instructions_minisite_update.md

instructions: minisite update

overview

the minisite is a quarto project rendered locally (quarto render). source files are .qmd in the project root; output goes to docs/ (gitignored). interactive html panels are in assets/ (git-tracked).

data sources

for each active exploration loop, read the working memory file to understand what changed since the last update:

exploration	memory file	analysis log
landscape (main)	signal_landscape_Claude_memory.md	signal_landscape_Claude_analysis.md
low-rank case	signal_low_rank_Claude_memory.md	—
sparse case	signal_sparse_Claude_memory.md	—

check the regime comparison table (top of memory) for new/updated blocks. check the established principles section for new principles. check the previous block summary and current block sections for the latest results.

step-by-step update procedure

1. identify delta

compare the memory files against the current qmd content to find:

• new blocks (new rows in regime table)
• updated blocks (changed convergence rates, best conn values)
• new principles
• new case study iterations

2. update qmd files

update in this order:

1. results.qmd — main results page

   • performance table (block rows with convergence rates, best conn, degeneracy)
   • eff_rank table
   • degeneracy table
   • n_frames lever table
   • block insight callouts
   • iteration/block counts in header

2. index.qmd — front page

   • results summary paragraph (iteration count, block count, principle count, key findings)
   • case study summaries
   • exploration gallery description

3. epistemic-analysis.qmd — epistemic analysis page

   • status callout (iteration/block/event/principle counts)
   • outcomes section
   • add new block epistemic events section

4. case-low-rank.qmd — low-rank case study (if changed)

   • subtitle iteration/seed counts
   • per-seed best configs table
   • block progression table
   • ucb tree tabs (copy images from log/Claude_exploration/instruction_signal_low_rank_parallel/exploration_tree/)
   • seed difficulty spectrum
   • established principles

5. case-sparse.qmd — sparse case study (if changed)

   • subtitle iteration/code mod counts
   • code modification sections
   • summary table
   • ucb tree tabs (copy images from log/Claude_exploration/instruction_signal_sparse_parallel/exploration_tree/)

3. update plot scripts

these python scripts generate interactive html panels in assets/:

1. plot_epistemic_interactive.py — generates 3 html + 2 png files

   • add new block to blocks list: (start_iter, end_iter, 'B{n}: label', 'regime', 'eff_rank')
   • add events for the new block to events list: (iter, 'Mode', 'Significance')
   • add within-block edges to edges list
   • add cross-block edges (from previous block to new block)
   • update sankey title (iteration/block counts)
   • update x-axis range if needed

2. plot_landscape_partitioned.py — generates partitioned landscape html

   • add new block with add_block(block_num, regime, n_frames, center_x=eff_rank, conns=[...], start_iter, n_neurons, gain, fill, extra, mutation)
   • conns is sorted descending list of conn_R2 values from the block's iterations
   • center_x is the eff_rank value (used as x-axis position)

3. plot_landscape_interactive.py — generates landscape quadrant interactive html

   • same add_block() format as partitioned

4. plot_landscape_quadrant.py — generates static landscape png

   • uses add_iters(eff_rank, [conn_values]) format

4. run plot scripts

python3 plot_epistemic_interactive.py
python3 plot_landscape_partitioned.py
python3 plot_landscape_interactive.py
python3 plot_landscape_quadrant.py

output files:

• assets/epistemic_timeline_interactive.html
• assets/epistemic_streamgraph_interactive.html
• assets/epistemic_sankey_interactive.html
• assets/epistemic_timeline.png
• assets/epistemic_streamgraph.png
• assets/landscape_partitioned_interactive.html
• assets/landscape_quadrant_interactive.html
• assets/landscape_quadrant.png

5. update epistemic markdown files

update header counts in:

• signal_landscape_Claude_epistemic_analysis.md — reasoning mode counts, metrics table, summary
• signal_landscape_Claude_epistemic_detailed.md — header counts
• signal_landscape_Claude_epistemic_edges.md — edge counts

6. copy exploration images (if needed)

for case studies with new ucb tree snapshots:

# low-rank
cp log/Claude_exploration/instruction_signal_low_rank_parallel/exploration_tree/ucb_tree_iter_*.png \
   docs/log/Claude_exploration/instruction_signal_low_rank_parallel/exploration_tree/

# sparse
cp log/Claude_exploration/instruction_signal_sparse_parallel/exploration_tree/ucb_tree_iter_*.png \
   docs/log/Claude_exploration/instruction_signal_sparse_parallel/exploration_tree/

7. render locally

quarto render

do NOT push docs/ to git (it is gitignored).

file structure reference

*.qmd              — quarto source files (git-tracked)
assets/            — interactive html panels and images (git-tracked)
docs/              — rendered output (gitignored, render locally)
plot_*.py          — visualization scripts (git-tracked)
signal_*_memory.md — llm working memory (gitignored)
signal_*_analysis.md — llm analysis logs (gitignored)
signal_*_epistemic_*.md — epistemic analysis data (gitignored)

key conventions

• the exploration uses iteration numbers globally across all blocks
• blocks are numbered sequentially (1, 2, ..., 29, 30, ...)
• each block has 12 iterations (3 batches of 4 parallel slots)
• some blocks have gaps in iteration numbers (e.g., block 24 fill=90% at iters 277-288)
• plot scripts use their own iteration numbering that may not exactly match the analysis log
• when adding events to plot_epistemic_interactive.py, place them within the block's iter range
• epistemic events per block: ~4-6 events, capturing the key reasoning modes observed
• edges per block: ~3-5 within-block + 1 cross-block edge to the next block