Mass conservation and depth cap
Stub README (2026-09-28): this directory predates the README convention. The summary below was reconstructed from its files; see them for detail.
Question. Does the trie depth cap distort path-probability convergence beyond the ordinary noise of mass-1 paths?
Answer. No. In the Shakespeare d=16 shared-path data, max-depth paths with mass ≥2 converge about as well as shallow paths with mass ≥2 (Spearman ρ 0.9095 vs 0.9211). Deep paths look noisy only because most of them are mass-1. The notes also give node mass conservation: count = child-edge counts + terminal count + cutoff count.
- max-depth paths (d=16), mass ≥ 2: Spearman ρ of log path probability between paired tries = 0.9095
- shallow paths (depth ≤ 8), mass ≥ 2: Spearman ρ of log path probability between paired tries = 0.9211
Sources. findings.md question/answer, results table and interpretation; notes.md conservation-law derivation; mass_filter_experiment.py reads rnd/convergence shared-path CSVs.
Caveats. rnd/TRIAGE.md lists this directory under AFFECTED-BY-BUG as 'directly relevant — redo'. But the analysis uses trie/corpus statistics only, with no trained model, and rnd/README.md calls it 'not obviously trainer-dependent', so the loss bug should not apply. The uniform-mass follow-up was not run. findings.md's reproduce command points at rnd/mass_filter_experiment.py, but the script is in rnd/mass-conservation/.