot rising above the protagonist and ending at his death; Game of Thrones storylines in separate communities) but call their evaluation preliminary and qualitative. The method uses verbal interactions only, so silent presence (as in a Favourite scene where one lead stands mute) is not captured; this limit is the worker's synthesis.
• articleHow are character networks extracted from screenplays and novels, how accurate are the steps, and has the quality of extracted networks been evaluated?
  Extraction is a pipeline (identify characters, detect interactions, build the graph) and each step is assessed separately. For screenplays, Agarwal et al. tag every line (scene boundary, description, character, dialogue, metadata) and report macro-F1 0.96 for an ML ensemble vs 0.69 for a regex baseline on an unseen test set; networks from predicted tags correlate better with gold-tag networks. For novels, Elson et al. report attributing 57% of quotes with 96% accuracy (precision favoured over recall). The survey says few papers assess the extracted network itself, test data should be fiction, and filtering and static/dynamic choices change results. All numbers are the authors' reports on their data, not rerun.
• articleWhat do Moretti's Hamlet networks show about protagonist centrality, clustering, weak-tie connectors and node-removal experiments, and which parts transfer to checking an ensemble outline?
  Moretti builds speech-act networks of Hamlet and treats the protagonist as the node nearest all others (mean distance 1.45). Removing nodes is his experiment: without Hamlet the graph nearly splits into court and "beyond"; removing Horatio, though slightly less central by distance than Claudius, fragments the rest because he sits in a low-clustering region and links the Ghost/Fortinbras side to the court. Claudius sits in a dense cluster and his removal changes little. Moretti flags the limits (unweighted, undirected, hand drawn, a few plays) and hedges his Horatio reading. For ensemble outlines the transferable idea is a cut-vertex / removal test; whether it indicates a defect is a judgment, since Horatio is both structurally key and, in Moretti's words, without motivation.
• articleWhat is a character network, and what can count as a link between two characters?
  A character network has characters as nodes and interactions as edges. The inspected sources use at least three edge rules: spoken exchange (Moretti, Elson), co-presence in a scene with dialogue (Agarwal/Weng), and broader co-occurrence; the Labatut-Bost survey lists five interaction definitions overall. Rules give different graphs (co-occurrence graphs are denser), and Moretti's own graphs were unweighted and undirected. For an outline, which rule to pick is a design choice; the sources do not name a best one.
• articleHow could an agent check a frame, UI or chart for CVD problems, and what does a minimal computed check show for common palettes?
  No inspected source defines a pass threshold. A defensible procedure assembled from the sources: (1) inventory colour meanings and add non-colour cues; (2) simulate protan, deutan and tritan with a pinned model in linear RGB; (3) compare each pair of meaningful colours before and after simulation; (4) treat the simulation as a prompt for human review, not as what a person sees. The author's own computation (synthesis, not a source result) applied Machado severity-1.0 matrices to six palettes: rainbow_5 collapses to a minimum CIE76 dE of 8.2-8.3 under protan/deutan from 52.9 normal, a mid red/green pair falls to 9.9 (deutan), and the Okabe-Ito set including R's added gray keeps min dE 7.1 under deutan because reddish purple and gray converge.
• articleWhich colour-blind-safe qualitative palette and sequential colour map have primary sources, and what are the exact Okabe-Ito colours and their stated design rules?
  Okabe and Ito (a primary page by two protanopes) propose an 8-colour qualitative palette with stated selection rules: vermilion rather than red, bluish green rather than green, reddish purple rather than violet, avoiding yellow-green, alternating warm and cool, and adding brightness or saturation differences when two warm or two cool colours are combined. Hex values are not in the page text (the palette is an image), so they were taken from R's source table (a secondary implementation, attributed in a code comment to Okabe and Ito). For continuous data, Crameri et al. 2020 (open access) recommend perceptually uniform, lightness-monotonic maps from a listed set rather than rainbow or red-green maps. Wong 2011 (Nature Methods) full text is paywalled.

More