Back to Index

Archive: 2026-01-31

Hypothesis-driven ES discovery session

Model: Elman RNN 256-128-256, ~5.69 bpc (100K sample)
Key Result: 5 ESs (Digits, Punctuation, Vowels, Whitespace, Other) explain 59.3% of compression
Data: enwik9 (first 100KB for testing)

Figures

fig-20260131-07: SN Layered Graph

Named ESs, compression framing, 59.3% explained

fig-20260131-06: SN Representation View

Sparse Network structure in UM formalism

fig-20260131-04: Hypothesis Test Results

Digits 0.9996, Punctuation 0.999, Vowels 0.988

fig-20260131-05: Coactivation Chain Analysis

Persistent neurons: h[34], h[48], h[62], h[71], h[113]

fig-20260131-03: Main Dashboard

Model overview, pattern strengths

fig-20260131-01: Theory Validation

Tests P-1 through P-8

fig-20260131-02: Activation Heatmaps

Per-neuron activation by context

Papers

RNN to UM Mapping

Doubled-E baseline achieves 0% bpc difference

Activation Probing Framework

Query language for hidden state analysis

Data Files

RNN checkpoint (329 KB)

Sparse Network representation

Supporting Files

hypothesis-test-20260131.txt README.txt