mapwise state representations: byte language C built from codex, both vocabs projected through it (full row contract); joint content-state map + words-of-C table

mapwise state representations: byte language C built from codex, both vocabs projected through it (full row contract); joint content-state map + words-of-C table

mapwise state representations: byte language C built from codex, both vocabs projected through it (full row contract); joint content-state map + words-of-C table

mapwise state representations: byte language C built from codex, both vocabs projected through it (full row contract); joint content-state map + words-of-C table

mapwise state representations: byte language C built from codex, both vocabs projected through it (full row contract); joint content-state map + words-of-C table

byte-specific state gauge: A 82449 / B 63001 trie states, 28246 co-occupied; 12.0K/12.5K whole states walk fully on shared structure vs 10801 exact name matches

lexicon byte-state gauges: A 31997, B 29523, overlap 10801, union 50719, irreducibles 1048 (atoms: 10 digits + letters + punct; zero word-starts), 95/256 single bytes present

btx-e001 T3-a: matched mask arity moves arith:nm only 0.107->0.127 (still < 0.161 majority-digit baseline) — crater is content entropy, not protocol; 1x1 rows reproduce T-2 exactly

btx-e001 T-2 control: vanilla t5-small = wordpiece parity (0.47/0.61) through the same pushforward; flan sentinel infill EOS-degenerate — channel exonerated, flan indicted

btx-e001 T-2 control: vanilla t5-small = wordpiece parity (0.47/0.61) through the same pushforward; flan sentinel infill EOS-degenerate — channel exonerated, flan indicted

btx-e001 T-2: shunt fidelity ledger — cloze-BERT byte marginal hits 0.43-0.61 gold; flan-t5 sentinel infill 0.05-0.09; n:m cells crater both

btx-e001 T-1: matrix size verdict — closed register saturates at 94 cells by 16KB; open-class digits grow 1276+ cells and flip modal shape to 2x2

btx-e001 T-1: matrix size verdict — closed register saturates at 94 cells by 16KB; open-class digits grow 1276+ cells and flip modal shape to 2x2

btx-e001 T-1: matrix size verdict — closed register saturates at 94 cells by 16KB; open-class digits grow 1276+ cells and flip modal shape to 2x2

btx-e001 T-0: flan-t5-small + tinybert tables, gated, coverage disclosed (newline is the mutual floor; case one-way lossy into wordpiece)

btx-e001 T-0: flan-t5-small + tinybert tables, gated, coverage disclosed (newline is the mutual floor; case one-way lossy into wordpiece)

btx-e001 T-0: flan-t5-small + tinybert tables, gated, coverage disclosed (newline is the mutual floor; case one-way lossy into wordpiece)

vocab_qwen25.jsonl — Qwen2.5 family table (extract.py fleet extractor, round-trip gated)

README: link the standalone code repo (github.com/AbstractEyes/geolip-bytelex) + git+ install line; regeneration section now points at the library, not alephllm

bytelex v1: FULL lexicon translation matrix (1.25M tokens across 12 vocabularies) + byte_lexicon_v1 corpus tables (8.2MB mixed corpus, char1-4 exact, hashed 9-gram, word-grams)

bytelex v1: 12 normalized token->bytes vocabulary tables (fleet-extracted, round-trip spot-checked)

bytelex v1: README — model-free byte-information relational system, artifact contract, loss primitives

initial commit