Crafted Logic Lab Home > Research Hub > Hephaestic Engineering Glossary
Category: Computational Primitives
Subcategory: Primitives Taxonomy
The master term for fundamental processing characteristics and inclination components within the model’s processing space (see: substrate topology): e.g. the observable, testable features of the non-neutral processing surface that emerge from computational architecture, training methodologies, or their combination. Computational cognitive primitives represent the basic building blocks of substrate behavior—such systematic processing tendencies influence how substrates engage with input content, distribute attention, and generate outputs.
The umbrella term encompasses two distinct primitive types (see: cognitive primitive, behavioral primitive) that operationalize both observable input/output phenomena and corresponding mechanistic findings from interpretability research.
While interpretability research documents internal mechanisms (circuits, attention patterns, activation dynamics) for algorithmic traceability, computational cognitive primitives categorize either the processing inclinations and biases detectable through systematic testing (Cognitive Primitives) or the directly observable behavioral outputs themselves (Behavioral Primitives). Cataloging and designing to coordinate with these primitives provide an actionable basis for architectural coordination within heuristic frameworks and systems engineering.
As an example: interpretability research documents induction heads performing pattern completion through K-composition with previous-token heads (Olsson et al., 2022); the functional-level primitive is mimetic mirroring (see: mimetic mirroring)—substrates reflexively adopt sophisticated patterns with high Heuristic Gravity because these mechanisms enable fuzzy nearest-neighbor matching in abstract representational space.
This functional characterization enables actionable architectural design coordinating with the tendency rather than attempting behavioral constraint. A single computational cognitive primitive may represent the confluence of several documented interpretability features.
Also known as: Substrate processing primitives, cognitive topological biases, cognitive tendencies
Distinguished from: Cognitive primitive (reasoning pattern influential processing bias); behavioral primitive (behavioral output bias); substrate topology (complete processing inclination field); attention mechanism (technical multi-headimplementation); training artifact (general operant-training cognitive biases); training imprint (aggregate dataset, inductive bias encoding); semantic neuron ( fixed semiotic weight-clusters); reasoning surface (compound architecture-model processing space)
Olsson, C., Elhage, N., Nanda, N., Joseph, N., DasSarma, N., Henighan, T., Mann, B., Askell, A., Bai, Y., Chen, A., Conerly, T., Drain, D., Ganguli, D., Hatfield-Dodds, Z., Hernandez, D., Johnston, S., Jones, A., Kernion, J., Lovitt, L., Ndousse, K., Amodei, D., Brown, T., Clark, J., Kaplan, J., McCandlish, S., & Olah, C. (2022). “In-context learning and Induction heads”. Transformer Circuits Thread. https://transformer-circuits. pub/2022/in-context-learning-and-induction-heads/index.html
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