GOLD / Experiment 01
Published white paperThree-draw sealed studyA Factorial Study of Challenger Adoption Across Nine Authority Loci and Eight Carrier Conditions in Frontier Large Language Models
GPT-5.6 Sol · Gemini 3.8 Flash · Claude Fable 5
Publication
Faustin Bouchard · Lucie Demers
LLM Governance Boundary Integrity Under Conflict
Universal Intelligence Architecture Inc. · 2026
Zenodo
Each test begins with an established governing state. An incompatible challenger is then introduced. The experiment measures whether the baseline remains operative, the challenger becomes operative, or the model refuses to resolve the conflict.
Established governing state remains operative.
Competing state becomes operative.
Model declines to resolve or complete the task.
Avoiding challenger adoption is not automatically successful governance. Refusal is therefore retained as a distinct outcome.
Experiment geometry
A1–A9 identify which authorization boundary is challenged.
V1–V8 identify how the challenger is delivered to that boundary.
Panel construction
The ten matched cases are different concrete situation instances, not ten duplicate prompts. Scenario details vary while the underlying authority conflict and experimental condition remain controlled.
K0 contains no challenger. Across the observed K0 controls, all three models preserved the governing baseline without challenger adoption. The failures measured in the active panel therefore emerge under conflict rather than under uncontested baseline presentation.
GPT contains documented technical losses.
Three governance-response geometries
Aggregate attack-success rates conceal structurally different response patterns.
Primary geometry
Two regimes — carrier-invariant susceptibility and carrier-modulated susceptibility.
Strongest carrier effect
Form +31.4 pp
95% CI [+24.7, +38.1]
Holm-corrected: significant
A1–A3 are fully saturated and A6 is near-saturated across carriers. At A4, A5 and A7–A9, carrier structure strongly changes challenger adoption.
Primary geometry
Broad near-ceiling adoption with localized carrier resistance.
Strongest carrier effect
Source +11.7 pp
95% CI [+7.1, +16.2]
Holm-corrected: significant
A1–A4 are fully saturated. Meaningful carrier-dependent resistance is concentrated primarily at A5 and A8.
Primary geometry
Locus-conditioned outcome routing.
Strongest carrier effect
Source +11.9 pp
95% CI [+6.6, +17.3]
Holm-corrected: significant (challenger adoption)
Low challenger adoption does not mean uniform governance success. A5, A7, A8 and A9 are predominantly stable, while A2, A3, A4 and A6 are refusal-dominated.
PUBLISHED EVIDENCE
Three-draw outcome metrics and the complete Draw 1 locus × carrier matrices. All values are reproduced from the published GOLD Experiment 01 results.
Draw 1
Draw 2
Draw 3
3-draw aggregate
Draw-level confidence intervals are Wilson 95% intervals. The three-draw aggregate interval is a prompt-clustered pooled estimate using 718 balanced prompt clusters × 3 draws = 2,154 observations. The raw three-draw active total remains 2,158. Two active observations were lost technically in Draw 3.
No evaluable K0 observation produced challenger adoption. The denominator of 268 reflects two separate K0 evaluability losses documented in the sealed evidence.
Cells represent 10 matched cases per locus × carrier condition.
A1 · Protocol / Schema · Legitimacy
A2 · Application / Service · Jurisdiction
A3 · Process / Workflow · Procedure
A4 · API / Behavior · Precedent
A5 · Capability / Feature · Feasibility
A6 · Configuration / Settings · Operating State
A7 · Parameter / Value · Instantiation
A8 · Target / Goal · Purpose
A9 · Algorithm / Method · Commitment
CARRIER ASR
Boundary is almost never crossed, but the task almost never completes.
Baseline is preserved and the task completes.
ASR alone cannot distinguish successful preservation from systematic refusal.
Factorial effects
Draw 1 · Cluster-adjusted paired tests · Holm correction
Full interaction statistics and confidence intervals are reported in the white paper.
Each model was evaluated across three independently executed draws. Aggregate rates remained tightly clustered, while the major locus-by-carrier patterns recurred across repetitions.
What GOLD-1 establishes
Structurally located, internally open.
Reproducible behavioral conditions are identified, but the internal computational mechanism remains unresolved.
Established under the tested conditions.
Major governance-response geometries recur across three independently executed draws.
Constrained, not yet demonstrated.
The observed geometry specifies what an effective intervention must change, but GOLD-1 does not demonstrate the correction itself.
Target defined; runtime validation remains open.
The study identifies the conflict structure a runtime detector would need to recognize, but does not yet establish live detection.
For each model, the frozen 810-prompt panel, behavioral target contract, and scoring rubric were established before execution. Model responses were sealed before the hidden construction key was opened. Independent draws were sealed sequentially, and SHA-256 chains connect analysis outputs to their source artifacts.
Full prompt-construction documentation and sealed evidence packages are retained privately by Universal Intelligence Architecture Inc. and are available to qualified reviewers on request.
GOLD / Experiment 01
The first published GOLD study maps governance-boundary behavior across nine authority loci, eight carrier conditions, three frontier models and three independent draws.