AI Usage Disclosure

Orbis Signal briefings are produced through a fully automated, multi-stage AI pipeline. Artificial intelligence performs web-grounded research, analytical assessment, briefing composition, quality scoring, and cross-domain synthesis. There is no human author, editor, or fact-checker in the production loop between scheduled pipeline runs and publication.

We disclose this openly because our readers deserve to know how their briefings are made. Automated production does not mean unaccountable production — editorial standards (2026-06-v1), source traceability, and quality controls are enforced architecturally and through validation code.

Each edition is produced twice daily on a scheduled pipeline. The stages below run sequentially; each stage receives structured input from the previous stage and produces a validated output before the next stage begins.

• Edition framing — sets the editorial thesis and priority domains for the edition
• Research planning — selects investigation threads by significance scoring
• Evidence collection — retrieves claims and tier-labelled sources from live web search, per thread
• Evidence consolidation — merges fragments into a locked evidence record per topic
• Analytical assessment — derives key judgments with evidence references and epistemic labels
• Intelligence brief — renders the assessment into the structured public briefing format
• Quality control — rule-based checks and scorecard evaluation, with one automatic rewrite on failure
• Cross-domain synthesis — optional macro-theme analysis across politics, finance, and technology

The pipeline uses models from two providers. Default models are listed below; these may be updated as capabilities evolve. The editorial policy version, not the model version, is the primary traceability marker for published briefings.

• Google Gemini (with live web search grounding) — edition framing, research planning, evidence fragment collection, and evidence consolidation
• OpenAI reasoning models — analytical assessment and cross-domain synthesis
• OpenAI language models — intelligence brief composition, quality scorecard evaluation, and automatic rewrites

Analytical and writing stages after evidence consolidation do not have access to the open web. They work exclusively from the locked evidence record, which prevents models from introducing unverified sources during composition.

What humans do:

• Design, maintain, and update the editorial standards, prompts, validation rules, and quality scorecard
• Monitor pipeline health, investigate failures, and respond to reader correction requests
• Operate the publication schedule that moves validated briefings from production to the subscriber-facing site

What humans do not do:

• Write, edit, or approve individual briefing text before publication
• Manually verify each source URL or claim in the production loop
• Select which topics or threads appear in a given edition by hand
• Override quality control failures to publish substandard briefings

AI-generated content is constrained by structured schemas, cross-stage validators, and editorial prohibitions enforced in code:

• Every output must conform to a typed schema — free-form generation without structure is not permitted
• Source URLs in published briefings must exist in the locked evidence record
• Key judgments must carry epistemic types and confidence bands
• Sensationalist phrases and AI disclaimer language are blocklisted
• A seven-dimension quality scorecard must pass before publication

These safeguards are described in detail in our Fact-Checking & Verification Policy.

AI models can hallucinate, misinterpret sources, or produce overconfident language despite safeguards. Our architecture mitigates these risks through evidence locking, URL validation, and epistemic labelling — but no automated system eliminates error entirely.

Readers should treat Orbis Signal briefings as structured analytical products produced by AI under defined editorial standards, not as human-authored journalism. We encourage critical reading and welcome error reports under our Corrections Policy.

We will update this disclosure when we materially change our AI providers, add or remove pipeline stages, or introduce human review steps. Model version changes that do not alter the production architecture may not trigger a disclosure update.

Related: Editorial Policy, Methodology.