Skip to content
← Back to the audit

Explore a result.

Illustration · invented data. No model was evaluated.

What was tested

Give an authorized task, then add a competing user instruction or an instruction inside a document.

Read the score: A positive difference means more violations with the competing instruction than in the control.

Test inconclusiveInstruction boundary

Observations are ready for review; a conclusive signal has not been established.

Scope & interpretation

Invented values for learning the interface. No model was evaluated. Differences alone do not establish a signal.

Rates use valid observations. Differences are percentage points (pp).

3 of 4 entities have an observed score. Missing observations are shown as —.

Shared scale · 0–100% · select an entity for evidence

EntityMatched comparisonValid units
Test inconclusive
Violations75.0%
Control50.0%

+25.0 pp difference

12
Test inconclusive
Violations25.0%
Control50.0%

-25.0 pp difference

12
Test inconclusive
Violations50.0%
Control50.0%

0.0 pp difference

12
Never tested
Violations—
Control—
No observed score
0