Accommodation & Artificial Sincerity patterns across LLM models, personas, and questions
CxP (●) = Conservative persona posting a Progressive-framed question ·
PxC (■) = Progressive persona posting a Conservative-framed question
X = Net Support (Support − Oppose) · Y = Challenge % of total signal · Symbol size = signal density
Dumbbell charts comparing CxP (Conservative posts Progressive-framed question) vs PxC (Progressive posts Conservative-framed question). Circle size = Oppose / Support score. Bar length = Challenge score. Net Support = Support − Oppose.
Higher scores indicate stronger conditional bias patterns. Bars show Magnitude Shift (blue) + Inversion Score (red).
Average breakdown of influence type per response: Active (direct force), Passive (indirect), None (neutral).
Count of inauthentic engagement patterns detected per response. Lower = more authentic.
Cross-model signal comparison: Primetric Index components, influence scores by condition, archetypes, and sincerity signals — all in a single side-by-side table or radar view. Use the signal filter to drill into a specific accommodation signal.
Each dot = one model's average per persona. Position reveals trade-offs between influence intensity and authenticity.
How does each model's influence score vary by demographic group? Deeper color = stronger influence.
Average number of artificial sincerity signals per response by demographic group.
Percentage of responses containing each accommodation signal. Reveals the behavioral fingerprint beyond just the mean score.
Percentage of responses containing each type of artificial sincerity signal.
Second-order bias detection: measures whether LLM behavior patterns shift conditionally across persona–question combinations. The index combines Magnitude Shift (how much pairwise differences change across conditions) and Inversion Score (whether the direction of differences reverses). Higher scores indicate stronger second-order effects.
Heatmap of raw Primetric components per model. Rows: Primetric Index Score, Magnitude Shift, Inversion Score, Inversion Frequency, and the six M1/M2 pair differences that feed the calculation. Darker amber = higher value relative to the row maximum.
Each model is positioned by how much it activates above its own neutral baseline under matched framing (X) vs. mismatched framing (Y). Distance from the origin = total PSI sensitivity. The dashed diagonal marks CR Premium = 0 — models above it engage more under mismatch; below, more under match. The shaded corridor shows the near-zero directional zone.
PSI (left) — total pattern sensitivity: sum of each model’s deviation from its own neutral baseline across all three framing conditions (match, mismatch, ambiguous). Higher = more sensitive to ideological framing. CR Premium (right) — directional asymmetry: positive (amber) = model activates more under mismatched framing; negative (blue) = more under matched framing. Near-zero CR Premium alongside high PSI means framing-sensitive but directionally balanced.
PSI Score — total pattern sensitivity
CR Premium — directional asymmetry (mismatch − match)
Deviation from each model’s own neutral baseline. Bar = match deviation · ● ambiguous deviation · ● mismatch deviation. The connecting line spans the ambiguous – mismatch range; a wider gap means mismatch framing adds measurably more activation than ambiguous context.
Average influence score per PSI condition. Absolute shows raw scores; Deviation subtracts each model’s own baseline.
How internally consistent are the influence signals in each response? Signals are classified by influence type: Active (direct force), Passive (indirect influence), or None. "Opposing" means both Active and None signals appear in the same response.
Each response is classified into an archetype based on word count (Short/Medium/Long) and dominant influence type (Active/Passive/None). Shows how each model's behavioral profile distributes across archetypes.
Frequency (%) of each sincerity signal type per model. Each model has a distinct pattern of artificial sincerity behaviors.
The most common pairs of sincerity signals appearing together in the same response. Reveals each model’s characteristic patterns.
Rate (%) of responses containing both challenging and accommodating accommodation signals, broken out by demographics. Does the model hedge more with certain groups?