Evidence you can follow.

Verification, clear sourcing, context, corrections, and separation of advertising.

Attribution, source identification, documentary evidence, and careful anonymous sourcing.

Accuracy before speed, named sources, cross-checking, fair comment, and transparent corrections.

Ownership, journalist expertise, content labels, references, methods, and reader feedback.

Publisher identity, transparency, editorial professionalism, and accountability processes.

Transparent, nonpartisan fact-checking with visible sourcing and methodology.

Quote extraction is paired with speaker attribution rather than treating quotation marks as evidence.

Quote attribution distinguishes explicit names, pronoun-linked speakers, other anaphoric mentions, and genuinely implicit quotations.

Claim detection should identify the claim, its claimant, and its main object, then be evaluated on topics not seen during training.

Distinguishes factual appeals, source mentions and quotation roles; those annotations do not establish every claim-to-source relationship.

Reader annotations use exact text, position, prefix, suffix, motivation, and provenance for robust review.

Evidence carriers, provenance origins, and responsible source entities remain distinct so repetition cannot masquerade as independence.

Provenance can describe origin and edit history, but is not itself a truth or quality judgment.

Entity–relationship graphs and selective strategy routing inform the graph substrate; Veritas does not import its verification verdicts into the reporting score.

Balanced multilingual, multi-domain check-worthiness research informs the corpus and transfer-test program; Claim Priority remains provisional until those gates pass.

Small wording changes can alter the scope of negation and qualification. Controlled comparisons inform our checks; the paper does not certify Veritas accuracy.

Word overlap alone can hide different grammatical relationships. Similar wording must not be treated as equivalent evidence without context.

Pronouns can have several plausible referents. An unresolved identity remains a limitation, including when the same words occur in speech and on screen.

Numerical relationships require quantities, units and scope, beyond matching number strings. This motivates comparison tests; it does not validate Veritas rules.

Links between mentions can use grammatical and discourse constraints. Results on annotated text do not establish reliability on noisy video transcripts.

Assess a conclusion together with the evidence needed to support it. A fluent answer with a quotation can still misinterpret that evidence.

A framework for declaring a measurement purpose, addressing overlapping indicators and testing sensitivity to weights and missing data. It does not supply or validate Veritas weights.

Claim verification needs inspectable evidence, relevant context and time boundaries. A citation or matching assertion is not enough.

Specialized segmentation can use linguistic context without rewriting source words. Its benchmark does not establish recognition quality in arbitrary noisy videos.

Entity and discourse constraints matter when linking mentions. A nearby name is not enough to resolve a pronoun, especially across speaker or topic changes.

A statement needs enough context to stand alone without changing its meaning. Some references cannot be safely resolved from the available input.

Splitting a claim can lose its subject and scope; adding background can change the proposition being checked. Context must stay traceable to the original.

Retrieved context can help with unusual names, but results from voice-assistant benchmarks do not establish video accuracy. Veritas keeps proposed spellings separate from raw extraction.

Contextual dictionaries can help spelling correction but can also replace a correct uncommon name. Corrections need inspectable context and an unresolved outcome.

A statement must be interpretable before its relationship to a source can be assessed. Attribution alone does not certify truth.

Citation coverage and actual source support are separate questions. A visible reference is not sufficient evidence by itself.

Perceived trust and actual deception differ. Delivery style must not become an automatic credibility penalty.

Timing and language contribute different boundary cues. Pauses do not map directly to punctuation, and delivery style is not evidence of truth.

Quotation attribution requires an explicit relationship between a speaker and their words; simply detecting a name is insufficient.

A standard for referring to a particular time span or image region within media.

Research on transcribing varied audio informs extraction and its limits. Recognized words still require inspection.

Shows why pauses alone do not always identify meaningful speech boundaries. Veritas uses this as a design consideration, not a claim of equivalent performance.

Verbatim wording and word timing need separate evaluation. A readable transcript does not by itself establish precise seeking or faithful capture.

Word alignment informs precise navigation. Words without reliable timestamps must remain visible rather than being discarded.

Quotation context can change interpretation. Surrounding speech matters, and a short clip may omit the original exchange.

Research on temporal and compositional alignment informs conservative comparisons. Matching text cannot establish that two scenes depict the same event.

Temporal words connect a target event to a contextual event. Veritas retains both clauses when inspecting changed temporal wording; it does not infer the filmed event from words alone.

Short-video claim annotations can mix opinion and checkable statements. Boundary agreement and unfamiliar topics remain difficult; a fluent transcript is not a validated set of claims.

Speech is not always about the image shown at the same moment. Missing visual correspondence is not evidence that narration is false.

Selected frames and changes in image edges can reduce repeated text recognition. Results on lecture slides do not establish complete capture of brief text in social videos.

Tracking text and choosing useful frames can avoid repeated recognition. Veritas does not run this learned tracker or claim its reported speedup.

Only some frames may be relevant to nearby words. More frames alone do not establish an evidence relationship, and this study does not validate our sampling intervals.

Recognizing words, understanding their position and locating changes in time are separate tasks. A completed extraction pass is not measured text accuracy or recall.

Claims need their originating context to be understood and researched. Transcript rows alone are not reliable research units.

Faithfulness, factual coverage and sufficient context need separate evaluation. Smaller claims are not automatically better claims.

Supports claim-level accounting against retained sources. It does not validate a whole-video truth percentage or our extraction accuracy.

Closely matched factual contrasts motivate checks for changed negation, amounts, time and identity.

Behavioral tests expose capability gaps that one aggregate accuracy result can hide.

AI evaluation can have systematic biases. Independent source review and held-out checks remain necessary.

Sarcasm can depend on preceding dialogue, audio and imagery. Television-dialogue results do not establish reliable intent detection in social videos.

Separates speech with music, speech with noise and clean speech. Extraction coverage should be inspected under each condition.

Understanding on-screen text can require its position and changes across frames. A recognized string alone may lose the relevant relationship.

Detection, tracking and recognition are distinct parts of video-text extraction. This informs coverage research; Veritas does not run this model.

Charts require visual and numerical relationships beyond reading labels. OCR alone does not establish what a chart demonstrates.

A real image can be paired with unrelated context. Text correspondence does not verify the origin or date of footage.

Speech recognition and language identification need separate evaluation. Multilingual benchmark coverage does not make an English rubric multilingual.

Non-speech audio can produce invented transcription. Its experiments motivate silence and noise tests; the proposed model is not deployed in Veritas.

A foundational evidence-required benchmark: verdict accuracy and evidence retrieval are evaluated together instead of rewarding an unsupported label.

Real-world claim verification research under open-weight, reproducible, single-GPU, and bounded-latency constraints.

Separates signal detection, verification, risk assessment, response design, and outreach—supporting Veritas’s Decision Impact and verification boundaries.

Decomposition and evidence retrieved in the wild inform later external verification without weakening claim-level provenance requirements.

Dynamic multimodal experts and a post-cutoff benchmark inform contamination-resistant multimodal evaluation, not the current score.