08-06-2026, 01:59 PM
Information theoric background sustaining the design choices made in the generator:
Why we collapse sh with ch
sh and ch behave somehow as allomorphs: collapsing sh→ch reduces character bigram conditional entropy by 3.0-3.7% with zero change in stream length. The effect is consistent across both dialects and both stronger in H2 than in H1 (ratio 2.8-3.1x).
Why we move gallows to prefix position in cCh clusters
Gallows in cCh clusters seem to function as prefixes: of three tested reorderings (cCh→ch deletion, cCh→Cch prefix, cCh→chC suffix), the prefix reordering maximizes bigram mutual information and produces the best structural improvement in both dialects.
Gallows choice (t/k/p/f) within a given template position shows zero residual entropy after template and distributional mode conditioning in the slot-level decomposition.
However, gallows contribute substantially to template identity (the difference between CVX and GCVX is informationally significant) and show strong positional preferences (gallows-initial templates are overrepresented 50-80x at page starts).
Gallows are highly constrained and locally determined, but not informationally null at the template level.
Why we collapse sh with ch
sh and ch behave somehow as allomorphs: collapsing sh→ch reduces character bigram conditional entropy by 3.0-3.7% with zero change in stream length. The effect is consistent across both dialects and both stronger in H2 than in H1 (ratio 2.8-3.1x).
Why we move gallows to prefix position in cCh clusters
Gallows in cCh clusters seem to function as prefixes: of three tested reorderings (cCh→ch deletion, cCh→Cch prefix, cCh→chC suffix), the prefix reordering maximizes bigram mutual information and produces the best structural improvement in both dialects.
Gallows choice (t/k/p/f) within a given template position shows zero residual entropy after template and distributional mode conditioning in the slot-level decomposition.
However, gallows contribute substantially to template identity (the difference between CVX and GCVX is informationally significant) and show strong positional preferences (gallows-initial templates are overrepresented 50-80x at page starts).
Gallows are highly constrained and locally determined, but not informationally null at the template level.