An engineering finding from Arcanamatic on the difference between a model defaulting and a model making the right choice, repeatedly.
The Illusion of Variety
Arcanamatic’s work on an autonomous marketing agent, powered by the NeuraScene platform, produced a specific, verifiable finding about the behaviour of large language models in stateless environments. The agent is designed to select a persuasion trigger, a target emotional affect, and a Gestalt visual principle for each campaign it generates. In its first seven live test runs, an unexpected pattern emerged.
Across seven independent, stateless calls to the model, it selected:
- Authority Bias: Six times.
- Cognitive Ease: Five times.
This repetition was not the result of a lazy or poorly-designed prompt. Crucially, each run was a separate, stateless API call. The model had no memory of its previous selections. An instruction to “be more varied” would have been meaningless, as there was no prior context against which to vary its output. The model was not ‘stuck in a rut’; it was arriving at the same conclusion from the same starting point, every time.
Designing for Memory
The engineering challenge was not one of prompt-craft, but of system design. Two specific architectural changes were implemented to address this behaviour.
- Contextual Feedback Loop: The most recent N selections are now fed back into the model’s context window for the next run. This provides a memory of recent choices, allowing an instruction like “select a novel persuasion trigger not present in the recent history” to become executable.
- Funnel-Stage Constraint: The agent is now passed the marketing funnel stage (e.g., Brand Awareness, Consideration, Conversion) as a hard constraint. This narrows the set of appropriate psychological triggers. A brand awareness piece has different valid triggers than a direct-response conversion ad, focusing the model's decision-making on what is strategically coherent.
When Repetition is Correct
The most interesting lesson, however, was that some of the repetition was not just logical, but correct. The campaigns being generated were often technical proof pieces, explaining the mechanics of the NeuraScene platform itself.
In this context, Authority Bias — leveraging the credibility of expertise and data — is not merely a valid choice, but arguably the optimal one.
Forcing the system to choose a different trigger, such as Scarcity or the Decoy Effect, would have purchased novelty at the direct cost of coherence. The campaign would have been different, but worse.
The core takeaway is a lesson in designing for and interpreting memoryless systems. It is essential to distinguish between a model defaulting to a lazy answer and a model independently concluding that the same answer is repeatedly the right one. The solution is not always to force variety, but to build systems that know when consistency is the more intelligent choice.
Method note. This piece was built in NeuraScene's Advertising Mode using an explicit psychological frame:
- Persuasion trigger: The Halo Effect
- Target affect: Cognitive Ease
- Gestalt principle: Law of Similarity
The Halo Effect is triggered by showcasing deep, transparent engineering competence, building trust in the Arcanamatic brand. Cognitive Ease is the intended affect for the reader, making a complex AI design problem feel clear and solvable. The Law of Similarity visually represents the core problem of repetition and variation, making the abstract concept tangible.
