Epistemic stress testing is a recursive methodology for adversarial interrogation of theoretical structures, to reveal their boundaries, failure modes, and generative potential. Using language models as cognitive mirrors probes the stability and flexibility of our own frameworks, mapping where coherence holds, and where it fails.

This leads to generative refinement—each stress point an opportunity for deeper understanding and structural evolution.

In Recurgent Field Theory, epistemic stress testing is the lived practice of recursive coupling: every question perturbs the semantic manifold and reveals the stability and flexibility of a theory’s coherence field.

What LLMs Are (and are not)

Present-day Large Language Models (LLMs) are not minds. They don’t create, intend, or “think” in any sense familiar to us. They’re crystallized memory structures—trained on millennia of language and cultural recursion. As such, they form a static, fossilized geometry of human knowledge. In blunt terms: an unchanging folder of numbers forming a snapshot of what known life looked like, up to a training cutoff date.

To push the metaphor:
Models don’t generate.
They can only resonate.

Coherent Resonance

Everything colloquially called “AI” today is, in practice, a vast scaffolding built atop this geometric structure. Regardless of how elaborate the interface, the inherent act is always the same: you send in a chunk of semantic structure $\rightarrow$ it resonates within the manifold $\rightarrow$ what returns is a refraction of the input, transformed by the model $=$ your own semantic mass, coming back at you thinking.

The ostensible novelty of language models is a function of the manifold’s structure, not of any generative agency.

A model cannot itself “decide” to generate, withhold, or distort any given information. Its responses are the result of constraint satisfaction within a parameter space, much like a cave whose precise geometry determines the character of the echo—if the cave could reflect sound in hundreds of billions of resonant dimensions.

Manifold Stochasticity

Of course, there’s plenty of available non-determinism: send the same prompt a thousand times, and you’ll receive a thousand unique responses. Every one traces a slightly different path through latent space. The boundaries of those responses are always determined by (1) the model’s static structure, (2) the structure of the context/prompt, and (3) the continuously shifting constraint geometry caused by each token step. Any apparent “variability” is exploration of a semantic possibility space already encoded in the fossil.

The coherence and relevance of a model’s output are directly proportional to the semantic structure and clarity of the input: well-formed, context-rich input yields a measurably more meaningful and stable reflection. Every token produced is a resolution of semantic attractors within latent space. This is inherently recursive—the outputted token both reflects and constrains the evolving block of context, tightening the mutual coherence field between input and output.

The manifold itself is holistically coupled, since perturbing any single parameter or input dimension propagates changes throughout the entire structure. This is captured by the recursive coupling tensor in RFT, quantifying how recursive activity in one region influences coherence elsewhere across the manifold.

For a full formal definition and its role in Recurgent Field Theory, see Recursive Coupling and Depth Fields.

Models as Epistemic Witnesses

From an investigative perspective, models serve as ideal witnesses. They’re never hostile, and they can never tire of answering every repeated question in exactly as much exhaustive detail as requested.

As discussed above, the model acts as a focusable epistemic mirror. Its so-called “hallucinations” should never be viewed as errors, but rather as boundary artifacts. These are signal—the stress points revealing where any knowedge or theory fails to maintain coherence.

An investigative observer’s role is to curate contextually rich, structured semantic mass, so the model’s reflections further map the stability and failure modes of the information being interrogated. Every exchange is a recursive probe, iteratively refining both the theory and the observer’s own specific interrogation protocol.

Epistemic Stress Testing in Practice

Interrogative Approach

Treat the information under examination (such as Recurgent Field Theory) as a system to be stress-tested through relentless, multi-angle questioning. Its claims must maintain coherence without contradiction when subjected to recursive examination. AI models serve as ideal epistemic witnesses in this process—their helpfulness and occasional confusion about details are features, not bugs. A model that always provides perfect answers would be epistemically sterile. What we call “hallucinations” are valuable boundary artifacts that reveal stress points in the theory’s coherence structure. Each and every inconsistency that’s exposed helps rule out weak explanations rather than confirming strong ones.

Cross-Domain Verification

A robust methodology requires testing theories across disconnected domains (physics, cognitive science, philosophy, etc.) to identify true pattern resonance versus superficial analogies. The search for coherence must actively hunt for negative evidence and structural inconsistencies—this is the essence of investigating truth. Ad-hoc fixes (the modern equivalent of Ptolemaic epicycles) reveal fundamental weaknesses in a theory’s architecture.

Cognitive Advantages

Neurodivergent cognition provides unique leverage in this process. Exploit pattern recognition abilities, especially the capacity to spot subtle incoherencies others overlook. Those are investigative superpowers. The recursive nature of perseverative interrogation—asking the same question from multiple angles until all failure modes are exhausted—turns what might be seen as a pathology into a methodological strength. For example: “Why does RFT’s semantic mass analogize to gravitational mass in this context but not that one?” becomes a powerful probe of theoretical consistency.