How Veritas evaluates a news article

Primary journalism standards repeatedly emphasize clear attribution, source material, original evidence, corroboration, fair response, transparent anonymous sourcing, and faithful representation of uncertainty. Engine 5.10.0 derives article text, HTML, paragraphs, links, evidence locators, and Reader media placement from one atomic canonical article document; candidate agreement counts independent extraction lineages rather than correlated DOM fallbacks. A stable independent consensus controls the canonical-boundary verdict, so weaker losing candidates cannot veto a complete article through the raw disagreement average. A substantial contained subset may confirm a longer complete candidate, while a tiny unrelated page-shell, metadata, or media-card fragment remains diagnostic and cannot overrule it. Two substantial competing article bodies still make the boundary unstable and withhold scoring. Multi-topic newsletters, collection pages, consent-only overlays, and access-verification challenges are rejected before Reader or scoring, even when their shell contains article-like markup. Before a real article reaches Reader or scoring, structural boundary pruning removes clustered headline/deck/byline/date preambles, sharing and follow controls, related-story and editorial-highlight modules, embedded media cards and captions, app, audience and contact prompts, consent and sign-in dialogs, and audio/video player copy while preserving real metadata in its separate role. A bounded recommendation list is removed only when its heading, links, limited prose, and container agree, so later story paragraphs stay in the article. Exact repeated full-body copies collapse to one article. Literal or escaped anchor markup contributes only its visible linked words to this canonical text; validated destinations remain separate Reader metadata and cannot become claims, annotations, Atlas nodes, report text, or comparison text. Extraction then owns one immutable canonical string: Reader blocks, claims, edits, media positions, hashes, evidence relations, and W3C locators use that same snapshot without a second cleanup pass changing its coordinates. Standalone structural headings, media captions, and publication or update timestamps remain Reader context rather than scored claims; real assertions that mention a date or use update language remain eligible. Completeness follows the detected page format, so a complete short brief can pass while an unfinished liveblog, transcript, gallery, or video-first page abstains. Standard analysis can inspect bounded React/Next and application-state payloads, portable rich-text trees, same-site static script data, declarative Shadow DOM, lazy templates, explicit continuations, and verified print, AMP, or canonical alternatives without executing publisher code. It reports unknown hyperlink coverage and cannot see content that exists only inside a live browser runtime. It separates a claim’s subject or object from the person, organization, record, or publication that supplied it: a generic article URL is neutral, and a linked item becomes documentary evidence only when it identifies an evidence carrier and the retained sentence uses it through attribution or a same-clause evidence predicate. A source must be the grammatical supplier of the attributed statement; a reporting-word homograph inside a continuing title or institution phrase cannot turn a person, party, topic, or other sentence object into a source. Stance predicates attach to their grammatical supplier while short quoted characterizations remain labels. Complete spoken quotations use one bounded attribution grammar in automatic detection and Reader review: explicit speaker tags may appear before or after a quote, bounded transcript speaker turns may introduce speech with or without decorative quotation marks, and tags may carry a short recipient or publication context; pronoun and participial continuations resolve only to the closest unambiguous eligible earlier source in the same paragraph, or the immediately preceding paragraph when a quotation opens the next one. Short quoted labels do not dilute quotation coverage. The addressed outlet or audience remains a recipient rather than a speaker. Dates, temporal lead-ins such as “On Monday” and “Late Friday,” discourse markers, contractions, places, and incomplete role phrases cannot become named evidence sources without a grammatical reporting role; explicit dictionaries, archives, registries, handbooks, reports, filings, and records remain documentary carriers rather than people. Guided Atlas keeps Why it scored to the same three highest-priority decisive claims as Summary's Claims to inspect, Trace includes every mapped claim with a real named-source and evidence-passage path, and Find gaps includes every claim connected to an explicit Evidence Gap. Expert presents the complete canonical graph as a map, chain, timeline, or list, and paired reports preserve the same per-article tasks alongside every aligned or article-specific comparison path. A comparison with no eligible claim alignment is reported as zero overlap; it preserves independent score and dimension comparison without inventing a claim relationship. Atlas also separates claim support and evidence quality from headline framing, keeps captions and editorial labels outside the scored claim set, preserves coordinated predicates with their grammatical subjects, binds article metadata and methods to the selected source documents, and keeps deduplicated picture inspection score-neutral; a visible caption may only contextualize an adjacent claim in Atlas. Equivalent repeated propositions remain readable where they legitimately recur but contribute only one scoring unit. Discourse markers, quantified placeholders, and generic collectives cannot become named evidence sources. A direct quotation can support the narrow attribution claim that an identified speaker made the statement, but not the statement’s truth. Identified official documents receive one scoring opportunity across equivalent tracking variants. Summary, Claims, Guided and Expert Atlas, comparison, and the grade derive support and gaps from one serialized relation set; a claim cannot retain support without its evidence relation. Claim Priority routes attention but does not change the grade without calibration. Extraction confidence never silently raises or lowers a score. If extraction is partial, appears cut off, lacks a stable canonical boundary, or exceeds the declared 240-claim, 450-source-signal, or 4,000-relation capacities, every public surface withholds both score and grade. English is the current supported analysis language, so non-English results display a warning and cannot be shared.