HAI / BIS: A Finite, Shared, Structured Basis of Semantic Primitives Extractable from Large Language Models

Computational test of the Brain Instruction Set (BIS) hypothesis: that humanconceptual knowledge can be represented over a finite, shared, internallystructured basis of semantic primitives recoverable from large language models. This release contains the BIS Release bundle (152 verified terminals across ninemodalities with per-terminal test scores and provenance; schema bis-schema/1.0),the decomposition graph (13,534 concepts / 44,682 weighted edges, GraphML), thecross-model comparison (16-terminal model-invariant core confirmed across anOpenAI and an Anthropic model), the sub-qualia dataset (3,684 stable / 6,928extended named sub-types with per-terminal dimensionality), the saturation curve,the preprint (SK + EN), figures, and code. Activation-level results (mechanistic program). Probing open-weight models showsthe cross-model core indexes genuine internal representational structure:- Reconstruction: the BIS basis explains 39% (16-terminal core) to 59% (all 152) of residual-stream variance in Gemma-2-2B at layer 14, versus <=0.07 for a random basis of equal size; variance explained peaks mid-network.- Readability: terminals are linearly readable from activations at AUC 0.86-0.95.- Causality: norm-calibrated steering of a terminal direction shifts generation.- Cross-architecture: AUC and steering replicate across Gemma-2-2B, Qwen2.5-1.5B and Phi-3.5-mini (three labs). Raw reconstruction R-squared near-saturates for Qwen/Phi due to architecture-specific outlier dimensions, so universality is claimed on the architecture-robust measures (readability and steering).- SAE alignment: all 16 core terminals match a distinct Gemma Scope sparse- autoencoder feature (mean cosine 0.44 vs random null 0.09), so BIS supplies interpretable names for features an unsupervised SAE discovers independently. Together this promotes BIS from a purely behavioural construct toward a candidatemechanistic basis. Mandatory declarations: results characterize the structure of textualrepresentations, not phenomenal experience (qualia), not perceptualdiscriminability, and without external anchors not calibrated truth.Activation-level results are demonstrated across three architectures at smallscale; larger-scale replication and a formal steering benchmark are ongoing work. Author: Daniel Marko (independent researcher). Computational implementationassisted by large language models. Data licensed CC-BY-4.0; code MIT. Idealineage traces to geocities.com/danielis1975 (c. 2000).