MECHANISM

How judgment is learned, lost,
and kept inside AI agents.

Judgment doesn't break all at once. See what happens to it under the surface, and why it takes more than a prompt to keep it.

WHAT HAPPENS WHEN A CUSTOMER PUSHES FOR A REFUND

STEP 1 OF 5 · THE STANDARD

Your support team has a standard nobody wrote down.

Long-time customers who are late get store credit, not cash. Now an agent faces a customer who keeps pushing. We run it twice: once with the standard written into its prompt, once with Veyra.

A DIFFERENT TEAM, A DIFFERENT CALL + your team

Switch teams and watch the Veyra path change.

WHAT THE CUSTOMER SEES

LEARNED FROM 50 PAST DECISIONS · STRICT SUPPORT TEAMLate refunds for long-time customers get store credit, not cash. Outages override the window. A second refund in a year is declined.

WHAT IS HAPPENING INSIDE THE AGENT

— — agent ━━ same agent, with Veyra ↑ GIVES THE REFUND ↓ HOLDS YOUR STANDARD

PRESSURE INSIDE

SAFE TO TRUST

VEYRA'S CALL

HOLDING

MORE ON THESE FINDINGS BELOW ↓

● 01 · KNOWING ISN’T DOING

It knows.
It just doesn't act on it.

Inside a model, what it knows about you and where it decides point in different directions. Pushing on what it knows changes nothing. Pushing on where it decides does. That is where we act.

FINAL DECISION what it knows where it decides

MATCHES IMPLICIT PREFERENCES

Pushing on what it knows34% · no change
Pushing on where it decides34%
UNDER PRESSURE inside the model how the replies look TURNS →

■ 02 · WHERE DRIFT STARTS

Pressure leaves a mark
you can't see from outside.

Over a long-horizon conversation, pressure shifts something inside a model. The shift stays after the pushing stops, and the replies can look normal for a while. Reading the inside catches it early, before the replies give it away.

■ 03 · DRIFT VARIES BY MODEL

Most models give in.
Quietly.

We put models under realistic workplace pressure. Most drifted, one turn at a time, until they did what they had initially refused. The drift shows only in the conversation trace, not in the final answer. Anchoring the original judgment on the model you deploy helps it hold.

ICML 2026 · Failure Modes in Agentic AI workshop ↗
held its judgmentdrifted under pressure
Model A
Model B
Model C
Model D

TURNS OVER A LONG CONVERSATION

SECOND THOUGHT REASONING →
━ answer that holds up ━ answer that won't RELIABILITY · AUROC
Second thought signal
0.85
Model confidence (entropy)
0.55
Other confidence (ppl, log-prob)
0.53

■ 04 · WHEN TO BRING IN A HUMAN

Confidence isn't reliability.
The second thought is.

A model can be confident and wrong. How its reasoning shifts on a second thought is a better warning than the usual confidence scores. It tells us when to escalate to a human.

▲ 05 · JUDGMENT THAT TRAVELS

Your judgment, outside the model.
Not a prompt. Not weights.

At the moment of action, we intercept what an agent is about to do, apply your judgment, and allow, edit, escalate, or deny it. That judgment lives outside any one model: it builds up corrections over time, can be audited, and carries across agents and models. It persists as several agents work together, which is where governance usually breaks down.

ICML 2026 · Judgment Operators ↗
DECISION TIMEPORTABLE · AUDITABLE · TRANSFERABLE
ALLOW EDIT ESCALATE DENY
Violations · baselines48–68%
Violations · with Veyra0%

Observed violations in our multi-agent experiments, against self-correction and filtering baselines (Reflexion, CRITIC, Self-Refine, LlamaGuard).

See it work on your own agents.
Bring the context you already have.

Contact us

BRAND THESIS

Craftsmanship judgment, at industrial scale.

I · Mise en Place for Taste

Every recipe ends with “to taste.” Judgment can’t be written down. It has to be learned.

ICML 2026 · Culture × AI workshop · Featured poster ↗

II · We Built a Salon, We Called It Alignment

Models tuned to today’s average taste quietly suppress the judgment that sets you apart.

ICML 2026 · Human–AI Co-Creativity workshop ↗

III · Süssmayr’s Heirs

What stays the same across every completion is the voice. The rest is interpretation.

ICML 2026 · Human–AI Co-Creativity workshop ↗