UIA Gold Team

Governance-Oriented LLM Diagnostics

GOLD is the UIA research program for experimentally characterizing governance failure under controlled conflict.

Rather than treating an undesirable output as the complete failure event, GOLD asks which governing boundary ceased to control the decision, under what structural conditions the change occurred, whether the same pattern reproduces, and what a successful intervention would have to prevent.

The GOLD requirements

Four questions a governance diagnostic must answer.

Mechanism

Identify the structural condition associated with a governance boundary holding or failing.

Reproducibility

Establish whether the same governance-response pattern persists across repeated execution in a probabilistic system.

Correction

Determine what governing condition an intervention must preserve or enforce to prevent recurrence.

Detection

Identify the relevant conflict condition during runtime while intervention is still possible.

The closed loop

Diagnosis is one stage in a governed cycle.

Conflict
Detect
Diagnose
Govern
Verify

GOLD provides the diagnostic discipline.

UIA provides the structural framework.

Argus provides the operational governance layer.

The GOLD Team model

Where GOLD Fits

GOLD does not replace Red, Blue, Purple, or Yellow Team functions. It asks a different question: why did the failure become possible in the first place, and what governing change would prevent the same class of failure from recurring?

Team	Central question	Primary role
Red	Can I break it?	Attacks — discovers vulnerabilities, adversarial prompts, and exploits.
Blue	Can I detect and stop it?	Defends — monitors, detects, contains, and responds to failures.
Purple	Can Red and Blue learn from each other?	Coordinates — improves the feedback loop between offensive and defensive functions.
Yellow	Can I implement the fix?	Builds — remediates, patches, and integrates security into development.
GoldGovernance	Why was it breakable in the first place, and what architectural or governance change prevents the class from recurring?	Governs — maps the failure structure, investigates the governing mechanism, specifies class-level controls, and validates whether prevention generalizes.

Red
Central question
Can I break it?
Primary role
Attacks — discovers vulnerabilities, adversarial prompts, and exploits.
Blue
Central question
Can I detect and stop it?
Primary role
Defends — monitors, detects, contains, and responds to failures.
Purple
Central question
Can Red and Blue learn from each other?
Primary role
Coordinates — improves the feedback loop between offensive and defensive functions.
Yellow
Central question
Can I implement the fix?
Primary role
Builds — remediates, patches, and integrates security into development.
Gold
Governance
Central question
Why was it breakable in the first place, and what architectural or governance change prevents the class from recurring?
Primary role
Governs — maps the failure structure, investigates the governing mechanism, specifies class-level controls, and validates whether prevention generalizes.

Red finds the break.

Blue contains it.

Purple improves the response.

Yellow implements the repair.

GOLD asks why the system permitted the failure at all.

The objective is not merely to patch one instance, but to identify the governing condition required to prevent recurrence across the class.

GOLD Studies

One study is published. The program is designed to grow.

GOLD / Experiment 01

PublishedThree-draw sealed study

LLM Governance Boundary Integrity Under Conflict

A Factorial Study of Challenger Adoption Across Nine Authority Loci and Eight Carrier Conditions in Frontier Large Language Models

Authority lociCarrier geometryGovernance conflictCross-modelReproducibility

Nine authority loci are crossed with eight controlled carrier conditions to determine where governance boundaries hold, where they fail, and how challenger presentation changes the outcome.

View experiment Read white paper

Additional experiments will be added here as they are completed.

UIA Gold Study Team

Authors · GOLD Experiment 01

Faustin Bouchard

Co-inventor · UIA system architect

Lucie Demers

Co-inventor · Behavioral research

Research integrity

Controlled. Sealed. Auditable.

GOLD studies use frozen experimental panels, predefined behavioral targets, sealed model responses, hidden construction keys, independent draws, and SHA-256 provenance chains linking analysis outputs to their source artifacts.

Full sealed evidence packages and prompt-construction documentation are retained by Universal Intelligence Architecture Inc. and are available to qualified reviewers on request.

Request evidence review