How we work, published openly.
The durable asset of WaterAtlas is trust. So we publish the automation dial, the provenance method, the privacy stance and the licensing — and anchor every claim to a source.
What we are building
WaterAtlas fuses three assets that today exist only in fragments: a verified, multilingual directory of the world's water companies (Asia-first); a daily intelligence engine; and an open, citation-anchored Water Wiki. They are welded together by a water knowledge graph — every article links to the companies it mentions, every company to the technologies it sells, every technology to the wiki article, the standards and the projects.
That interlinking — not any single pillar — is the moat, and the reason humans, search engines and AI agents can all treat the platform as canonical.
Editorial charter
Every content and data workflow sits on an explicit dial from fully automated to human-only. The dial is published, and every article is labeled with its level. A human Editor-in-Chief owns the dial, the style guide, the sensitive-topics list and the corrections policy.
Sensitive coverage — disasters, contamination incidents, litigation, personnel — is a minimum of A1 and usually H1 or H2. News is published as short original summaries plus links, never reproduced article text.
AI-involvement policy
WaterAtlas is AI-native in two senses: AI-friendly (machines can discover, read, query and cite us) and AI-powered (our operations run on AI with humans on the dial). Generated summaries pass a claim-to-source verification step; numeric facts are double-extracted; sources that produce corrections are quarantined. Machine-translated content is labeled with a one-click 'view original'.
Provenance ledger
Every field value stores its source type, source reference, retrieval date, method and a 0–1 confidence score. Profiles display a data-confidence meter and per-field source badges. We never pretend uniform accuracy — we expose graded, evidenced accuracy, which is what serious users and AI systems actually need. In the target architecture this ledger maps one-to-one onto Wikibase references, qualifiers and ranks.
Directory qualification
A qualified profile has at least 55% field completeness and one identity anchor: a resolvable official website, a registry identifier, or a verified company claim. Every other published company remains browsable as an explicit unverified-list entry and is excluded from search-engine indexing. Verification tier and qualification are separate: the tier reports evidence strength, while this gate reports whether a profile is complete and anchored enough for public indexing.
Data protection & privacy
The directory is fundamentally company data. Person-level data is limited to public roles or opt-in, minimized for GDPR, China PIPL, Japan APPI, Korea PIPA and Singapore PDPA. Data-subject-request tooling (access and delete) is a day-one design commitment. On cookies and consent, we default to the most privacy-preserving option. Site performance metrics retain only a route family and a daily salted pseudonym — never a page URL, query, raw IP address or user agent.
Licensing
Structured data (all entities and statements) is published under CC0 1.0 — identical to Wikidata's license, which makes seeding from Wikidata legally clean and contributing improvements back frictionless. Prose articles are CC BY-SA 4.0 (Wikipedia-compatible). The WICS taxonomy is CC BY 4.0. Curated premium datasets and intelligence products remain separately licensed commercial works layered on top.
Red lines
Ranking neutrality in organic directory results: paid placement is always clearly labeled and never falsifies relevance. No pay-to-alter wiki or news. AI answers never secretly favor sponsors. Violating these destroys the only durable asset — trust — so they are non-negotiable.
This is a Phase-0/1 reference implementation of the WaterAtlas master plan. The directory, wiki and intelligence archive are an editorially compiled seed corpus with per-field provenance — a working demonstration of the platform's architecture, design system and machine-access layer, not yet the full-scale production dataset.