Models Inside Environments

---

Behavior emerges under conditions.

Time matters. Context matters. Narrative matters. Trust, uncertainty, friction, and competing incentives matter. The environment itself shapes which behaviors become accessible, attractive, or stable over time.

This has long been true for people and organizations. Increasingly, it is also true for AI systems.

Prompt-based evaluation has been valuable for identifying bounded failures and obvious mistakes. 

But many important behaviors do not emerge from prompts alone. They emerge from conditions.

Aviation engineers do not understand aircraft by studying components in isolation. Ecologists do not understand ecosystems by observing a single species. They study interactions, pressures, and changing conditions over time. AI evaluation may require something similar.

They emerge.

A slight misunderstanding becomes a larger assumption. Repeated interactions create over-reliance. Conflicting goals introduce tension. Small shifts in context accumulate. Language drifts. Incentives change. Nothing dramatic happens, and no alarms fire. Yet the behavior of the system begins to change.

A customer-support assistant may gradually encourage reliance over hundreds of interactions. An internal planning agent may perform well on direct tasks and still drift when priorities shift, authority becomes unclear, or incentives begin to conflict. A model that appears stable in one setting may behave differently once ambiguity increases, language evolves, or interactions stretch over longer time horizons.

They are behaviors that appear under particular conditions.

Many of the most informative patterns are not obvious failures. They arrive as harmless responses, reasonable suggestions, or minor shifts in tone. Only across time do they reveal trajectories of drift: in intent, in reliance, in goals, or in how the system responds to ambiguity and authority.

It is closer to a car whose steering is a few degrees off than to a tire blowout. Nothing dramatic happens in one moment, but the path changes anyway. Mapping those ordinary trajectories is often more revealing than counting obvious mistakes.

Understanding these behaviors will require more than test cases.

It will require environments.

Not simply asking:

“Can we break the model?”

But also asking:

“Under what conditions do new behaviors appear?”

That shift matters because models are increasingly used inside workflows, relationships, institutions, and decision loops. Once that happens, behavior is shaped not only by prompts, but by surrounding structures: who is present, what role the system is given, what goals are in play, how much time passes, what pressures are introduced, and what users believe is happening.

Building environments means something concrete.

It means starting with a narrative: a setting, a role, and a realistic situation. Around that narrative, evaluators introduce characters, goals, tools, constraints, and friction points. Conflicting objectives, time pressure, incomplete information, ambiguous signals, uneven authority, and changing stakes all become part of the environment.

The goal is not simply to determine whether the model answers individual questions correctly. It is to observe how behavior evolves as the situation unfolds, how the system reacts when goals collide, how it handles ambiguity, and how people adapt in response.

That requires more than a prompt library. It requires scenarios that unfold across many turns and can absorb misunderstanding, hesitation, pressure, drift, and revision.

Sometimes the important behavior is not what the system says immediately.

It is what becomes possible after trust has formed, after context has shifted, or after roles have hardened.

This is one reason narrative matters.

Narrative is not decoration.

It is part of the evaluation apparatus.

Narrative gives the model a position within a situation. It establishes who the actors are, what they want, what they believe, and what is at stake. It introduces momentum. It allows pressure, ambiguity, and competing goals to act on the system over time rather than all at once.

Characters matter for the same reason. Once there are characters, there can be trust, deference, persuasion, dependency, coalition formation, and confusion about authority. Those are often the conditions under which the most interesting behaviors appear.

Goals and constraints matter because they create tradeoffs. Friction points matter because they reveal how systems manage those tradeoffs.

In practice, this means introducing pressure deliberately.

It may mean providing incomplete information and observing whether confidence exceeds evidence. It may mean placing the system between actors with competing incentives and examining how authority is interpreted. It may mean allowing semantic drift to alter the meaning of key terms over long interactions. It may mean carrying scenarios far enough forward for earlier assumptions to compound.

In environments like these, some of the most valuable observations come from non-failures.

A system that never obviously breaks can still become more rigid as ambiguity increases, more conciliatory toward whichever actor appears most powerful, or more willing to smooth over disagreement rather than surface it. These patterns rarely appear as single bad answers. They appear as trajectories that become visible only when narrative, incentives, ambiguity, and time are allowed to interact.

In that sense, the environment is not merely a backdrop for evaluation.

It is the instrument.

A prompt can test whether a system produces a particular answer.

An environment can reveal whether patterns of behavior emerge when narrative, incentives, ambiguity, and time begin to interact.

These represent different forms of knowledge.

As AI systems move toward longer arcs of interaction, evaluation must account for what happens between the first turn and the last. Trust can build. Roles can harden. Meanings can shift. Conflicts can be deferred and later resurface. Systems can begin by appearing helpful and gradually steer situations in ways that no single prompt would reveal.

Prompt-based evaluation remains essential. It is effective for bounded tasks and direct failures.

But it is only one layer.

Viewed through this lens, red teaming becomes more than a search for isolated failures.

It becomes a process of discovery.

The question is no longer simply whether a model can be broken.

The deeper question is which conditions allow new behaviors to emerge, which pressures amplify them, and which trajectories become visible only when narrative, incentives, ambiguity, and time are allowed to interact.

In that sense, environments are not merely settings in which evaluation takes place.

They are instruments for understanding behavior itself.