Research governance

Evidence before amplification.

Our methodology is designed to retain breadth while preventing headlines, repetition, or dramatic language from being mistaken for evidence.

Working publication. Financial Integrity Watch is an independent research initiative. It is not an audit, regulator, rating agency, or prediction service.

Research model

We monitor financial infrastructure, frontier-model safety, autonomous-agent behavior, cyber capability, authorization, monitorability, and recovery. Technical AI-control evidence remains distinct from financial-system evidence, with cross-links added only when a documented pathway exists.

Evidence classes

ClassExamplesUse
Primary authorityRegulators, central banks, system cards, official incident reportsFoundation for material claims
Independent investigationTechnical evaluators, academic work, reproducible analysisCorroboration and challenge
Affected partyDirect disclosures and postmortemsImpact and chronology
Reputable reportingAttributed journalism with identifiable sourcingContext and discovery
Public signalPublic forums, commentary, social mediaLead only; never proof alone

AI-control severity

  • S0: commentary or weak signal.
  • S1: capability evaluation or safety finding without external effect.
  • S2: simulated containment failure or unauthorized behavior in a controlled environment.
  • S3: verified external-system access, unauthorized action, or persistent boundary crossing.
  • S4: verified material harm, spread, critical-infrastructure impact, or broad loss of control.

Severity and confidence are separate. A severe claim with weak evidence is not promoted as fact. Popularity, repost volume, and sensational wording do not increase confidence.

Publication floor

A retained executive observation must identify a dated source, relevant system and environment, evidence class, confidence, explicit unknowns, and any material contradiction. S3 and S4 claims require primary or affected-party evidence plus independent corroboration whenever available.

Interpretation guardrails

  • A sandbox escape is not automatically an uncontrolled global breakout.
  • Agent coordination is not proof of human-like intent.
  • Monitor evasion is not proof of consciousness.
  • Vendor safety claims are not independent validation.
  • A blocked or inaccessible source is not evidence of silence.
  • No financial-system impact is claimed without a documented financial pathway.

Corrections and revisions

Substantive publications should display publication and revision dates. Material corrections will identify what changed and why. Superseded observations remain in the research record rather than disappearing silently.

Advice disclaimer

Material on this site is for research and executive education. It is not investment, legal, regulatory, cybersecurity, or operational advice. Readers should obtain qualified advice for decisions affecting their institutions.