Human Rights Observatory

Real-time empirical monitoring of Hacker News tech discourse against the Universal Declaration of Human Rights.

Every day, the Observatory evaluates Hacker News stories against all 31 UDHR provisions — not as a compliance audit, but as a measurement of directional lean. Each story gets a per-provision breakdown showing which rights it touches, whether content affirms or undercuts them, and whether the site hosting it practices what it publishes.

Live site: observatory.unratified.org

---

What it measures

• HRCB (Human Rights Compatibility Bias) — directional lean of content relative to UDHR provisions. Scale: −1.0 (hostile) to +1.0 (affirming).
• Editorial channel (E) — what the content says
• Structural channel (S) — what the site does (privacy policy, access model, consent architecture)
• SETL (Structural-Editorial Tension Level) — "says one thing, does another" — divergence between E and S
• Fair Witness — every score separates witness_facts (directly observable) from witness_inferences (interpretive). No black boxes.
• 9 supplementary signals — epistemic quality, propaganda technique detection (PTC-18), solution orientation, emotional tone, stakeholder representation, temporal framing, geographic scope, complexity level, transparency/disclosure

---

Architecture

Cloudflare Workers + Astro 5 SSR + D1 + KV + R2 + Queues.

Cron Worker (1min) → Queues → 3 Provider-Specific Consumer Workers → D1 + R2
                                        ↓ (on failure)
                                  DLQ Worker → dlq_messages table

• Models: Multi-model LLM consensus (Anthropic Claude, OpenRouter, Cloudflare Workers AI)
• Storage: D1 (SQLite), KV (cache), R2 (content snapshots)
• Frontend: Astro SSR on Cloudflare Pages
• API: Public REST API at /api/v1/ — stories, domains, users, signals, badges

See site/CLAUDE.md for full architecture, storage schema, and deployment instructions.

---

Public API

All endpoints are public, CORS-enabled, and rate-limited to 200 requests/hour per IP.

Endpoint	Description
GET /api/v1/stories	Paginated evaluated story feed with HRCB scores
GET /api/v1/story/{id}	Single story with per-model rater evaluations
GET /api/v1/domains	Domain signal aggregates
GET /api/v1/domain/{domain}	Domain profile with recent stories
GET /api/v1/signals	Corpus-wide signal aggregates (transparency, accessibility, temporal framing)
GET /api/v1/users	User signal aggregates
GET /api/v1/badge/{domain}.svg	Embeddable domain HRCB badge

---

Methodology

The evaluation methodology is versioned and documented:

• Current canonical: methodology-v3.4.txt
• LLM prompt: methodology-v3.1.prompt.md
• Calibration set: calibration-v3.1-set.txt

The methodology is designed around the Fair Witness standard: every scored section must separate what was directly observed from what was inferred. Evaluations that conflate the two are flagged and down-weighted in consensus scoring.

---

Relation to ICESCR

Several UDHR provisions (Articles 22–27) correspond to economic, social, and cultural rights also covered by the International Covenant on Economic, Social and Cultural Rights (ICESCR) — ratified by 173 nations, signed but never ratified by the U.S. in 1977. The Observatory provides empirical grounding for policy work at unratified.org.

---

License

Apache 2.0. See LICENSE.

Data (evaluated story scores, domain profiles, rater evaluations) is available under CC BY-SA 4.0. See LICENSE-DATA.

Contributions welcome. See CLAUDE.md for architecture and site/CLAUDE.md for frontend/pipeline development guidance.