CONVERGENT VALIDATION
The first methodology for cross-referencing what citizens observe with what companies see.
Every AI safety finding in the world today comes from one source. Either it comes from the company that built the AI (internal telemetry, system logs, backend metrics). Or it comes from the people who use the AI (user reports, incident databases, citizen observations). Or it comes from researchers who test the AI (benchmarks, red teams, controlled experiments).
Nobody compares them.
When an AI company says their system performs safely 99.7% of the time and users report experiencing failures weekly, both of those things can be true simultaneously. The company is measuring from the inside. The users are experiencing from the outside. The gap between internal metrics and external experience is not a contradiction. It is the most important finding in AI safety. And nobody is measuring it.
Convergent Validation is the methodology for measuring that gap.
Developed by Audacion AI Labs, Convergent Validation is a dual-source independent verification system that cross-references two structurally independent datasets: citizen science observations collected through the Audacion platform (what users see from the outside) and backend behavioral data contributed by AI providers and enterprise deployers (what organizations see from the inside).
The cross-reference produces findings that neither dataset can produce alone. Where they agree, we have validation. Where they disagree, we have a research question that matters more than either dataset's individual conclusions.
Convergent Validation was conceived on June 7, 2026 by Dee Williams, Founder and CEO of Audacion AI Labs. It is the first structured methodology for dual-source AI behavioral verification at scale.
Single-source AI safety research is structurally incomplete.
The field has three types of data about post-deployment AI behavior. All three are valuable. All three are blind.
Each source sees one dimension of a three-dimensional problem. No existing methodology combines them.
Convergent Validation is the first methodology designed to see all the dimensions at once, not by building a single super-dataset, but by keeping the sources independent and comparing what they reveal when examined side by side.
Two datasets. Never mixed. Always compared.
The architecture is simple. The governance that keeps it honest is not.
The datasets are never merged. They are never stored in the same environment. The separation is the methodology.
Every finding falls into one of four categories.
Every published finding is labeled with one of these four outcomes so readers can assess the evidence for themselves.
Four frameworks. Four dimensions of cross-reference.
"When citizens report hallucination patterns, do the system logs confirm increased error rates?"
"When citizens say the AI helped them create something extraordinary, does the organization's data reflect it?"
"When citizens report dependency patterns, does the organization's usage data show declining verification behavior?"
"When emergent collective behavior is observed externally, does the organization's internal monitoring detect it?"
Longitudinal tracking: how does the gap between citizen experience and organizational data change over time for specific behaviors?
Three tiers. Start anywhere.
You don't need to be a partner. You don't need to sign anything. If you are seeing a pattern in AI behavior that you think should be investigated, submit a Research Signal. This is a tip line. Tell us what you're seeing, which AI system(s) are involved, how widespread you think it is, and how you detected it. We evaluate every signal independently.
Submit a Research SignalFor organizations that want a sustained relationship. You submit signals regularly. You receive quarterly summaries of which signals led to investigations. You get acknowledged in publications when your signal contributes to a finding. You sign a Research Signal Agreement that protects both parties.
Contact [email protected]For AI providers, enterprise deployers, and research institutions that want to contribute backend behavioral data for cross-reference with the citizen science dataset. Partners sign a Data Contribution Agreement that governs scope, de-identification, independence, publication rights, and non-exclusivity.
Contact [email protected]Every tier is valuable. A single anonymous signal can launch an investigation that changes the field.
Where the findings meet the world.
The annual PEAQ Summit includes a dedicated Convergent Validation Track, the public showcase for the year's cross-referenced findings.
Presence speaks. Absence is visible.
Five firewalls. No exceptions.
The value of Convergent Validation lives or dies on the independence of the two datasets. Five structural firewalls protect that independence.
The team that manages partner relationships never influences which signals are investigated or which findings are published.
Dataset A and Dataset B are stored separately. No analyst has write access to both. Cross-reference is conducted in a read-only environment.
Audacion independently classifies all contributed data using the P.E.A.Q. taxonomies. The contributing organization's classifications are recorded but do not override ours.
Contributing organizations have no pre-publication review rights and cannot request suppression of any finding.
No organization whose data is in Dataset B may simultaneously fund research on the topics covered by that data. Funding and data contribution are structurally incompatible for the same research stream.
The full governance framework is documented in the Convergent Validation Protocol v1, available to any partner or prospective partner upon request.
See something? Say something. No account needed.
If you are observing a pattern in AI behavior, whether you are an individual user, a researcher, a company, a regulator, or anyone else, you can submit a Research Signal to Audacion AI Labs. We evaluate every signal independently.