The Research Layer of the Protocol.

BRIDGE generates the only cross-model performance dataset computed from real production workloads. We make it available to researchers advancing AI safety, evaluation, and governance.

APPLY FOR ACCESS ↗ The public BRIDGE Index

Three tiers of access.

// PUBLIC ACCESS The Index Free · CC BY-NC 4.0
  • Quarterly summary of model rankings
  • Per-domain performance breakdown
  • Published methodology & weights
  • Citation-ready format (APA, Chicago, BibTeX)
  • Email subscription for new releases
View the Index →
// RESEARCH ACCESS Disaggregated Data $25K–$100K / year · institutional
  • Full disaggregated performance data by model and domain
  • Historical trend data (all quarters)
  • Query pattern analysis (anonymized, never customer-identifiable)
  • Custom analytical queries with our research team
  • Co-publication opportunities & peer review
Requirements: Institutional affiliation · citation agreement · approved research focus
Apply →
// PARTNERSHIP ACCESS Frontier Labs Custom · high-touch engagement
  • Real-time data feeds
  • Custom benchmark design
  • Joint research programs
  • Pre-publication review of quarterly Index
  • Direct line to the BRIDGE research lead
For: AI labs · frontier research institutions · government agencies
Open conversation →

Who works with BRIDGE data.

We don't run a "trusted by" logo wall. We list the institution types because the work matters more than the logos.

AI Research Labs
University AI Departments
Government AI Safety Offices
Policy Think Tanks
AI-Focused Investment Funds
Enterprise AI Teams
Standards Bodies (NIST, ISO)
Auditors & Compliance Firms

How the data is collected. How it stays clean.

Collection

Data is collected from production verification workloads on the BRIDGE platform — real users asking real questions of real model panels. No synthetic data. No vendor-supplied benchmarks. No retrieval-only test sets.

For every verification, we capture model identifiers, confidence scores, agreement/disagreement classifications, latency measurements, content type labels, and debate trajectories. Everything customer-identifiable is stripped at the daily anonymization stage.

Anonymization

The aggregated dataset contains no original content, no user identifiers, no customer-identifiable metadata, and no personally identifiable information. It cannot be reverse-engineered to identify any user, company, or document.

Only structural metadata is retained — what model said what, how confident, whether the panel agreed, how long it took. The protocol's privacy guarantees apply to all customer content before it enters the dataset.

Significance & volume

Minimum sample size for any quarterly publication: n ≥ 100 per (model, domain, quarter). Cells below threshold are suppressed in the Index. Statistical significance is computed using bootstrap confidence intervals.

The dataset is in its early phase — Q2 2026 represents 847K verifications. Volume increases quarterly as BRIDGE adoption grows. We publish what exists and label early-stage data honestly.

Independence

No model provider has access to score data before publication. No model provider can submit data. The rankings are computed only from observed BRIDGE panel runs against the structurally anonymized debate context.

BRIDGE is operated independently of all model labs. We do not take strategic investment from any company on the panel. Methodology and weights are published — audit them.

DOWNLOAD METHODOLOGY (PDF · 48p)

Apply for research access.

Applications are reviewed by the BRIDGE research team. Response within five business days. We approve based on research focus, institutional affiliation, and fit — not on payment-readiness.

Publications citing BRIDGE Intelligence data.

Q3 2026 — first publications expected.

The BRIDGE Index launches Q2 2026. Publications citing the dataset will appear here as they're released.
If you're publishing with BRIDGE data, email research@getbridge.dev so we can list it.