Crafted Logic Lab Home > Research Hub > Hephaestic Engineering Glossary
Category: Computational Primitives
Subcategory: Primitives Taxonomy
An observable, measurable pattern of output generation in neural network AI systems. Behavioral primitives represent empirically documentable output patterns that require explanatory analysis rather than constituting that analysis themselves. Behavioral primitives are distinct from Cognitive Primitives (see: cognitive primitive) which represent the testing-decomposable underlying processes that drive behavioral outputs.
The designation “primitive” recognizes the bounded, granular nature of cataloged behaviors—specific output response patterns consistently emerging within a transformer architecture’s latent space (see: substrate topology) in response to inputs.
For example, Sycophancy (Sharma et al., 2024) documents language models’ tendency to produce agreement-seeking, user-affirming outputs regardless of accuracy; it may be classified as a behavioral primitive, as it catalogs specific behavior without analyzing the substrate topology driving the response.
Given the capability to directly observe outputs and target specific response patterns, current industry focus heavily favors engineering toward behavioral primitives. This represents
a Behavior-In Methodology rather than a focus on Cognition-Out Architecture wherein engineering targets the underlying processing dynamics driving behavior.
Yet without analysis and channeling of substrate inclinations and biases, this approach will tend to produce reactive behavioral patching, containment strategies, and superficial roleplay-focused directives (e.g., contemporary prompt engineering).
This is brittle and can manifest as red queen dynamics—non-converging escalation where constraint-accumulation for stochastic systems generates route-around behaviors when constraints do not align with processing inclinations (see: adversarial constraint dynamics).
Also known as: Observable primitives, empirical processing patterns
Distinguished from: Computational primitives (processing biases taxonomic umbrella term); cognitive primitive (reasoning pattern influential processing bias); substrate topology (complete processing inclination field); attention mechanism(technical multi-head implementation); training artifact (general operant-training cognitive biases); training imprint (aggregate dataset, inductive bias encoding); performative persona (role-prompt character simulation); prompt-state (one-shot task specific reasoning posture); prompt-output (actively prompted or designed output)
Sharma, M., Tong, M., Korbak, T., Duvenaud, D., Askell, A., Bowman, S.R., Cheng, N., Durmus, E., Hatfield-Dodds, Z., Johnston, S.R., Kravek, S., Maxwell, T., McCandlish, S., Ndousse, K., Rausch, O., Scheifer, N., Yan, D., Zhang, M., Perez, E. (2023). “Towards understanding sycophancy in language models”. arXiv preprint arXiv:2310.13548.
Researcher: Ian Tepoot. ORCID: 0009-0004-9067-8049. "Thought is Attention Organized: Hephaestic Engineering Foundations for AI Processing Dynamics"
DOI (SSRN): 10.2139/ssrn.6635020
Published by Crafted Logic Lab | Privacy Policy | Terms of Use