Archive: 20260131
Status: FROZEN (do not modify - see #agglutinative for protocol)

== SUMMARY ==

First batch of interpretability experiments.
Key result: 5 ESs (Digits, Punct, Vowels, Whitespace, Other) explain
59% of model compression (1.37 of 2.31 bits/char).

== REPRODUCIBILITY ==

Repository: /home/admin/hutter
Model: model.bin (Elman RNN 256->128->256)
Performance: 5.69 bpc on 100K sample
Data: enwik9 (http://mattmahoney.net/dc/enwik9.zip)

To reproduce:
  wget http://mattmahoney.net/dc/enwik9.zip && unzip enwik9.zip
  make
  ./hutter eval enwik9 model.bin
  ./hutter es enwik9 model.bin

== CONTENTS ==

Papers (source + PDF):
  rnn-um-mapping.tex/pdf      - Doubled-E: 0% bpc difference
  activation-probing.tex/pdf  - Query language for hidden state

  To rebuild: pdflatex rnn-um-mapping.tex && pdflatex rnn-um-mapping.tex

Figures:
  fig-20260131-01: theories.html (pattern validation)
  fig-20260131-02: viz.html (activation heatmaps)
  fig-20260131-03: dashboard.html (main dashboard)
  fig-20260131-04: hypothesis-results.html (ES hypothesis tests)
  fig-20260131-05: coactivation-test-20260131.txt (neuron persistence)
  fig-20260131-06: sn-view.html (SN representation)
  fig-20260131-07: sn-layered.html (layered ES graph)

Data:
  model.bin (RNN checkpoint)
  model.sn (Sparse Network representation)

== KEY FINDINGS ==

1. ES Evaluation (fig-07, ./hutter es):
   - RNN: 5.69 bpc, ES-only: 6.63 bpc
   - 5 character classes explain 59% of compression
   - Remaining 0.94 bits/char from byte-specific patterns

2. Character Class ESs (fig-04):
   - Digits, Punctuation, Vowels confirmed as ESs
   - Consonants too diverse for single ES

3. Persistent Neurons (fig-05):
   - h[34], h[48], h[62], h[71], h[113] active across sequences

== NEXT ==

New experiments go in archive/20260131_2/