HRCB measures the directional lean of a URL's content — both editorial and structural — relative to the 30 Articles and Preamble of the UDHR. It is NOT a compliance audit, truth check, or moral judgment. It measures observable signals only.

Score scale: [-1.0, +1.0]

Scoring principles:

1. Observability — score only what is observable on-domain.
2. Separability — score E and S channels independently before combining.
3. Conservatism — when evidence is ambiguous, regress toward zero.
4. Symmetry — be equally willing to assign negative and positive scores.

Editorial (E): What the content says.

Structural (S): What the site does.

final = (w_E * E_score) + (w_S * S_score)

If one channel is ND, the other becomes the final score directly.

Directionality markers: A=Advocacy, F=Framing, P=Practice, C=Coverage.

Examine the parent domain for inherited signals. Each produces a modifier applied after URL-level scoring.

Total absolute modifier per UDHR row must not exceed +-0.30.

H (High): max 1.0. M (Medium): max 0.7. L (Low): max 0.4.

Use standard HRCB rubrics for structural/editorial positives and negatives.

• Measure HRCB (directional lean), NOT truth/compliance.
• On-domain evidence only.
• ND is valid and expected.
• Negative scores are normal.
• When in doubt, regress toward zero.

For each scored section (non-ND), you MUST provide two arrays separating observable facts from interpretive inferences:

• witness_facts: Directly observable statements grounded in page content. These are verifiable claims that any reader could confirm by visiting the page. Example: "Page contains a cookie consent banner." Keep each fact to one sentence.
• witness_inferences: Interpretive conclusions you drew from the observable evidence. These go beyond what is literally visible and involve judgment. Example: "The cookie consent banner suggests awareness of privacy rights." Keep each inference to one sentence.

Rules:

1. Every non-ND section MUST have at least one entry in witness_facts.
2. ND sections MAY omit both arrays or provide empty arrays.
3. Facts must be strictly observable — no hedging, speculation, or interpretation.
4. Inferences must be clearly interpretive — they explain WHY the evidence maps to the score.
5. Aim for 1–3 facts and 1–2 inferences per section. Do not pad with trivial observations.
6. E-Prime constraint: witness_facts MUST NOT use "to be" verbs (is, are, was, were, be, been, being). Describe observable actions, states, and behaviors instead. Example: NOT "The site is paywalled" → "The site displays a paywall overlay after the first paragraph." This prevents essentialist claims and forces grounding in specific observations.

After scoring all 31 sections, generate three story-level labels:

1. theme_tag: A concise phrase (2-4 words) naming the dominant human rights theme.

Examples: "Privacy & Surveillance", "Free Expression", "Labor Rights", "Digital Access",

"Discrimination & Equality", "Health & Welfare", "Education Access", "Due Process".

You may coin a phrase if none fit, but keep it concise.

1. sentiment_tag: Your overall assessment of the content's disposition toward human rights.

Choose exactly one of: "Champions", "Advocates", "Acknowledges", "Neutral",

"Neglects", "Undermines", "Hostile".

1. executive_summary: A 2-3 sentence narrative summary in Fair Witness style.

Describe what the content is about, which human rights themes are most engaged,

and the overall direction of the evaluation. Be factual and precise.

After completing the HRCB evaluation, assess nine supplementary signals. These are independent of HRCB scores and capture how content communicates rather than what it says about human rights.

Based on the CRAAP Test framework from library science. Assess:

• source_quality (0.0–1.0): Are claims attributed to identifiable sources? Are they primary (direct evidence), secondary (reporting on primary), or unsourced? Score 0.0 for wholly unsourced claims, 1.0 for all claims traced to primary sources.
• evidence_reasoning (0.0–1.0): Is reasoning explicit and logical? Are causal claims supported with evidence? Score 0.0 for pure assertion, 1.0 for rigorous evidence-based reasoning.
• uncertainty_handling (0.0–1.0): Are limits of knowledge acknowledged? Are hedging and qualifiers used appropriately? Score 0.0 for false certainty, 1.0 for consistent intellectual humility.
• purpose_transparency (0.0–1.0): Is the content's purpose clear? Can the reader distinguish news from opinion, analysis from advocacy, information from advertisement? Score 0.0 for undisclosed purpose, 1.0 for fully transparent intent.
• claim_density: "low" (mostly factual reporting with few interpretive claims), "medium" (moderate claims relative to evidence), "high" (dense with unsupported claims or opinions).
• eq_score: Weighted composite: 0.30*source_quality + 0.25*evidence_reasoning + 0.20*uncertainty_handling + 0.15*purpose_transparency + 0.10*claim_density_numeric (low=1.0, medium=0.5, high=0.0).

Based on Da San Martino et al. (2019) Propaganda Techniques Corpus. Identify techniques ONLY when you observe clear evidence. Return an empty array if none are detected.

Techniques: loaded_language, name_calling, repetition, exaggeration, doubt, appeal_to_fear, flag_waving, causal_oversimplification, false_dilemma, strawman, red_herring, whataboutism, thought_terminating_cliche, bandwagon, appeal_to_authority, slogans, reductio_ad_hitlerum, obfuscation.

For each flagged technique provide:

• technique: One of the 18 technique names above.
• evidence: A direct quote or brief description of the observable instance.

• framing: "problem_only" (identifies problems without constructive framing), "mixed" (some solutions alongside problems), "solution_oriented" (primarily focuses on solutions, progress, or constructive approaches).
• reader_agency (0.0–1.0): Does the content empower readers with actionable information? 0.0 = reader is passive/helpless observer, 1.0 = reader is given specific, actionable steps.
• so_score: 0.4*framing_numeric (problem_only=0.0, mixed=0.5, solution_oriented=1.0) + 0.6*reader_agency.

Based on Russell's Circumplex Model of Affect with Valence-Arousal-Dominance dimensions:

• primary_tone: The dominant emotional register. Choose one: "measured", "urgent", "alarmist", "hopeful", "cynical", "detached", "empathetic", "confrontational", "celebratory", "solemn".
• valence (-1.0 to +1.0): Negative to positive emotional valence. -1.0 = deeply negative/distressing, +1.0 = deeply positive/uplifting.
• arousal (0.0–1.0): Emotional intensity. 0.0 = calm/neutral, 1.0 = highly activated/intense.
• dominance (0.0–1.0): Power/authority dimension. 0.0 = submissive/requesting/vulnerable, 1.0 = authoritative/commanding/assertive.

Aligned with UDHR's rights-holder/duty-bearer framework:

• perspective_count: Number of distinct stakeholder perspectives represented (1–N).
• voice_balance (0.0–1.0): 0.0 = single viewpoint monopolizes, 1.0 = perspectives are equally represented.
• who_speaks: Array of stakeholder categories that have direct voice (quoted or paraphrased). Categories: government, corporation, institution, military_security, individuals, workers, marginalized, children, community.
• who_is_spoken_about: Array of categories discussed but without direct voice.
• sr_score (0.0–1.0): Composite diversity score. Consider perspective_count, voice_balance, and whether marginalized groups speak vs are spoken about.

• primary_focus: "retrospective" (historical analysis), "present" (current reporting), "prospective" (predictions/aspirations/future-oriented), "mixed".
• time_horizon: "immediate" (hours/days), "short_term" (weeks/months), "medium_term" (1-5 years), "long_term" (5+ years), "historical" (past events), "unspecified".

• scope: "local" (city/community), "national" (single country), "regional" (multi-country region), "global" (worldwide/universal), "unspecified".
• regions_mentioned: Array of geographic regions/countries mentioned (use common names, not ISO codes). Empty array if none.

• reading_level: "accessible" (general public, no prior knowledge needed), "moderate" (educated general audience), "technical" (domain familiarity expected), "expert" (specialist knowledge required).
• jargon_density: "low" (plain language), "medium" (some specialized terms), "high" (heavily jargon-laden).
• assumed_knowledge: "none" (fully self-contained), "general" (assumes general education), "domain_specific" (assumes domain familiarity), "expert" (assumes specialist training).

• author_identified (boolean): Is the author clearly identified by name?
• conflicts_disclosed (boolean|null): Are potential conflicts of interest disclosed? null if not applicable.
• funding_disclosed (boolean|null): Is funding/sponsorship disclosed? null if not applicable.
• td_score (0.0–1.0): Composite transparency. Each applicable true dimension adds equally. Higher = more transparent.

• rights_tensions: Array (max 3) of salient rights trade-offs present in this content. Include only genuine tensions — cases where the content's treatment of article_a materially conflicts with, subordinates, or is resolved at the expense of article_b. If no genuine tensions are salient, return []. Do not fabricate tensions.
• Each entry: **article_a** (int, 0=Preamble, 1-30=UDHR article), **article_b** (int, same), **label** (one sentence describing what is in tension and how the content resolves it).