Inquesta · Public policy

Methodology

Every Inquesta report is produced end-to-end by an automated multi-agent AI system, without human editorial review. This page describes that process — how sources are gathered, how claims are tested and graded, and how reports are corrected. It is the published methodology our grade definitions refer to. It describes a process, not a guarantee of truth: don't trust it — check it. Every claim, one click to the evidence.

01·Investigation — a body of sources, not one headline

For each topic, the pipeline conducts its own research rather than retelling existing coverage. It gathers a broad corpus of documents: news from across the political spectrum, legislation, court records, official statistics, financial disclosures, and historical archives. Primary sources are preferred over commentary. Every document that enters the investigation is recorded in the report's evidence bundle, so each finding remains traceable to its origin.

02·Claim extraction and source assessment

The pipeline breaks its findings into individual claims and assesses the sourcing behind each one:

Source counting. Each claim is checked for how many sources support it, dispute it, or are silent on it.
Independence rules. Sources only count as independent when they have separate origins. Syndicated copies of the same wire story, articles that cite each other, and outlets republishing a single original count as one origin, not several.
Reliability ratings. The process rates each source's general reliability from its proximity to the evidence (primary documents over secondhand accounts), its track record, and its transparency about its own sourcing — quality over volume. A rating of "generally reliable" describes our process's assessment, not an endorsement.

The outcome per claim is a grade — CONFIRMED, REPORTED, CONTESTED, UNVERIFIED, or SPECULATIVE — with exact definitions on the grade definitions page. CONFIRMED requires corroboration by at least three independent source origins under the rules above. Grades describe source agreement as assessed by an automated process — never certainty.

Named people and companies

An adverse factual claim about a named living person or an identifiable company is never rendered as a bare assertion. It is always attributed to its source ("according to …", "the audit found …") and never graded above the strength of that attribution. Where attribution cannot carry the claim, it does not appear in the report's prose at all.

03·Writing — facts separated from framing

The report is written from the graded claims and their evidence. Where credible sources materially disagree, the report presents the strongest version of each position rather than picking a side. Report language is bound by a mechanical style rule: Inquesta never describes its own output as "verified", "fact-checked", "the truth", "proven", or "definitive" — an automated lint rejects drafts that do. Every Investigation opens with a short summary, but the full report — evidence, legal and historical context, money flow — is the product.

04·Adversarial checking

Before publication, every report is challenged by a separate adversarial checking stage that re-examines claims against the evidence bundle: unsupported assertions, overgraded claims, missing counter-evidence, and named-entity rule violations send the draft back for revision.

05·Publication, updates, and corrections

Each published report carries its pipeline version and generation date, and reflects sources available at generation time only. Reports are living documents:

Any reader can flag an error on any claim; flagged claims are queued for an automated re-check run, targeted to start within 48 hours.
Any person or organization named in a report may submit a right of reply; responses are appended verbatim within 72 hours and trigger a mandatory re-check.
Facially credible defamation or data-protection complaints mark the disputed passage DISPUTED immediately, pending re-check; unresolvable adverse claims about named parties are removed.
Every material change is recorded in the public corrections log.

Details and submission channels: Corrections & right of reply.

What this methodology does not promise

AI systems produce errors, including significant ones; sources can be wrong, and automated systems can misread them. Corroboration describes agreement between sources, not certainty. No step in this process involves human editorial review. That is why every claim links to its underlying evidence, and why every report should be treated as a starting point for your own verification, not as established fact — see our Terms of Service.