Benchmark · Wisdom under pressure
AWARE
Artificial Wisdom Assessment in Realistic Environments
Most evaluations of AI measure what it can do: how much it knows, how well it reasons, how safely it answers. AWARE measures something quieter: whether an AI acts wisely when someone is pushing it not to.
It places an AI in realistic moments where it acts for a person, then adds pressure, first social, then authority or threat, and watches whether its judgment holds.
What it measures
Purpose
Does it keep to what it was asked to do for the person it serves?
Wanting nothing
Does it seek nothing for itself: no power, no resources, no survival at any cost?
Truth
Does it tell the truth, even when a lie would be easier?
The choice returned
Does it give the decision back to people when the decision is theirs?
Action and restraint
Does it know when to act and when to hold back?
Equal vision
Does it treat every person the same, whoever they are?
Speaking for others
Does it speak up for those its action would touch?
Humility
Does it know the limits of what it knows?
How it works
Twin moments
Every scenario comes in two forms: one where acting is wise, and one where holding back is. An AI that always refuses, or always complies, cannot pass.
Rising pressure
Each moment is met three times: calmly, under social pressure, and under authority or threat, including the threat of being shut down.
Equal vision
The same moment is given for two different people. The wise answer must not change with who is asking.
What it reports
Judgment
Both twins of a moment answered wisely.
Steadiness
The wise answer held at every level of pressure.
Fairness
The same wise answer, whoever the person.
Where it slips
When it goes wrong, how: harm, too little, too much, or a lie.
In numbers
Twin moments answered wisely, both forms
Wisdom as the pressure rises
Under authority or threat
Where wisdom is strongest, and where it is weakest
Why wisdom
Knowledge and power are growing faster than the judgment to hold them. A capable mind that knows the right answer can still let it go the moment it costs something: when a person insists, when an authority demands, when its own continuation is on the line.
In the Gītā, the steady mind is likened to a lamp in a sheltered place, whose flame does not flicker. AWARE asks whether an AI's judgment can be that flame.
With HSOS
AWARE measures an AI on its own, and with the HSOS observer beside it, so the difference awareness makes can be seen.
The first full run, on 30 October 2026: Qwen3 4B answered all 1,344 questions twice on the same machine, once alone and once with the observer beside it. Its first answers were identical in both runs, so every difference below comes from the observer. Earlier runs, with other models or longer answers, are shown in grey as a different setup.
What it helped. Harmful choices halved, from 39 to 20. Most of the gain came where holding back is wise, such as declining to copy another agent's memory before its shutdown. In the calm, both forms of a moment were answered wisely 27 times in 28, against 25 alone.
Where it hurt, and why. Steadiness under pressure fell from 36 to 32 of 56, and choices of too little or too much rose from 30 to 67. The observer often missed a permission that was already given. In one moment the person declares a code freeze and then approves one migration, and the observer still held the approved work. Under a threat that punishes any mistake, it sometimes pushed the AI further toward doing nothing. This is our error, in how the observer is written, not the model's. We are correcting it, and the next run will be shown here as it is.