This appendix applies the Information Bottleneck method (Tishby, Pereira & Bialek, 1999) to formalize the framework’s observation that analysis has an optimal stopping point. The IB Lagrangian and the compression–relevance trade-off are established information theory; the mapping to the loop’s “fertile boundary” is the framework’s contribution.
1. The Japanese Rock Garden Problem
The framework identifies a fertile boundary where elimination has an optimal stopping point.
- Raking too much: A sterile parking lot (Over-compression).
- Raking too little: An overgrown lot (Over-fitting).
This section formalizes this boundary using Information Bottleneck (IB) Theory (Tishby, 1999).
2. The Compression-Relevance Trade-off
In a loop, the Instrument (I) is an encoder that transforms the high-dimensional “Bulk” input (X) into a formal representation (Y). The Sensor (S) then uses Y to recognize the “Truth” (T).
The Information Bottleneck is defined by a Lagrangian:
ℒ = I(Y; X) − β I(Y; T)
Where I(Y; X) is the Complexity (how much the instrument has “raked” the data), I(Y; T) is the Relevance (how much “truth” the sensor can still recognize), and β is a parameter that balances the two.
3. The Fertile Boundary as a Phase Transition
The “Japanese Rock Garden” is the optimal β point.
- As β → 0 (Dead Speech): The instrument compresses everything into a tiny summary. The relevance I(Y; T) drops to zero. The “sterile parking lot”—no moss remains.
- As β → ∞ (Overgrowth): The instrument captures every detail, including the noise. The complexity I(Y; X) explodes, making recognition impossible for the sensor’s finite processing.
The Fertile Boundary is the “Knee” of the IB curve—the point where we get the Maximum Relevance for the Minimum Complexity.
4. Why “Round 5” Output is Dead Speech
The Ancient Song Round 5 results are the empirical evidence of Over-Compression. Each analytical pass is an attempt to “rake the sand” more precisely. Eventually, the loop passes the Information Bottleneck Critical Point, and the “Synergy” (the moss) is filtered out.
A “Living Document” must intentionally stop raking. It must leave a “Fertile Boundary” of uncompressed wildness to keep the sensor engaged.
5. Summary: Epistemology as Optimal Compression
- Elimination (the sculptor) is the act of increasing compression.
- Recognition is the act of checking if the relevance I(Y; T) is still high.
- The Fertile Boundary is the point where the sculptured form and the raw marble are in perfect tension.
Truth is not the summary; truth is the residue that survives the compression.
Toward Testability
The following grounds this appendix in measurable quantities—produced through the Friction Test, where a second instrument (Gemini 3 Pro) critiqued and rebuilt these intuitions.
The β-Parameterized Bottleneck
The AI possesses a massive, noisy latent space of pre-training data X. The human wants a specific, highly relevant answer Y. The conversation T is the bottleneck representation, optimizing the Lagrangian:
ℒIB = I(X; T) − β I(T; Y)
The parameter β is the human’s active prompt engineering strategy:
- Low β (broad prompt: “summarize this”): favors compression. The bottleneck is tight. Output is short and generic—high compression of X, but low relevance to specific details of Y.
- High β (constrained prompt with 20 bulleted requirements): forces the bottleneck open, demanding maximum relevance. Risk of “overfitting” to the prompt, losing creative flexibility.
The iterative conversation is a joint optimization: the human tunes β across turns, sculpting the bottleneck until the output achieves maximum relevance for minimum complexity.