About HRO

Classification Labels

Story-Level Labels

Sentiment Labels

Signal Channels

The weights depend on content type — news articles lean 65% editorial, product pages lean 60% structural. Combined: (w_E × Editorial) + (w_S × Structural). Channels are weighted by content type (e.g., Editorial=0.6, Structural=0.4). Combined: (w_E × Editorial) + (w_S × Structural).

Content Type Weights

Content Type Consensus Vote

Getting the content type wrong shifts the score. A news article misclassified as a product page moves 25% of the weight from what was written to how the site behaves. Content type classification directly determines the E:S weight split, so misclassification can meaningfully distort the final HRCB score. For example, an editorial article (ED: E=0.65, S=0.35) misclassified as a landing page (LP: E=0.40, S=0.60) shifts 25 percentage points of weight from the editorial channel to the structural channel.

When several AI models review the same story, they vote on the content type. The majority wins. Lite-mode models don't vote. When multiple models evaluate the same story, the system uses a majority vote to determine the consensus content type rather than trusting any single model's classification. Each full-mode rater contributes one vote; lite-mode raters do not vote (they don't classify content type). The plurality type becomes the consensus content type used for weight assignment.

Two accuracy checks run on the models dashboard: Validation metrics tracked on the models dashboard:

• Suspect rate — structural-heavy types (PO, LP, PR, AC) where the structural channel has no data, suggesting misclassification
• Cross-model disagreement — stories where raters assigned different content types, indicating ambiguous content

Per-Provision Pipeline

Each of the 31 human rights articles gets its own score for what the content says and what the site does. Each article tracks: Each of 31 UDHR provisions (Preamble + Articles 1-30) is scored independently on both channels. Per-provision fields:

SETL — Structural-Editorial Tension Level

SETL measures the gap between what a site says and what it does. A high score means the article's words point one way on human rights, but the site's design and tracking point another. SETL measures the divergence between editorial and structural signals for each article. It captures whether a site's words and infrastructure tell different stories.

Formula: sign(E-S) × √(|E-S| × max(|E|, |S|)). A large gap means the site's articles point one way on human rights, but the site's design and tracking point another. A high absolute SETL means the site's content and infrastructure are misaligned on human rights.

Fair Witness Evidence Layer

Every evaluation separates its evidence into two categories: Inspired by Heinlein's Fair Witnesses from Stranger in a Strange Land, every evaluation separates its evidence into two categories:

The FW Ratio (Fair Witness Ratio) tracks what share of the evidence is directly verifiable — things anyone could confirm by visiting the page. High = grounded in visible facts. Low = relies more on interpretation. Toggle Fair Witness on any story page to see the evidence breakdown. The FW Ratio (Fair Witness Ratio) is the proportion of observable facts to total evidence items. A high ratio means the evaluation is well-grounded in verifiable observations; a low ratio means more interpretive weight. Toggle Fair Witness on any item page to switch between the standard view and a stripped-down view showing only the evidence breakdown.

FW Ratio

FW Ratio = observable_facts / (observable_facts + inferences)
Scale [0, 1]. Higher = evaluation more grounded in verifiable observations; lower = more interpretive weight.

E-Prime Constraint

Observable facts follow an E-Prime constraint — they cannot use "to be" verbs (is, are, was, were). This forces the evaluator to describe specific actions and behaviors rather than making identity claims. "The site is paywalled" becomes "The site displays a paywall overlay after the first paragraph." Observable facts follow an E-Prime constraint — witness_facts must not use "to be" verbs (is, are, was, were, be, been, being). This prevents essentialist claims and forces grounding in specific observable actions, consistent with operationalist epistemology. The constraint applies to facts only; inferences may use "to be" since they are explicitly marked as interpretive.

Evidence Strength & Confidence

Confidence shows how well-supported the score is across all 31 human rights articles. More articles with strong evidence push it higher. Articles with no data don't count. Confidence is an evidence-weighted aggregate across all 31 provisions: H=1.0, M=0.6, L=0.2, ND=0.0. Higher confidence means more provisions had strong evidence.

Directionality Markers

Volatility

How consistently does a piece of content treat different rights? Low volatility = treats all rights about the same. High = strong on some rights, weak on others. Standard deviation of per-provision combined scores. Measures how uniformly content aligns across different rights.

Consensus (Multi-Model)

When several AI models evaluate the same story, scores are averaged with weights. Full-mode evaluations count more than lite. Higher self-reported confidence counts more. Truncated content counts less. When multiple models evaluate the same story, a consensus score is computed as a weighted mean. Each rater's weight combines three factors: prompt mode (full = 1.0, lite = 0.5), self-reported confidence (floored at 0.2 so no model is silenced), and a content truncation discount.

How Evaluations Work

1. Fetch the page and identify its content type (news, marketing, academic, etc.) — this sets the scoring weights
2. Build a background profile of the site: privacy policy, tracking, access model, and other site-level signals
3. Score all 31 human rights articles — what the content says, and what the site does
4. Record the evidence; generate a theme tag, an overall sentiment rating, and a 2–3 sentence summary
5. Calculate the final score, confidence level, say-do tension (SETL), and overall classification

1. A URL from HN's top 7 pages (plus top best/ask/show) is fetched as the unit of analysis (other stories can be manually triggered)
2. The page's content type is classified (Editorial, Community, HR, etc.); when multiple models evaluate the same story, a majority vote determines the consensus type
3. Channel weights are assigned based on the consensus content type
4. A Domain Context Profile (DCP) is constructed from 8 domain-level elements
5. Each of 31 UDHR provisions is scored for editorial and structural signals
6. Fair Witness evidence (observable facts + inferences) is recorded per provision
7. Story-level labels are generated (theme tag, sentiment, executive summary)
8. Ten supplementary signals are assessed (EQ, PT, SO, ET, SR, TF, GS, CL, TD, RTS)
9. Aggregate statistics are computed (weighted mean, evidence profile, confidence, SETL, etc.)
10. A final classification is assigned (Strong positive to Strong negative)

Evaluation Modes

Two modes balance cost vs. depth. Stories in lite mode show ~lite in the feed. Two evaluation modes serve different cost/quality trade-offs. Stories with only a lite evaluation show ~lite in the feed.

Lite item pages display an editorial+structural summary card instead of the full heatmap. Lite scores use holistic two-dimension scoring (editorial + structural) blended with content-type weights.

Psychological Safety Quotient (PSQ) experimental

PSQ measures how safe content feels to read — separate from whether it supports human rights. Torture exposés score high on HRCB (rights-affirming) but low on PSQ (threatening to read). A privacy-violating app landing page scores low on HRCB but high on PSQ (pleasant to read). PSQ is an independent signal measuring reader psychoemotional safety on a 0-10 scale, orthogonal to HRCB's rights-stance construct. The instrument adapts Kline's Psychological Safety Quotient framework (originally validated on social media content against LLM scores, not human ground truth) to news content. Construct independence confirmed by Phase A validation: PSQ threat_exposure and HRCB editorial showed expected divergence on known-construct stories.

Three dimensions are scored per story:

• Threat Exposure — intensity of harmful content the reader encounters (violence, exploitation, psychological manipulation). Lower = safer.
• Trust Conditions — whether the content creates conditions for reader trust (source transparency, logical coherence, respect for autonomy). Higher = safer.
• Resilience Baseline — whether the content leaves readers more or less psychologically equipped (empowerment, coping resources, constructive framing). Higher = safer.

PSQ is experimental until validation criteria are met: (1) pilot distribution across ≥50 stories without clustering, (2) inter-rater ICC > 0.7 across 3 models on 30 stories, (3) CFA factor structure confirmation on N≥100, (4) criterion validity r > 0.4 against human "felt reading safety" ratings, (5) test-retest ICC > 0.8. Scores are displayed normalized to [-1, +1] (via (psq-5)/5) for visual consistency with E and S channels. Scores should be treated as ordinal — rank-order is meaningful, but distances between scores are not. PSQ is currently scored by Workers AI (Llama 4 Scout PSQ); an external DistilBERT scorer at psq.unratified.org collects data for comparison research but lacks sufficient score breadth for production use.

Supplementary Signals

Ten supplementary signals capture how content communicates, orthogonal to HRCB which measures directional lean. These are grounded in established psychometric and information quality frameworks. See the live signal dashboard for global averages and distributions, or the full glossary for all terms and UDHR article links.

Domain Signal Profiles (Factions)

The Factions page clusters domains by editorial character — how they cover topics, not just what they score. Instead of using the 31-dimension UDHR fingerprint, it clusters on 8 normalized supplementary signal dimensions:

Each dimension is z-normalized (zero mean, unit variance) across all domains with 3+ evaluations. Cosine similarity on the resulting 8D vectors measures editorial character similarity. Domains are clustered using agglomerative hierarchical clustering with average linkage at golden ratio thresholds:

Eight domain-level elements provide context modifiers:

Each element shifts per-article scores by up to ±0.30. DCP profiles cached 7 days in KV, persisted to D1. Only used in full evaluations (not lite).

Browser-verified signals supplement LLM-inferred DCP with machine-observed data:

A headless browser visits each domain to directly measure tracker scripts, security headers, accessibility attributes, and consent patterns — grounding structural scores in observation. Cloudflare Browser Rendering (Puppeteer) audits each domain via CDP network interception (tracker counting), response header extraction (HSTS, CSP), DOM evaluation (lang attr, skip-nav, alt text coverage), and consent UI pattern matching (cookie banners, dark pattern heuristics). Results produce four br_* DCP elements with calibrated modifiers, refreshed every 7 days.

A pre-evaluation content classifier that identifies non-evaluable pages before they enter the evaluation queue. Pure regex — no LLM calls.

Runs at two points: cron pre-fetch (primary gate) and consumer (safety net for KV cache misses). Writes gate_category and gate_confidence to the stories table.

RSS Feeds (Atom)

Subscribe to HRCB-evaluated stories via /feed.xml. Filter by stance, UDHR provision, or domain — combine freely.

Filters combine: /feed.xml?article=12&filter=negative = negative stories affecting privacy.
Download OPML to import all 31 UDHR provision feeds into your reader at once.

Embeddable Score Badges

Embed a domain's HRCB score anywhere — GitHub READMEs, blog footers, documentation.

Version History

Technical Details